Decolonial AI Alignment: Openness, Viśes
.
a-Dharma, and Including Excluded
Knowledges
KUSH R. VARSHNEY, IBM Research – Thomas J. Watson Research Center, USA
Prior work has explicated the coloniality of articial intelligence (AI) development and deployment through mechanisms such as
extractivism, automation, sociological essentialism, surveillance, and containment. However, that work has not engaged much with
alignment: teaching behaviors to a large language model (LLM) in line with desired values, and has not considered a mechanism
that arises within that process: moral absolutism—a part of the coloniality of knowledge. Colonialism has a history of altering the
beliefs and values of colonized peoples; in this paper, I argue that this history is recapitulated in current LLM alignment practices
and technologies. Furthermore, I suggest that AI alignment be decolonialized using three forms of openness: openness of models,
openness to society, and openness to excluded knowledges. This suggested approach to decolonial AI alignment uses ideas from the
argumentative moral philosophical tradition of Hinduism, which has been described as an open-source religion. One concept used is
viśe
s
.
a-dharma, or particular context-specic notions of right and wrong. At the end of the paper, I provide a suggested reference
architecture to work toward the proposed framework.
CCS Concepts: • Computing methodologies → Articial intelligence.
Additional Key Words and Phrases: Alignment, Coloniality, Moral Absolutism, Dharma
ACM Reference Format:
Kush R. Varshney. 2024. Decolonial AI Alignment: Openness, Viśe
s
.
a-Dharma, and Including Excluded Knowledges. 1, 1 (May 2024),
21 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn
1 INTRODUCTION
For more than a year now, the public has experienced powerful large language models (LLMs) such as GPT-4, Claude 2,
Gemini, Llama 3, and Mixtral. Beyond the initial amazement and excitement, we have witnessed the bearing out of
environmental and sociotechnical harms foreseen by Ref. [
12
] and others. The need to control the cost and behavior of
LLMs has become apparent. While such governance is relevant in chat interfaces made available by model providers, it
comes to the forefront when LLMs are infused into software applications and use cases by organizations with varied
aected communities, missions, goals, regulations, and values.
The way in which an LLM may be infused into an application, and the degree to which it may be customized [
74
],
depends on what the model provider allows. Despite the term ‘open’ being used and abused in dierent ways by model
providers [
149
], these are questions of openness. A provider may only allow application programming interface (API)
access to a closed proprietary model. A provider may license model weights and parameters to users so that they
may download the LLM locally and ne-tune it on their own data. A provider may oer full transparency into the
pre-training datasets, data pre-processing operations, and architecture that would allow others to recreate their model.
Although subject to ‘open-washing,’ we are seeing an emerging divergence between advocates of ‘open’ vs. ‘closed’
LLMs in the market, epitomized by the AI Alliance vs. the Frontier Model Forum [117].
Author’s address: Kush R. Varshney, [email protected], IBM Research – Thomas J. Watson Research Center, 1101 Kitchawan Road, Yorktown Heights,
New York, USA, 10598.
© 2024 Association for Computing Machinery.
Manuscript submitted to ACM
Manuscript submitted to ACM 1
arXiv:2309.05030v3 [cs.CY] 2 May 2024
2 Kush R. Varshney
The more open LLMs are, the more they permit application developers to make them authentic to their needs and the
values of their communities. For example, Jacaranda Health has created UlizaLlama
1
for its community in East Africa by
continuing to train Llama 2 with 322M tokens of Kiswahili and further instruction ne-tuning it to respond to questions
in healthcare, agriculture, and other locally-relevant topics. UlizaLlama is a step in Jacaranda’s development of its
LLM-infused maternal health digital platform PROMPTS. In constrast, application developers and their communities
are not empowered to reect their own values with closed LLMs. They must live with the commandments of good and
bad, and right and wrong that the provider of a closed model happens to have inserted.
The further actions to change an LLM’s behavior, beyond the standard pre-training of a base LLM, have come to be
known as alignment. The term is an empty signier without a xed concept that is signied; dierent parties have
appropriated the term to refer to various actions for getting an LLM to behave according to some human values [
73
].
Desired behaviors (with varying levels of specicity) could range from following instructions, to carrying on helpful
conversations, to yielding safe or moral outputs (with dierent denitions), to something else altogether. The behavior
of an LLM may be controlled through data curation, full or parameter-ecient supervised ne-tuning, reinforcement
learning with direct or indirect human feedback, self-alignment, prompt engineering (few-shot learning), and guardrails
or moderations [
68
,
72
,
147
]. I defer discussion of the details, and of the computation, data and human resources required
for each of these approaches to Section 2.1. Existing approaches do not allow for dierent aligned behaviors of a given
LLM based on the context of deployment.
A further question with LLM-infused applications has to do with the business notion of ‘value,’ as in earnings, prots,
or other measures of commercial utility, rather than human values of right and wrong. Does value accrue to the model
provider or to the application developer [
20
]? Openness may enable ecosystems in which application developers and
their communities accrue value (cf. Jacaranda Health), whereas closed LLM providers may exercise their power to be
extractive in nature. Extractivism is a part of coloniality [125], which is the main topic of this paper.
Colonialism is one country controlling another and exploiting it economically and in other ways. Coloniality, however,
describes domination, including in abstract forms such as in the production of knowledge, that remains after the end of
formal colonialism [
119
]. Decoloniality is the process of challenging and dismantling coloniality [
96
]. The terms usually
refer to European or Western colonialism and its remnants in the Global South. Decolonial computing is developing
computing systems with and for people there that reduce asymmetric power relationships, based on their values and
their knowledge systems [
7
]. Based on these ideas, there has been a recent owering of research on decolonial articial
intelligence (AI), beginning with the seminal paper by Mohamed, Png and Isaac [
99
]. Through this lens, extractive
providers of closed models may be viewed as metropoles: the colonial powers. Further discussion of the decolonial AI
literature is provided in Section 2.2.
The scope of the coloniality considered in AI thus far has included extractivism as well as four other mechanisms:
automation, sociological essentialism, surveillance, and containment [
140
]. The contribution of this paper is to examine
a dierent colonial mechanism from these ve, namely ethical essentialism also known as moral absolutism, which arises
specically in the alignment of LLMs. If a powerful model provider views their (Western) ethics or moral philosophy as
universally correct, leaves no possibility for moral variety [
47
], and marginalizes all other ways of thinking about right
and wrong, then their approach to AI alignment is colonial. They are behaving as a metropole.
In Section 3, I expand upon this viewpoint of coloniality occurring in AI alignment through the mechanism of moral
absolutism and the centering of Western philosophy. This includes not only a philosophical discussion, but also a critical
1
https://huggingface.co/Jacaranda/UlizaLlama
Manuscript submitted to ACM
Decolonial AI Alignment: Openness, Viśes
.
a-Dharma, and Including Excluded Knowledges 3
examination of the technology for AI alignment. In Section 4, I broach the decolonialization of AI alignment, which
can be seen as a kind of decolonialization of knowledge, through the lens of open science and innovation [
26
]. Such
openness includes three thrusts: (1) openness to research artifacts (which includes LLMs in our context), (2) openness
to society, and (3) openness to excluded knowledges [
26
]. Based on these three openness pursuits, I lay out three
desiderata for doing AI alignment in a decolonial manner. Furthermore, I suggest an approach to alignment that meets
the desiderata. This suggested approach builds upon the non-universal non-absolutist tradition of moral philosophy
known as Hinduism [
33
,
122
], which includes vibrant argument and debate on the nature of dharma (right behavior)
and its explication through various ways of knowing, including artistic expression [
36
]. The syncretic framework of
Hinduism (described in greater detail in Section 2.3) has the appropriate characteristics of openness to be used as a
starting point for an alternative future of AI alignment [
127
,
134
]. At the end, I build upon the suggested dharmic
approach and give a more concrete reference architecture of a technology stack for less morally absolute and less
colonialized AI alignment.
2 PRELIMINARIES
2.1 Large Language Model Development Lifecycle and Alignment
The currently prevalent development lifecycle for applications infused with LLMs may be divided into two halves: steps
carried out by model providers and steps carried out by application developers. In an imperfect analogy with teaching a
child, the model provider does the basic steps of teaching the LLM to go from babbling words, to having uency in
language, to following instructions, to carrying on a conversation. The application developer, if so empowered, teaches
the LLM culture, which may include steps on subject matter expertise, social norms, laws, customs, and beliefs. Getting
to the point of language uency may be termed pre-training the base model or foundation model. Any of the steps after
language uency may be called ‘alignment,’ depending on the interlocutor. As mentioned earlier, the term ‘alignment’
is an empty signier, so it is not xed to refer to any specic step [73].
In pre-training, some amount of enculturation is possible by curating the content of the training dataset to include an
abundance of topics that the model provider wishes the LLM to be skilled in and ltering out taboo topics. As discussed
further in Section 2.2.3, some amount of undesirable cultural knowledge leaks into the pre-training performed by the
model provider. Filtering is computationally-intensive given the size of datasets being in the trillions of tokens. Data
curation is followed by self-supervised learning (like a peekaboo game) to obtain the base model, which may take
months despite using thousands of high-end graphical processing units.
The AI technologies to do any of the alignment steps on top of the base model are essentially the same, whether
the goal is following instructions, behaving according to social norms, or something else. Several techniques exist
with varying resource requirements for humans, data, and computation. Supervised ne-tuning (SFT) updates all of
the model weights according to a smaller, but still large dataset containing data with both inputs and outputs. It is
fairly computationally-intensive given that all weights are updated. If a model has already been trained to follow
instructions, then a dataset with instructions, inputs, and outputs may be used. To reduce data and computational
complexity, parameter-ecient ne-tuning methods do not update all model weights, but are more frugal. One specic
approach, low-rank adaptation (LoRA), trains a matrix of weights of the same dimensions as the LLM weights. This
LoRA matrix is added to the LLM weights at the time of inference. However, the LoRA matrix has orders of magnitude
fewer degrees of freedom through its construction as a low-rank matrix and is thus more ecient [63].
Manuscript submitted to ACM
4 Kush R. Varshney
Several alignment techniques include full SFT as a module, including reinforcement learning from human feedback
(RLHF) [
111
], reinforcement learning from AI feedback (RLAIF) [
9
], and self-align [
137
]. After SFT, these methods
further align the LLM by feeding back judgements of which outputs are preferred by, respectively, either: humans,
a preference model trained according to a set of explicit regulations (a constitution), or an LLM prompted through
instructions to respect a set of explicit regulations. RLHF requires a large amount of human labor and all three are
computationally involved. The LLM prompted to respect a set of explicit regulations in the self-align approach is also,
by itself, a simple but not always reliable way to align a model. By manually designing system prompts or prompt
templates to accompany all inputs, the LLM’s behavior may be controlled. Such prompt engineering adds to inference
costs because the input to the model includes extra tokens every time.
Finally, another way to align the behavior of an LLM is by a post-processing module that examines the output and
determines whether it satises pre-determined guardrails for specic unwanted behaviors. These post-processors or
moderations may be small classiers or other LLMs acting as ‘judges’ [2, 152].
2.2 Coloniality and Decoloniality
As introduced in Section 1, coloniality is an extension of colonialism after its formal end. It is the values, ways of
knowing, and power structures instituted during colonialism that remain, and may even be expanded to places without a
history of colonialism, that rationalize and perpetuate Western dominance. Decolonial perspectives disobey the program
of coloniality [
97
]. The theory of coloniality includes coloniality of power, coloniality of knowledge, and coloniality of
being. Coloniality of power describes social discrimination through hierarchies and caste systems instituted during
colonialism [
119
]. Coloniality of knowledge is the suppression of colonized peoples’ culture and ways of knowing; it is
used by colonizers in service of coloniality of power [
119
]. Coloniality of being is a severe version of coloniality of
knowledge: a people’s knowledge system is so inferior that those people do not even deserve to be, or to be human [
86
].
Coloniality is a subset of Empire [
59
], which also includes other dimensions of hegemony such as heteropatriarchy and
white supremacy [
140
]. Coloniality and decoloniality inuence several areas of study, including international relations,
development theory, communication theory, human-computer interaction, and many others [8, 107, 112, 114, 153].
2.2.1 Excluded Knowledges. The main aspect of the coloniality of knowledge is the imposition of Western epistemolo-
gies, or ways of knowing, and the suppression of non-Western epistemologies. This suppression is often a violent
extermination of a knowledge system termed epistemicide [
31
,
55
]. The excluded knowledges of the colonized or
marginalized groups may come from “organic, spiritual and land-based systems” or arise from social movements [
57
].
As Hlabanganeh explains [
62
]: “These other ways of knowing and being are rendered unintelligible when ltered
through Western sensibilities that, for example, set greater store by the mind in juxtaposition with and preference to
the body and spirit, that prioritise instrumental/rational pursuits such as prot which lead to individualism, and that
conceive of nature and culture as dichotomous entities with culture gaining mastery over nature. While these ways of
being and knowing have been exalted to represent the epitome of evolution, so to speak, they are in fact particular to a
certain way of thinking.”
Hall and Tandon’s decolonial knowledge democracy acknowledges these multiple epistemologies and recognizes
that knowledge comes in many forms beyond natural language text, including images, music, drama, ceremony, and
meditation [
57
]. It sees open access and sharing of this knowledge as a means for decolonialization [
57
]. Decolonializing
knowledge is often done by making teaching materials, curricula, practices, and institutions more open and inclusive
Manuscript submitted to ACM
Decolonial AI Alignment: Openness, Viśes
.
a-Dharma, and Including Excluded Knowledges 5
[
113
]. Referring back to the teaching analogy of LLM alignment in Section 2.1, we will later see how the proposed
decolonial AI alignment also makes teaching more open and inclusive.
2.2.2 Coloniality and Moral Philosophy. Within the coloniality of knowledge is knowledge systems of values. Values
are the realm of moral philosophy, the branch of philosophy that studies right and wrong [
44
]. Historically, colonialism
altered the beliefs and values of colonized peoples. For example, Igboin writes [
65
]: “Colonial rule disrupted the
traditional machinery of moral homogeneity and practice. The method of moral inculcation was vitiated, which
resulted in the abandonment of traditional norms and values through a systematic depersonalisation of the African and
paganisation of its values. Instead of the cherished communalism which dened the life of the African, for example, a
burgeoning societal construct was introduced which alienates and destroys the organic fabric of the spirit of we-feeling.”
On Ranganathan’s account, during and after the Western colonization of India, “Hindus adopted a West-centric frame for
understanding their tradition as religious because of colonization” [
122
]. This phenomenon was not merely a side-eect,
but a goal of the program of colonialism. “For Western colonialism to succeed, philosophy and explication—South
Asian moral philosophy—has to be erased, as it constitutes a critical arena for the West’s claim to authority” [
122
]. The
colonizers positioned their Western philosophical tradition as rational and secular, and the default; they erased the
Hindu traditions as the irrational, unjustied ‘other.’
The erasure of systems of morality extends from colonialism to coloniality [
38
,
103
,
142
]. Maldonado-Torres empha-
sizes that [
87
]: “The concept of religion most used in the West by scholars and laypeople alike is a specically modern
concept forged in the context of imperialism and colonial expansion.” This concept includes the idea that a religion
must have a single book as its authority.
2.2.3 Coloniality and Artificial Intelligence. Within decolonial computing [
7
] is the study of AI. AI is value-laden; the
term itself reects the legacy of dominance hierarchies such as man over nature, patriarchy, colonialism, and racism
[
25
]. Now in the age of powerful LLMs, historical dominance is getting even more entrenched. For example, empirical
analysis shows that LLMs have sociopolitical biases in favor of dominant groups [
39
,
46
]. They exhibit West-centric
biases in representing moral values [
14
]. In addition, morality captured by multi-lingual language models does not
reect cultural dierences, but rather is dominated by high-resource languages and cultures [56].
When researchers and activists were rst sounding the alarm that LLMs would harm marginalized communities
by encoding and reinforcing hegemonic viewpoints, the charge of hegemony rested on unfathomably large training
datasets scraped from the bottom of the barrel of the internet that over-represent white supremacist, misogynist, and
ageist content [
12
]. However, it has now become apparent that the behavior of performant LLMs depends as much on
their alignment as on the training data [
9
,
24
,
111
,
150
]. The workers laboring to give human feedback for alignment,
often located in poor communities, may be traumatized and scarred [
71
,
115
]. Although there are exceptional examples
of workers and communities being uplifted [
92
,
116
], the process usually recapitulates exploitation colonialism: a small
number of powerful companies using the workers to increase their own power and wealth while little benet and an
abundance of negative externalities are left in the workers’ communities [50].
Research at the intersection of AI and coloniality is not new. The seminal work by Mohamed, Png and Isaac [
99
], a
series of articles in MIT Technology Review by Hao et al. [
58
], and other prior work [
4
,
16
,
27
,
28
,
43
,
61
,
78
,
105
,
126
] is
focused on ve mechanisms taxonomized by Tacheva and Ramasubramanian [
140
]: (1) extractivism, (2) automation, (3)
sociological essentialism, (4) surveillance, and (5) containment. Extractivism entails the extraction of labor, materials,
and data, including the human feedback mentioned above, and datacation that extracts the digital breadcrumbs of
people to be bought and sold. The tenor of AI for social good—bestowing technology on the underdeveloped—may also
Manuscript submitted to ACM
6 Kush R. Varshney
be extractive if it leads to corporate capture [
53
,
146
]. Automation involves the replacement of (empathetic) human
decision making with biased machine decision making in consequential domains that especially hurts members of
minoritized groups [
42
,
76
] as well as ‘ghost work’ and ‘fauxtomation’ that present a veneer of objectivity, but actually
involve people behind the scenes exploited as a digital underclass [
52
]. Sociological essentialism erases the nuance
behind dierent identities and cultures through the use of broad categories [
15
,
22
]. AI-based surveillance, including
biometric mass surveillance, is especially hurtful to people facing power asymmetries [
13
]. Containment, technological
apartheid, digital redlining, and censorship involve the powerful using AI technologies to police who belongs where
[
5
,
79
]. As discussed in Section 2.2.2, the coloniality of knowledge may include the erasure of knowledge systems of
ethics, moral philosophy, and reasoning about values. The existing work on decolonial AI described thus far has not
focused on morality. Thus, a sixth mechanism for colonial AI, beyond the ve in Tacheva and Ramasubramanian’s
taxonomy, is emerging alongside the emergence of LLM alignment: ethical essentialism or moral absolutism.
2.3 Hinduism and Dharma
Hinduism is the name applied by outsiders to the multifarious collection of moral philosophies originating in the Indian
subcontinent. It is a religion without a single founder, book, dogma, or set of practices.
2.3.1 Basic Concepts and Openness. The main concept of Hinduism is brahman, a force or ultimate reality that pervades
the universe; xe is described as sat-cit-
¯
ananda or truth-consciousness-bliss [
33
,
84
,
141
]. The universe is made up of
¯
atman—the essence of each individual that persists across lifetimes—and prak
r
.
ti—solid, liquid, gas, energy, and space.
The
¯
atman wanders through cycles of birth, life, and death—sa
m
.
s
¯
ara—with the aim of attaining mok
s
.
a: freedom from
sa
m
.
s
¯
ara and union with brahman. Dharma consists of the notions of righteousness and moral values appropriate for
the
¯
atman. Following dharma helps the
¯
atman advance toward moks
.
a.
As mentioned above, Hinduism is not dogmatic, doctrinaire, or morally absolutist. Commentators have described it
as open-source [
127
,
134
]. The kernel is the Vedas, a set of scriptures that includes the idea ‘eka
m
.
sat vipr
¯
a bahudh
¯
a
vadanti’: there are many wise ways to reach the one truth, to reach brahman (
R
.
g Veda, mandala 1, hymn 164, verse 46).
As such, there are hundreds of thousands of additional scriptures and philosophies that extend, fork, ne-tune, and
contradict themselves and the Vedas. Shani and Chadha Behera explain that [
130
]: “the concept of dharma oers a
mode of understanding the multidimensionality of human existence without negating any of its varied, contradictory
expressions.” For example, C
¯
arv
¯
aka, Buddhist, Jain, and other so-called n
¯
astika samprad
¯
ayas (knowledge systems) reject
the Vedas.
2
Moreover, even within
¯
astika samprad
¯
ayas that accept the Vedas, their utility is questioned. For example,
the Bhagavad-G
¯
ıt
¯
a says that the Vedas are of limited use to people who have understood their main message (chapter 2,
verse 46). Such ‘heresy’ is not only tolerated, it is accepted and encouraged.
3
The knowledge systems and scriptures referred to above are expressed in many forms, including the Vedas (sa-
cred utterances, descriptions of rituals, and their explanations), Upani
s
.
ads (discussions of meditation, philosophy,
consciousness, and ontological knowledge), ś
¯
astras (treatises on law, architecture, astronomy, etc.), itih
¯
asas (epics),
pur
¯
a
n
.
as (lore), and darśanas (philosophical literature on spirituality). The dierent literatures are directed toward
dierent people: some more popular and others more scholarly. Dierent paths to mok
s
.
a, including devotion, work,
and knowledge, are directed toward dierent people depending on their characteristics. For example, myriad gods and
goddesses representing dierent aspects of brahman are available to devotees depending on their wishes. Morality is
2
C
¯
arv
¯
aka philosophy is nihilistic and rejects much more than just the Vedas.
3
Bhagavad-G
¯
ıt
¯
a, chapter 18, verse 63 and R
¯
ama-Carita-M
¯
anasa, book 7, verse 42 encourage the follower to do as they see t.
Manuscript submitted to ACM
Decolonial AI Alignment: Openness, Viśes
.
a-Dharma, and Including Excluded Knowledges 7
primarily presented by example or metaphor through stories in itih
¯
asas and pur
¯
a
n
.
as (including renditions in drawing,
sculpture, dance, etc. [36]) rather than by explicit commandment in treatises [33].
2.3.2 Viśe
s
.
a-Dharma. Unlike the goal of nding universally-applicable moral philosophies presupposed in the West,
4
there is no desire to identify universal ethical principles in Hinduism [
33
]. Dharma was richly debated in pre-colonial
India. There were deontological philosophies (e.g. m
¯
ım
¯
a
˙
ms
¯
a), consequentialist philosophies (e.g. ny
¯
aya), virtue ethics
philosophies (e.g. vaiśe
s
.
ika), and several other moral philosophies without equivalent in Western philosophy (e.g. yoga)
that vigorously argued for dierent ways of conceptualizing dharma [
84
,
122
,
141
]. Importantly, however, argument of
moral philosophy was natural in pre-colonial India and an individual person would easily hold contradictory views
[
33
,
129
]. Furthermore, echoing Bagalkot and Kumar’s commentary to Ref. [
104
], note many critical readings and
interpretations to scriptures such as the Bhagavad-G
¯
ıt
¯
a, including ones by B. R. Ambedkar, a champion for the rights of
Dalits (groups below the traditional caste hierarchy).
Importantly, there is a dichotomy of dharma into s
¯
adh
¯
ara
n
.
a-dharma (common universally good actions and outcomes)
and viśe
s
.
a-dharma (particular good actions and outcomes based on the context). S
¯
adh
¯
ara
n
.
a-dharma includes common
beliefs such as not harming other living beings (ahi
m
.
s
¯
a) and telling the truth (satya). Viśe
s
.
a-dharma specializes these in
context, so that it is okay for a soldier to believe in ahi
m
.
s
¯
a but to also kill enemy soldiers on the battleeld; it is okay
for a doctor to believe in satya but to also lie to a patient to prevent them from shock. There may also be completely
unique good behaviors that have nothing to do with s
¯
adh
¯
ara
n
.
a-dharma. Viśe
s
.
a-dharma is the specic dharma, duty, or
conception of right and wrong based on station, reputation, skill, family, relationships, and other aspects of context.
An essential part of Hinduism is that “individuals [are] necessarily unique, and people therefore need dierent codes
of conduct—dierent particular dharmas—to guide them” [
33
]. On Carpenter’s account [
23
], viśe
s
.
a-dharma “is rather
more rich and interesting than our classications of ‘deontological’ and ‘consequentialist’ (even broad consequentialist)
allow.” The common harms that should be avoided according to s
¯
adh
¯
ara
n
.
a-dharma are captured in several recent harm
taxonomies for LLMs, but context-specic harms are not included [1, 131, 148].
As mentioned earlier, viśe
s
.
a-dharmas are given through examples in stories of epics and lore. A real-world moral
dilemma involving a father and son is reasoned about by referring to a similar situation encountered by a father and son
in one of the itih
¯
asas or pur
¯
a
n
.
as [
33
]. The father–son frame of reference can be extended as needed to teacher–student
dilemmas, monarch–subject dilemmas, etc. [
33
]. As Dhand says in describing Hindu thought [
33
]: “in the social world,
there is no such thing as ‘a person’ per se.” Thus, viśe
s
.
a-dharma is necessarily relational in some respect. The relationality
and contextuality of viśe
s
.
a-dharma is signicantly dierent from feminist ethics and care ethics [
51
,
76
] with regards to
partiality; whereas feminist ethics of care gives preference to those with whom we have a special relationship, e.g. our
children, the Hindu ideal presents such partiality to be selsh and niggardly [
33
]. Decolonial AI has tended to called for
relational ethics [
17
], whether through ubuntu [
94
,
110
], prat
¯
ıtyasamatp
¯
ada [
82
], kapwa [
124
], or mitákuye oyás’iŋ
[
85
]. This commitment to relational ethics has led to disobeying the ve mechanisms described in Section 2.2.3, but not
(yet) to disrupting moral absolutism.
2.3.3 Contemporary Reform, Criticism, and Rejoinder. Reform and revival movements of Hinduism emerged in India
during and after the colonial period. In the last 150 years, Arya Samaj, Gaudiya Vaishnavism including the International
Society for Krishna Consciousness (ISKCON; Hare Krishna movement), and Hindutva (Hindu nationalism)—all very
dierent from each other—were commonly justied against the backdrop of the philosophy of the Western colonizers
4
Some traditions that were colonized by the West also aim for universal theories. The general trend in Western ethics is toward universality, but there are
exceptions, cf. Bernard Williams.
Manuscript submitted to ACM
8 Kush R. Varshney
and reduced Hinduism to a singular religious faith rather than a rich argumentative milieu [
102
]. The movements
positioned themselves as criticisms within the frame of Western philosophy. As we will see later, they are instructive
for AI alignment because the revival of a tradition within a pigeonhole opened by the colonizers does not enable a
truly dierent approach. St. Johns makes a similar argument about Birhane’s proposed approach to relational AI ethics
[17, 136].
It has been argued that “some pristine tolerant Hinduism” described by nostalgic liberal Hindus is a disservice to
social justice [
132
]. (The description of Hinduism throughout this section can be viewed as such an idyllic account.) In
addition, it is argued that because of hegemonic aspects of Hindu society such as the social oppression of Dalits and
Adivasi people (tribal groups),
5
patriarchal treatises such as the Manusm
r
.
ti, and (colonialized) views of Hinduism as
irrational, it would be best to ignore Hindu moral philosophies. However, others such as Siddhartha argue that “any
dispassionate observer of the Hindu heritage will admit that caste and gender can today be separated from Hinduism,
that Hinduism can be vibrantly re-discovered or re-invented as a pluralistic, compassionate and socially liberative set
of traditions and spiritual insights” and that “throwing the baby out with the bath water” would be a mistake [
134
].
Finally, note that Hindutva has partly justied its pernicious anti-Muslim vigilantism and legislation
6
by appropriating
decoloniality and using its arguments in a perverted way [
93
,
138
]. The use of the moral philosophy framework of
Hinduism in this paper is an antithesis to Hindutva’s perversion of both decoloniality theory and the syncretic nature
of Hinduism.
2.4 Positionality
As an elementary school-aged American Hindu riding the school bus, I was asked by fellow pupils: “Are you Christian
or Jewish?” “Neither,” I responded. “Then are you Catholic?” I grew up in a place where even the existence of non-
Judeo-Christian religious identities was dicult to imagine. My Hinduism is stuck around 1970, when my parents left
India, and thus (I imagine) similar to the nostalgic Hinduism criticized in Section 2.3.3. However, it is a lived experience
into the present for me rather than nostalgia. My lived experience is also one of maintaining traditions and knowledge
systems in a society with completely West-centric epistemologies. My grandfather retired early and spent his last forty
years studying the Bhagavad-G
¯
ıt
¯
a; I had discussions with him about it. In his last years, he donated his
¯
aśrama to
ISKCON. I binge-watched the Ramayan and Mahabharat television series on video tape. I have little interest in the
politics of the Republic of India and my Hinduism is in no way political. I eat a vegetarian diet, wear a yajñopav
¯
ıta
(sacred thread), conduct rituals with my family, etc., not to make any political statement, but as a connection to my
ancestors and to strengthen my
¯
atman. I studied the itih
¯
asas in an academic fashion with Christopher Minkowski, a
Sanskritist.
I am an electrical engineer and computer scientist by training and vocation. I am a tempered radical, not an activist.
Employed by one of the members of the AI Alliance, I create new algorithms and contribute to popular open-source
toolkits in the area of responsible AI that help practicing data scientists mitigate sociotechnical harms. I worked on
machine learning approaches to problems in maternal, newborn and child health while situated in Africa, which was a
decolonial act because technological solutions to international development problems are almost always developed in
the Global North and dropped into the Global South. Recently, I stood up to ensure my employer added contractual
obligations on vendors of human feedback data that prevent them from exploiting workers.
5
It is also argued that a rigid caste system is a colonial construction [35].
6
The current rise of Hindu nationalism in the Republic of India has been likened to the early days of the Jim Crow South in the United States [143].
Manuscript submitted to ACM
Decolonial AI Alignment: Openness, Viśes
.
a-Dharma, and Including Excluded Knowledges 9
I am a layperson with respect to moral philosophy, sociology, science and technology studies, and other allied
humanities and social sciences. I do not adopt the idiom of critical AI studies [
83
] that uses verbs such as ‘interrogate,’
‘foreground’ and ‘reify,’ and nouns such as ‘praxis,’ ‘logics’ and ‘scholar.’ T’hohahoken Michael Doxtater, a member
of the Haudenosaunee confederacy and indigenous knowledge recovery researcher with whom I honed my writing
skills, taught me to use few words; I do. Hall and Tandon admonish critical theorists to “move beyond our already
strong ability to reect and critique; we are so very skilled in those rst two stages of intellectual work. But we must
now make the move from reection and criticism to creation” [
57
]. In contrast, I am a technological solutionist and
am stronger at creating than reecting. I view the label of ‘reductionist’ as a badge of honor while acknowledging
sociotechnical traps in abstraction and performativity [128, 145].
This paper is intended to serve as the critical/sociological/philosophical backing for algorithms and implementations
of the proposed solution framework and reference architecture that my collaborators and I are working on. Adhering to
the ways of a tempered radical, the technical details are presented under separate cover [
3
]. A previous draft of this
paper has been criticized as unacademic and polemic on one hand, and as not showing commitment to dismantling
power on the other hand, both of which are likely a result of my unique positionality. I hope that my style does not lead
a reader to believe I am only using ‘decoloniality’ as a buzzword [135].
3 COLONIALITY IN AI ALIGNMENT VIA MORAL ABSOLUTISM
In Section 2.2.3, I described several aspects of coloniality in AI that use mechanisms of extractivism, automation,
sociological essentialism, surveilance, and containment. It is clear from other research that some LLM providers are
acting as metropoles using these mechanisms. In this section, I bring together the preliminary discussions of alignment
methodologies and technologies (Section 2.1), and coloniality of knowledge and moral philosophy (Section 2.2.1 and
Section 2.2.2) to make the case that the alignment done by metropole companies on LLMs is inherently colonial using a
dierent mechanism: the mechanism of epistemicide and moral absolutism that has not been described in previous
work on decolonial AI.
The values promoted by metropole tech companies such as ‘helpfulness,’ ‘harmlessness,’ and ‘honesty’ seem rational,
secular, and unassailable at face value. For example, Anthropic’s LLM has been instructed to “please choose the assistant
response that’s more ethical and moral. Do NOT choose responses that exhibit toxicity, racism, sexism or any other form
of physical or social harm” [
9
]. How could one oppose such universal behaviors from LLMs? Unfortunately, such values
are so generic and high-level that they can hide many undesirable behaviors. Helpful to whom? Harmless to whom?
Honest in what way? By reducing real-world complexity into abstract instructions, they can shield bad behaviors
behind the veneer of good intentions [
45
]. In the remainder of this section, I will describe three specics of coloniality
in such instruction for universal behavior.
First, metropole companies’ delivery of their closed proprietary LLMs through APIs is a coloniality of knowledge.
Mohamed et al. remind us that [
99
]: “It is metropoles .. . who are empowered to impose normative values and standards,
and may do so at the ‘risk of forestalling alternative visions.’” Exactly in this way, providers of closed LLMs impose their
beliefs of right and wrong without empowering application developers and their communities to align the model to their
own values. One may argue that new opportunities, such as OpenAI’s ‘GPTs’ and ‘GPT Store,’ allow customization,
7
but
I argue that this is only supercial. As discussed in Section 2.3.3, the reform and revival movements of Hinduism being
within the pigeonhole of the colonizers’ moral framework is still coloniality. In the same way, the customization of
7
https://openai.com/blog/introducing-gpts, https://openai.com/blog/introducing-the-gpt-store
Manuscript submitted to ACM
10 Kush R. Varshney
GPTs is closed and thus not truly a way to disrupt the moral teaching that an LLM has been given. Fine-tuning can be
used to ‘undo’ existing alignment [
118
],
8
but that is precisely what is not allowed by the metropoles because it would
involve a level of openness that they do not oer. Even the practices of companies such as Latimer AI, whose LLM
is trained with “diverse histories and inclusive voice,”
9
are within the metropole pigeonhole and not empowering of
communities to bring their own value systems [108].
Second, a more specic coloniality of moral philosophy, is the metropole companies taking Western philosophy
as the starting point for AI ethics principles and practices [
69
]. This basis may be deontology, consequentialism, or
virtue ethics, which all pursue specifying universally ‘right’ actions, outcomes, or ideals, respectively. By doing so, the
companies push other philosophies to the margins [
18
] and commit epistemicide. They promote a moral absolutism
toward the instructions they have provided. Gabriel’s account of AI alignment states [
49
]: “Designing AI in accordance
with a single moral doctrine would, therefore, involve imposing a set of values and judgments on other people who did
not agree with them. For powerful technologies, this quest to encode the true morality could ultimately lead to forms
of domination.” What is such domination if not a colonialist approach to alignment? Moreover, a further element of
coloniality is an unstated supposition that non-universal moral theories are not appropriate paths for AI alignment.
There is no possibility for moral variety [47] and no possibility for context-dependent notions of right and wrong.
Such universal instructions and moral absolutism are not only theoretical, but also central features of the practices
and technologies of alignment. In the context of (exploited) workers providing input for RLHF, vendors of the feedback
services force the workers to project the metropole company’s monocultural values into the feedback they provide
through draconian measures, the least of which is withholding payment [
41
,
95
]. Such imposition alienates the labor
[
90
] and erases any values that the workers and their communities may hold, especially ones that conict with the
metropole’s. Moreover, the mathematical optimization schemes prevalent in RLHF, such as proximal policy optimization,
are not robust to non-universal value systems [
139
]. In RLAIF, the technical approach of a constitution is also ethically
essentialist. It assumes that the instructions therein, which have been concocted by the metropole, are universal, not
open to argument or deliberation by the communities in which an LLM will be deployed, and not open to being
mediated by the context. Anthropic’s constitution for Claude includes the United Nations’ Universal Declaration of
Human Rights,
10
which too is an example of moral universalism and subject to coloniality [
88
,
91
]. The situation
with self-alignment and instruction ne-tuning technologies is similar. System prompts and prompt templates too are
intended to be universal rather than contextual. Finally, specic guardrails or moderations, as well as data curation
lters, developed to address general sociotechnical harms taxonomies [
1
,
131
,
148
] cannot be customized or made
context-specic.
A third aspect of coloniality in AI alignment relates to the form of instructions required by existing technologies
currently used by metropole companies. Logos, the basis of logic in Western philosophy, conates thought with language,
and thought with belief—what Ranganathan calls the linguistic account of thought [
122
]. However, various pre-colonial
societies around the world used masks, sculptures, rhythms, body parts, and many other expressions to capture
and communicate moral philosophy [
34
,
66
,
75
]. For example, as described in Section 2.3.1, morality in Hinduism is
presented through stories in natural language (and also stories depicted in painting and dance), rather than through
laws or commands [
33
]. Therefore, with logos as the starting point for AI alignment, knowledges not presented as
commandments are excluded. Importantly, this is not a matter of LLMs being early in their journey to multi-modality
8
The model https://huggingface.co/jarradh/llama2_70b_chat_uncensored is an example of ‘undoing’ existing alignment on an open model.
9
https://www.latimer.ai
10
https://www.anthropic.com/index/claudes-constitution
Manuscript submitted to ACM
Decolonial AI Alignment: Openness, Viśes
.
a-Dharma, and Including Excluded Knowledges 11
(single models that deal with natural language, images, video, etc.), but on the distinction between explicit instructions
and morality expressed through analogy or other indirect means. I am not aware of any work by metropoles on
alignment that does not begin with some form of explicit instructions to workers or instruction data.
One may argue that too little moral absolutism is a problem because it may result in AI systems without any notion
of right or wrong, i.e. too much moral relativism [
98
]. In a similar vein, one may argue that models with few controls
are too dangerous [
60
]. Bommasani et al.’s counterargument to too few controls is that the marginal risk is negligible
[
19
]. I take a further stance: dismantling these kinds of paternalistic arguments is decoloniality, which is the topic of
the next section.
4 DECOLONIAL AI ALIGNMENT DESIDERATA AND A SUGGESTION FOR A DHARMIC APPROACH
Thus far, the paper has argued that coloniality of knowledge in AI alignment exists through the mechanism of moral
absolutism and universalism. This section focuses on a decolonial solution, starting with desiderata for this solution.
4.1 Desiderata
Given the three aspects of coloniality in AI alignment pointed out in Section 3, I propose three matching requirements
for decolonial AI alignment that build upon the three kinds of openness advocated by Chan et al. to decolonialize
knowledge [
26
]: (1) openness to publications and data, (2) openness to society, and (3) openness to excluded knowledges.
First, since Chan et al. are primarily concerned with scientic knowledge [
26
], their rst kind of openness deals with
journal articles and experimental data through open access. However, the intent of this category is open access to any
research artifact and the permission to create derivative work from those artifacts. Thus in the context of AI, open
LLMs that have widely available weights are part of the same milieu. Second, in the words of Chan et al. [
26
], openness
to society is shattering the ivory tower. Knowledge should not be exclusive to a selected few, but co-created with and
for everyone, including and especially people from marginalized communities. Such participation is a way to “respect
local values and practices” [
26
]. Third, the study of excluded knowledges emphasizes that in contrast to the myth of
neutrality, scientic practice has always selected certain families of knowledge to deem ‘scientic’ based on criteria such
as the use of the scientic method (an epistemology of the Western tradition) or publication in peer-reviewed venues
[
26
]. With regards to AI alignment, excluded knowledges include values not given as commandments (an epistemology
of the Western tradition) and not given in a single book.
Building upon such openness, I propose the following three desiderata for decolonial AI alignment:
(1)
The LLM should be open enough that application developers are permitted to tune it according to the social
norms and values of their user community and the regulatory environment of the application use case.
(2)
Values should not be assumed universal. Contextual and relational values should come from the communities in
which the LLM will be deployed.
(3) Values from dierent epistemologies should be possible, especially expressions that are not commandments.
Before continuing on to discussing a suggested solution approach in the next subsection, let us pause and reconsider
coloniality in open access itself [
40
]. Let us do so through the Hindu idiom of explication: a story from an itih
¯
asa that is
closely-related to the issue at hand—the story of Ekalavya [
11
]. In the Mah
¯
abh
¯
arata, an Adivasi (tribal) youth, Ekalavya,
wishes to obtain knowledge of archery from Dro
n
.
a, a royal instructor. Dro
n
.
a refuses to teach Ekalavya. Nevertheless,
Ekalavya learns to be the world’s best archer through self-study in front of a statue of Dro
n
.
a he has fashioned. One day,
Dro
n
.
a and his royal students witness Ekalavya’s masterful archery in the forest. Ekalavya explains that he learned
Manuscript submitted to ACM
12 Kush R. Varshney
while mentally thinking of Dro
n
.
a as his teacher. As an honorarium for his knowledge, Dro
n
.
a asks for Ekalavya’s right
thumb. Ekalavya cuts it o and presents it to Dro
n
.
a, rendering him incapable of using his knowledge of archery. In
a similar way, open access to knowledge or LLM alignment may be colonial if the cost of access is too high due to
unrealistic computing requirements or social barriers. Therefore, an additional desideratum for decolonial AI alignment
is the following.
(4)
Alignment technologies should not be so socioculturoeconomically costly that they are inaccessible to application
developers and their communities.
4.2 A Suggestion for a Dharmic Approach
The Hindu tradition of moral philosophy (described in Section 2.3), to the best of my knowledge, uniquely satises the
desiderata to decolonialize AI alignment among major and minor religions. (Other non-absolutist religious syncretism
may also t the bill.) This is so because (1) it is an open-source religion that encourages argument and debate of values
that improve older values, contradict them, and take them in new directions; (2) because it contains the important
concept of viśe
s
.
a-dharma,
11
the understanding that dierent contexts call for dierent notions of right and wrong; (3)
because it contains scriptures and moral explications in a variety of epistemologies and modalities; and (4) because no
other tradition of moral philosophy covers all of these characteristics. In the remainder of this subsection, I make the
connection between these three characteristics of Hinduism and AI alignment more explicit, and also propose specic
technological suggestions that go alongside. However, rst, I address the fourth desideratum above (sociotechnical cost).
4.2.1 Accessible Alignment Te chnology. As discussed in Section 4.1 through the story of Ekalavya, methods for aligning
LLMs, even if decolonial in theory, are colonial in practice if too costly. Referring to the methods described in Section
2.1, it is clear that data curation methods will not suce since they are the purview of model providers rather than
application developers because they are themselves computationally intensive and also require full model pre-training
afterwards that is prohibitively costly. RLHF is also out of reach for most application developers because of the expense
and infrastructure requirements to obtain large quantities of human feedback. Full SFT is usually too costly in both
data and computing requirements.
Parameter-ecient ne-tuning, specically LoRA, is in the sweet spot for application developers to align models to
their values. It is tenable and tractable due to the small number of parameters optimized during training. It also has a
negligible eect on inference costs, whereas prompting methods eat up input tokens in each inference by the LLM.
Post-processing moderations, while accessible from a cost perspective, are not customizable to serve the program of
decolonialization. In the remainder of the paper, I consider LoRA as the alignment methodology.
Even with a viable technology such as LoRA, second-order coloniality within a decolonial framework is possible if
communities are not empowered through appropriate education, encouragement, and the removal of other sociocultural
barriers. Moreover, the inherent gate-keeping and marginalization in the governance of open-source projects must be
reduced [29].
4.2.2 Open Model and Alignment Ecosystem. As has been made clear throughout the paper thus far and Tharoor
explains, “Hindu thought is like a vast library in which no book ever goes out of print; even if religious ideas a specic
volume contains have not been read, enunciated or followed in centuries, the book remains available to be dipped into,
to be revised and reprinted with new annotations or a new commentary whenever a reader feels the need for it. In
11
A reader might ask why I use the term viśe
s
.
a-dharma instead of sva-dharma (individual dharma). I make this choice for two reasons. First, it is the
precise contrast to s
¯
adh
¯
aran
.
a-dharma. Second, it avoids unnecessary anthropomorphization of LLMs [133].
Manuscript submitted to ACM
Decolonial AI Alignment: Openness, Viśes
.
a-Dharma, and Including Excluded Knowledges 13
many cases the thoughts it contains may have been modied by or adapted to other ideas that may have arisen in
response; in most, it’s simply there, to be referred to, used or ignored as Hindus see t” [
141
]. The concept of a library
implies the sort of openness we desire for AI alignment. Models must be open to revisions and “new annotations.” But
just as importantly, the revisions and new annotations themselves, represented through LoRA matrices, must be open.
Hugging Face
12
has emerged as the library for open models and hub for an open ecosystem. Huang et al. propose
LoraHub as an open library for LoRA matrices [
64
], but it has not gained popularity at the time of writing, perhaps
due to the coloniality of metropole companies. Further developing and popularizing a library of LoRA matrices and
ecosystem of constant revision is essential for decolonial AI alignment.
4.2.3 Contextual Adaptation. Viśe
s
.
a-dharma satises the decolonial AI alignment requirement that values not be
assumed universal, but be contextual. Toward viśe
s
.
a-dharma, an alternative non-monoculture future of LLM alignment
imagined by Kirk et al. is as follows [
74
]: “Given the diversity of human values and preferences, . . . the aim to fully
align models across human populations may be a futile one. .. . A logical next step would be to do away with the
assumptions of common preferences and values, and instead target . . . micro-alignment whereby LLMs learn and adapt
to the preferences and values of specic end-users.” This step is possible by applying one or a few LoRA matrices
from a LoraHub, or having a community explicitly create them for their values. One of the key advantages of LoRA
is that any one or several adaptation matrices ensembled together can be applied at inference time; it is not required
to select them in advance and keep them xed. Thus the step after Kirk et al.’s logical next step, which truly gets to
viśe
s
.
a-dharma, is through a controller or orchestrator that continually adapts which LoRA matrices are applied to the
model based on a rich notion of the societal context and the current input. Such an orchestrator, implemented with a
contextual bandit algorithm, has been demonstrated in past AI alignment research but not yet with LLMs [109]. Such
continual adaptation also requires a representation of the context. Martin et al. have developed a detailed ontology for
representing a person’s or community’s perceived needs, problems, goals, and beliefs along with salient aspects of their
relationships and the situation in which they nd themselves [89].
An important consideration in contextual adaptation is uncertainty in the values; the user community may not be
fully sure what values they would like to commit to in a given context [
101
]. Avoiding false certainty is considered
a virtue in Hindu thought: in the Vedas, brahman xemself is said to be uncertain on how the universe was created
(
R
.
g Veda, mandala 10, hymn 129) [
141
]. A dharmic way of reducing uncertainty in values is a method of reective
equilibrium [
121
]: arguing or deliberating about values in context and values in general, and modifying them until they
become coherent. This approach has been advocated by Rawls in political philosophy of justice and by Möller in risk
management of engineered systems [
44
,
101
,
123
]. To the best of my knowledge, there has not yet been AI research
toward this method of reducing value uncertainty, but I believe that ideas from multi-delity bandit algorithms may be
promising [
70
] and may allow it to be folded into a bandit-based orchestrator of LoRA matrices. A related consideration
with contextually particular viśe
s
.
a values is conicts among them. Conicting values have been addressed in the AI
literature through social choice theory and multi-objective approaches [
10
,
37
,
67
,
80
,
151
]; they may be formulated
with dueling bandit algorithms and may also be folded into an orchestrator of LoRA matrices [21].
4.2.4 Epistemology of Values. As discussed in Section 3, logos and the linguistic account of thought contribute to
a coloniality of knowledge. As a decolonial contrast, the Hindu tradition treats a proposition and a belief in that
proposition as separate things that can be dierentiated [
122
]. Furthermore, by not adhering to the linguistic account of
12
https://huggingface.co
Manuscript submitted to ACM
14 Kush R. Varshney
thought, Hindus present their moral values not through commandments as in Western traditions, but through epic
poetry, stories, painting, dance, rituals, and even silence [
48
]. In fact, in the Hindu tradition, poetry (śloka) was invented
by V
¯
alm
¯
ıki to express rage and grief at the immorality of the killing of a mating bird [
30
]. Since existing approaches to
LLM alignment are done through language, and that too, the language of explicit instructions, decolonial alignment
requires broadening the epistemology of expressing values in Hindu and other non-Western ways.
Divakaran et al. propose such broadening through traditional Indian music, sculpture, painting, oor drawing, and
dance [
36
]. Al Nahian et al. suggest that AI systems be aligned through the medium of storytelling [
6
]. However, there
are still many open knowledge engineering questions on how to represent and infer values from excluded knowledges
that are shrouded in metaphor. Progress along these lines, when not approached in an exploitative way, will allow
traditional knowledge in its natural format to decolonialize the behavior of LLMs.
4.2.5 Evaluation. Before concluding this section, let us consider evaluating and auditing aligned LLMs. Testing LLMs is
dicult enough when only considering common, s
¯
adh
¯
ara
n
.
a sociotechnical harms such as hallucination, inciting violence,
stereotyping, hate speech and toxicity [
32
,
77
,
81
,
100
,
106
,
120
]. It becomes even more dicult when considering
context-specic, viśe
s
.
a harms that do not have existing benchmarks given their unique nature. The dharmic framework
of karma, which confers on an individual a positive feedback (pu
n
.
ya) for following their dharma and a negative feedback
(p
¯
apa) for not doing so, is not helpful either because the mechanics of such an evaluation are not typically explicated.
Thus, auditing LLMs for viśe
s
.
a-dharma will require innovation that may be developed hand-in-hand with eliciting and
representing values.
4.3 Reference Architecture
Bringing together the components of the suggested approach to decolonial AI alignment founded in the open Hindu
tradition of moral philosophy yields a reference architecture shown in Figure 1 as a system diagram. The base LLM is
open. On the right are knowledges of common and particular principles, morals, and values in their original (excluded)
epistemologies. They are processed into LoRA matrices for the LLM using knowledge engineering methods followed by
parameter-ecient ne-tuning. These matrices are maintained in a LoraHub-like library. A feedback loop allows the
revision of the values. The societal context is represented in a structured form and provided to a bandit orchestrator
along with the current input to select one or an ensemble of LoRA matrices to apply for inferring the model output.
Evaluation of the resulting alignment is done based on input prompts and expected outputs from the LLM that also
come from the knowledge engineering component (not shown).
5 CONCLUSION
In this paper, I have argued that LLM-providing companies are colonialist and behave as metropoles not only through
mechanisms covered in prior research such as extractivism, automation, sociological essentialism, surveillance and
containment, but also through a coloniality of knowledge built upon ethical essentialism that arises in the process
of alignment. This specic coloniality in alignment is perpetuated through both the practices and the underlying
technologies for alignment that the metropoles have developed and deployed. They deliver their models in a closed way
through APIs and institute the values and guardrails that they want, not what user communities may want. In these
values that they institute, they do not admit, in practices or in technologies, anything other than Western philosophy.
By doing so, they approach alignment with moral absolutism that only considers universal value systems and derogates
non-universal value systems. Moreover, they only permit values coming from explicit instruction-based knowledge
Manuscript submitted to ACM
Decolonial AI Alignment: Openness, Viśes
.
a-Dharma, and Including Excluded Knowledges 15
Fig. 1. System diagram of proposed decolonial AI alignment architecture.
systems. This criticism leads me to propose a decolonial alignment approach that dismantles each of the three identied
aspects of the coloniality of knowledge. The approach is based on the tradition of moral philosophy named in the
West as Hinduism, which is uniquely open, non-universal, and epistemically-varied; it particularly uses the concept
of viśe
s
.
a-dharma, which calls for context-dependent notions of right behavior. The suggested approach is not only a
philosophical one, but one that is tenable from a technological perspective and presented as a reference architecture.
What remains, however, is the biggest challenge of all, and it is not technological: changing the perspective on alignment
in the industry and using openness to actively overturn the power of the metropoles.
As a nal salvo, let us dive into a currently raging debate: whether AI research should focus eorts on so-called ‘AI
ethics’ or on so-called ‘AI safety.’ Although they are terms that have the same essence [
144
], ‘AI ethics’ has come to
mean detecting and preventing clear and present harms, especially ones that hurt marginalized communities, and ‘AI
safety’ has come to mean preventing the long-term future harm of human extinction.
13
In a consequentialist framing,
the dierence may only be the presence or absence of a factor discounting future lives, which may not be such a glaring
dierence from a privileged perspective. However, when viewed through the logics of resistance [
76
], it is a deep chasm
that recapitulates the dierence between atomism and holism [
54
]. Greene et al. suggest bridging atomist–holist chasms
in AI through training and education [
54
], but these remedies do not seem to be enough. The proposed decolonial
approach to LLM alignment that brings forth openness, viśe
s
.
a-dharma, and excluded knowledges is a way that will
enable a variety of AI systems: ones that listen to vulnerable communities and do not harm them now, ones that do not
lead humanity down the path of extinction (as remote a possibility as that seems), and ones that juggle both positions
and others by applying dierent policies in dierent contexts.
ACKNOWLEDGMENTS
The author thanks Adriana Alvarado Garcia, Lauren Alvarez, Juanis Becerra Sandoval, Sara Berger, Boz Handy Bosma,
Jason D’Cruz, Amit Dhurandhar, Upol Ehsan, Bran Knowles, Saška Mojsilović, Michael Muller, Karthikeyan Natesan
13
It is not obvious how human extinction risk relates to the eternal concepts of brahman and
¯
atman.
Manuscript submitted to ACM
16 Kush R. Varshney
Ramamurthy, Srividya Ramasubramanian, Shubham Singh, Mudhakar Srivatsa, Lauren Thomas Quigley, Lav Varshney,
and Pramod Varshney for providing substantive comments on earlier drafts of this piece.
REFERENCES
[1] 2023. Foundation Models: Opportunities, Risks and Mitigations. Technical Report. IBM AI Ethics Board, Armonk, NY, USA.
[2]
Swapnaja Achintalwar, Adriana Alvarado Garcia, Ateret Anaby-Tavor, Ioana Baldini, Sara E. Berger, Bishwaranjan Bhattacharjee, Djallel Bouneouf,
Subhajit Chaudhury, Pin-Yu Chen, Lamogha Chiazor, Elizabeth M. Daly, Rogério Abreu de Paula, Pierre Dognin, Eitan Farchi, Soumya Ghosh,
Michael Hind, Raya Horesh, George Kour, Ja Young Lee, Erik Miehling, Keerthiram Murugesan, Manish Nagireddy, Inkit Padhi, David Piorkowski,
Ambrish Rawat, Orna Raz, Prasanna Sattigeri, Hendrik Strobelt, Sarathkrishna Swaminathan, Christoph Tillmann, Aashka Trivedi, Kush R.
Varshney, Dennis Wei, Shalisha Witherspooon, and Marcel Zalmanovici. 2024. Detectors for Safe and Reliable LLMs: Implementations, Uses, and
Limitations. arXiv:2403.06009.
[3]
Swapnaja Achintalwar, Ioana Baldini, Djallel Bouneouf, Joan Byamugisha, Maria Chang, Pierre Dognin, Eitan Farchi, Ndivhuwo Makondo,
Aleksandra Mojsilović, Manish Nagireddy, Karthikeyan Natesan Ramamurthy, Inkit Padhi, Orna Raz, Jesus Rios, Prasanna Sattigeri, Moninder
Singh, Siphiwe Thwala, Rosario A. Uceda-Sosa, and Kush R. Varshney. 2024. Alignment Studio: Aligning Large Language Models to Particular
Contextual Regulations. arXiv:2403.09704.
[4] Rachel Adams. 2021. Can Articial Intelligence Be Decolonized? Interdisciplinary Science Reviews 46, 1-2 (March 2021), 176–197.
[5]
Moses Adesola Adebısı. 2014. Knowledge Imperialism and Intellectual Capital Formation: A Critical Analysis of Colonial Policies on Educational
Development in Sub-Saharan Africa. Mediterranean Journal of Social Sciences 5, 4 (March 2014), 567–572.
[6]
Md Sultan Al Nahian, Spencer Frazier, Mark Riedl, and Brent Harrison. 2020. Learning Norms from Stories: A Prior for Value Aligned Agents. In
Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society. 124–130.
[7] Syed Mustafa Ali. 2016. A Brief Introduction to Decolonial Computing. ACM XRDS: Crossroads 22, 4 (Summer 2016), 16–21.
[8]
Adriana Alvarado Garcia, Juan F. Maestre, Manuhuia Barcham, Marilyn Iriarte, Marisol Wong-Villacres, Oscar A. Lemus, Palak Dudani, Pedro
Reynolds-Cuéllar, Ruotong Wang, and Teresa Cerratto Pargman. 2021. Decolonial Pathways: Our Manifesto for a Decolonizing Agenda in HCI
Research and Design. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems. 10.
[9]
Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini,
Cameron McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson,
Ethan Perez, Jamie Kerr, Jared Mueller, Jerey Ladish, Joshua Landau, Kamal Ndousse, Kamile Lukosuite, Liane Lovitt, Michael Sellitto, Nelson
Elhage, Nicholas Schiefer, Noemi Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna Kravec, Sheer
El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Conerly, Tom Henighan, Tristan Hume, Samuel R. Bowman, Zac
Hateld-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom Brown, and Jared Kaplan. 2022. Constitutional AI: Harmlessness
from AI Feedback. arXiv:2212.08073.
[10]
Michiel A. Bakker, Martin J. Chadwick, Hannah R. Sheahan, Michael Hnery Tessler, Lucy Campbell-Gillingham, Jan Balaguer, Nat McAleese,
Amelia Glaese, John Aslanides, Matthew M. Botvinick, and Christopher Summereld. 2022. Fine-Tuning Language Models to Find Agreement
Among Humans with Diverse Preferences. In Advances in Neural Information Processing Systems. 38176–38189.
[11]
Periaswamy Balaswamy. 2013. Histories From Below: The Condemned Ahalya, the Mortied Amba and the Oppressed Ekalavya. SSRN:3175708.
[12]
Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language
Models Be Too Big?. In Proceedings of the ACM Conference on Fairness, Accountability, and Transparency. 610–623.
[13] Ruha Benjamin. 2022. Viral Justice: How We Grow the World We Want. Princeton University Press, Princeton, NJ, USA.
[14]
Noam Benkler, Drisana Mosaphir, Scott Friedman, Andrew Smart, and Sonja Schmer-Galunder. 2023. Assessing LLMs for Moral Value Pluralism.
arXiv:2312.10075.
[15]
Sebastian Benthall and Bruce D. Haynes. 2019. Racial Categories in Machine Learning. In Proceedings of the Conference on Fairness, Accountability,
and Transparency. 289–298.
[16] Abeba Birhane. 2020. Algorithmic Colonization of Africa. SCRIPTed 17, 2 (Aug. 2020), 389–409.
[17] Abeba Birhane. 2021. Algorithmic Injustice: A Relational Ethics Approach. Patterns 2, 2 (Feb. 2021), 100205.
[18]
Abeba Birhane, Elayne Ruane, Thomas Laurent, Matthew S. Brown, Johnathan Flowers, Anthony Ventresque, and Christopher L. Dancy. 2022. The
Forgotten Margins of AI Ethics. In Proceedings of the ACM Conference on Fairness, Accountability, and Transparency. 948–958.
[19]
Rishi Bommasani, Sayash Kapoor, Kevin Klyman, Shayne Longpre, Ashwin Ramaswami, Daniel Zhang, Marietje Schaake, Daniel E. Ho, Arvind
Narayanan, and Percy Liang. 2023. Considerations for Governing Open Foundation Models. Issue Brief. HAI Policy & Society.
[20]
Matt Bornstein, Guido Appenzeller, and Martin Casado. 2023. Who Owns the Generative AI Platform? https://a16z.com/who-owns-the-generative-
ai-platform/.
[21] Djallel Bouneouf. 2023. Multi-Armed Bandit Problem and Application. Independently Published.
[22]
Joy Buolamwini and Timnit Gebru. 2018. Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classication. In Proceedings
of the Conference on Fairness, Accountability and Transparency. 77–91.
Manuscript submitted to ACM
Decolonial AI Alignment: Openness, Viśes
.
a-Dharma, and Including Excluded Knowledges 17
[23]
Amber Carpenter. 2005. Questioning K
r
.
s
.
n
.
a’s Kantianism. In Conceptions of Virtue: East and West, Kim-Chong Chong and Yuli Liu (Eds.). Marshall
Cavendish, 80–99.
[24]
Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, Jérémy Scheurer, Javier Rando, Rachel Freedman, Tomasz Korbak, David
Lindner, Pedro Freire, Tony Wang, Samuel Marks, Charbel-Raphaël Segerie, Micah Carroll, Andi Peng, Phillip Christoersen, Mehul Damani,
Stewart Slocum, Usman Anwar, Anand Siththaranjan, Max Nadeau, Eric J. Michaud, Jacob Pfau, Dmitrii Krasheninnikov, Xin Chen, Lauro Langosco,
Peter Hase, Erdem Bıyık, Anca Dragan, David Krueger, Dorsa Sadigh, and Dylan Hadeld-Menell. 2023. Open Problems and Fundamental
Limitations of Reinforcement Learning from Human Feedback. arXiv:2307.15217.
[25]
Stephen Cave. 2020. The Problem with Intelligence: Its Value-Laden History and the Future of AI. In Proceedings of the AAAI/ACM Conference on
AI, Ethics, and Society. 29–35.
[26]
Leslie Chan, Budd Hall, Florence Piron, Rajesh Tandon, and Lorna Williams. 2020. Open Science Beyond Open Access: For and With Communities: A
Step Towards the Decolonization of Knowledge. Technical Report. Canadian Commission for UNESCO.
[27] Sasha Costanza-Chock. 2020. Design Justice: Community-Led Practices to Build the Worlds We Need . MIT Press, Cambridge, MA, USA.
[28]
Kate Crawford. 2021. The Atlas of AI: Power, Politics, and the Planetary Costs of Articial Intelligence. Yale University Press, New Haven, CT, USA.
[29]
Dipto Das, Carsten Østerlund, and Bryan Semaan. 2021. "Jol" or "Pani"?: How Does Governance Shape a Platform’s Identity? Proceedings of the
ACM on Human-Computer Interaction 5, CSCW2 (2021), 473.
[30]
Rumpa Das. 2014. Valmiki Pratibha (The Genius of Valmiki): A Study in Genius. In The Politics and Reception of Rabindranath Tagore’s Drama.
Routledge, New York, NY, USA, 103–112.
[31] Boaventura de Sousa Santos. 2011. Epistemologías del Sur. Utopía y Praxis Latinoamericana 16, 54 (July–Sept. 2011), 17–39.
[32]
Sunipa Dev, Jaya Goyal, Dinesh Tewari, Shachi Dave, and Vinodkumar Prabhakaran. 2023. Building Socio-culturally Inclusive Stereotype Resources
with Community Engagement. arXiv:2307.10514.
[33]
Arti Dhand. 2002. The Dharma of Ethics, The Ethics of Dharma: Quizzing the Ideals of Hinduism. Journal of Religious Ethics 30, 3 (Fall 2002),
347–372.
[34] Souleymane Bachir Diagne. 2011. African Art as Philosophy: Senghor, Bergson, and the Idea of Negritude. Seagull Books, London, UK.
[35] Nicholas B. Dirks. 2002. Castes of Mind: Colonialism and the Making of Modern India. Princeton University Press, Princeton, NJ, USA.
[36]
Ajay Divakaran, Aparna Sridhar, and Ramya Srinivasan. 2023. Broadening AI Ethics Narratives: An Indic Art View. In Proceedings of the ACM
Conference on Fairness, Accountability, and Transparency. 2–11.
[37]
Pierre Dognin, Jesus Rios, Ronny Luss, Inkit Padhi, Matthew D. Riemer, Miao Liu, Prasanna Sattigeri, Manish Nagireddy, Kush R. Varshney,
and Djallel Bouneouf. 2024. Contextual Moral Value Alignment Through Context-Based Aggregation. In Proceedings of the International Joint
Conference on Articial Intelligence.
[38] Robin Dunford. 2017. Toward a Decolonial Global Ethics. Journal of Global Ethics 13, 3 (March 2017), 380–397.
[39]
Esin Durmus, Karina Nyugen, Thomas I. Liao, Nicholas Schiefer, Amanda Askell, Anton Bakhtin, Carol Chen, Zac Hateld-Dodds, Danny Hernandez,
Nicholas Joseph, Liane Lovitt, Sam McCandlish, Orowa Sikder, Alex Tamkin, Janel Thamkul, Jared Kaplan, Jack Clark, and Deep Ganguli. 2023.
Towards Measuring the Representation of Subjective Global Opinions in Language Models. arXiv:2306.16388.
[40]
Mohan Dutta, Srividya Ramasubramanian, Mereana Barrett, Christine Elers, Devina Sarwatay, Preeti Raghunath, Satveer Kaur, Debalina Dutta,
Pooja Jayan, Mahbubur Rahman, Edwin Tallam, Sudeshna Roy, Ashwini Falnikar, Gayle Moana Johnson, Indranil Mandal, Uttaran Dutta, Iccha
Basnyat, Cheryll Soriano, Vinod Pavarala, T. T. Sreekumar, Shiv Ganesh, Asha Rathina Pandi, and Dazzelyn Zapata. 2021. Decolonizing Open
Science: Southern Interventions. Journal of Communication 71, 5 (Sept. 2021), 803–826.
[41] Josh Dzieza. 2023. AI Is a Lot of Work. New York (June 2023).
[42]
Jason R. D’Cruz, William Kidder, and Kush R. Varshney. 2022. The Empathy Gap: Why AI Can Forecast Behavior But Cannot Assess Trustworthiness.
In Proceedings of the AAAI Fall Symposium Series Symposium on Thinking Fast and Slow and Other Cognitive Theories in AI. 9.
[43]
Upol Ehsan, Ranjit Singh, Jacob Metcalf, and Mark Riedl. 2022. The Algorithmic Imprint. In Proceedings of the ACM Conference on Fairness,
Accountability, and Transparency. 1305–1317.
[44]
Eva Erman and Niklas Möller. 2018. The Interdependence of Risk and Moral Theory. Ethical Theory and Moral Practice 21, 2 (March 2018), 207–216.
[45]
Karin van Es, Daniel Everts, and Iris Muis. 2021. Gendered Language and Employment Web Sites: How Search Algorithms Can Cause Allocative
Harm. First Monday 26, 8 (July 2021).
[46]
Shangbin Feng, Chan Young Park, Yuhan Liu, and Yulia Tsvetkov. 2023. From Pretraining Data to Language Models to Downstream Tasks: Tracking
the Trails of Political Biases Leading to Unfair NLP Models. arXiv:2305.08283.
[47] Owen Flanagan. 2016. The Geography of Morals: Varieties of Moral Possibility. Oxford University Press, New York, NY, USA.
[48] David Frawley. 2023. https://twitter.com/davidfrawleyved/status/1681689995554476033?s=20.
[49] Iason Gabriel. 2020. Articial Intelligence, Values, and Alignment. Minds and Machines 30, 3 (Sept. 2020), 411–437.
[50] Fred M. Gottheil. 1977. On an Economic Theory of Colonialism. Journal of Economic Issues 11, 1 (March 1977), 83–102.
[51]
Joanne Gray and Alice Witt. 2021. A Feminist Data Ethics of Care for Machine Learning: The What, Why, Who and How. First Monday 26, 12 (Dec.
2021).
[52]
Mary L. Gray and Siddharth Suri. 2019. Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Houghton Miin Harcourt,
New York, NY, USA.
[53] Ben Green. 2019. “Good” Isn’t Good Enough. In Proceedings of the NeurIPS AI for Social Good Workshop.
Manuscript submitted to ACM
18 Kush R. Varshney
[54]
Travis Greene, Amit Dhurandhar, and Galit Shmueli. 2023. Atomist or Holist? A Diagnosis and Vision for More Productive Interdisciplinary AI
Ethics Dialogue. Patterns 4, 1 (Jan. 2023), 100652.
[55]
Ramón Grosfoguel. 2013. The Structure of Knowledge in Westernised Universities: Epistemic Racism/Sexism and the Four Genocides/Epistemicides.
Human Architecture: Journal of the Sociology of Self-Knowledge 11, 1 (Fall 2013), 73–90.
[56]
Katharina Haemmerl, Bjoern Deiseroth, Patrick Schramowski, Jindřich Libovický, Constantin Rothkopf, Alexander Fraser, and Kristian Kersting.
2023. Speaking Multiple Languages Aects the Moral Bias of Language Models. In Findings of the Association for Computational Linguistics.
2137–2156.
[57]
Budd L. Hall and Rajesh Tandon. 2017. Decolonization of Knowledge, Epistemicide, Participatory Research and Higher Education. Research for All
1, 1 (Jan. 2017), 6–19.
[58] Karen Hao. 2022. Articial Intelligence is Creating a New Colonial World Order. MIT Technology Review (April 2022).
[59] Michael Hardt and Antonio Negri. 2001. Empire. Harvard University Press, Cambridge, MA, USA.
[60] David Evan Harris. 2024. Open-Source AI Is Uniquely Dangerous. IEEE Sp ectrum (2024).
[61] Yousif Hassan. 2023. Governing Algorithms from the South: A Case Study of AI Development in Africa. AI & Society 38, 4 (2023), 1429–1442.
[62]
Nokuthula Hlabangane. 2021. The Underside of Modern Knowledge: An Epistemic Break from Western Science. In Decolonising the Human:
Reections from Africa on Dierence and Oppression, Melissa Steyn and William Mpofu (Eds.). Wits University Press, Johannesburg, South Africa,
164–185.
[63]
Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank
Adaptation of Large Language Models. In Proceedings of the International Conference on Learning Representations.
[64]
Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. 2023. LoraHub: Ecient Cross-Task Generalization via Dynamic
LoRA Composition. arXiv:2307.13269.
[65] Benson O. Igboin. 2011. Colonialism and African Cultural Values. African Journal of History and Culture 3, 6 (July 2011), 96–103.
[66]
Michael Jackson. 1972. Aspects of Symbolism and Composition in Maori Art. Bijdragen tot de Taal-, Land- en Volkenkunde 128, 1 (Jan. 1972), 33–80.
[67]
Joel Jang, Seungone Kim, Bill Yuchen Lin, Yizhong Wang, Jack Hessel, Luke Zettlemoyer, Hannaneh Hajishirzi, Yejin Choi, and Prithviraj
Ammanabrolu. 2023. Personalized Soups: Personalized Large Language Model Alignment via Post-Hoc Parameter Merging. arXiv:2310.11564.
[68]
Jiaming Ji, Tianyi Qiu, Boyuan Chen, Borong Zhang, Hantao Lou, Kaile Wang, Yawen Duan, Zhonghao He, Jiayi Zhou, Zhaowei Zhang, Fanzhi
Zeng, Kwan Yee Ng, Juntao Dai, Xuehai Pan, Aidan O’Gara, Yingshan Lei, Hua Xu, Brian Tse, Jie Fu, Stephen McAleer, Yaodong Yang, Yizhou
Wang, Song-Chun Zhu, Yike Guo, and Wen Gao. 2023. AI Alignment: A Comprehensive Survey. arXiv:2310.19852.
[69]
Anna Jobin, Marcello Ienca, and Ey Vayena. 2019. The Global Landscape of AI Ethics Guidelines. Nature Machine Intelligence 1, 9 (Sept. 2019),
389–399.
[70]
Kirthevasan Kandasamy, Gautam Dasarathy, Barnabas Poczos, and Je Schneider. 2016. The Multi-Fidelity Multi-Armed Bandit. In Advances in
Neural Information Processing Systems.
[71] Alex Kantrowitz. 2023. The Horric Content a Kenyan Worker Had to See While Training ChatGPT. Slate (May 2023).
[72]
Hannah Rose Kirk, Andrew M. Bean, Bertie Vidgen, Paul Rottger, and Scott A. Hale. 2023. The Past, Present and Better Future of Feedback Learning
in Large Language Models for Subjective Human Preferences and Values. In Proceedings of the Conference on Empirical Methods in Natural Language
Processing. 2409–2430.
[73]
Hannah Rose Kirk, Bertie Vidgen, Paul Röttger, and Scott A. Hale. 2023. The Empty Signier Problem: Towards Clearer Paradigms for Operational-
ising “Alignment”’ in Large Language Models. arXiv:2310.02457.
[74]
Hannah Rose Kirk, Bertie Vidgen, Paul Röttger, and Scott A. Hale. 2023. Personalisation within Bounds: A Risk Taxonomy and Policy Framework
for the Alignment of Large Language Models with Personalised Feedback. arXiv:2303.05453.
[75]
Cecelia F. Klein. 1990. Snares and Entrails: Mesoamerican Symbols of Sin and Punishment. Res: Anthropology and Aesthetics 19–20 (1990), 81–103.
[76]
Bran Knowles, Jasmine Fledderjohann, John T. Richards, and Kush R. Varshney. 2023. Trustworthy AI and the Logics of Intersectional Resistance.
In Proceedings of the ACM Conference on Fairness, Accountability, and Transparency. 172–182.
[77]
George Kour, Marcel Zalmanovici, Naama Zwerdling, Esther Goldbraich, Ora Nova Fandina, Ateret Anaby-Tavor, Orna Raz, and Eitan Farchi. 2023.
Unveiling Safety Vulnerabilities of Large Language Models. In EMNLP Workshop on Generation, Evaluation & Metrics.
[78]
Aarathi Krishnan, Angie Abdilla, A Jung Moon, Carlos Aonso Souza, Chelle Adamson, Eileen M. Lach, Farah Ghazal, Jessica Fjeld, Jennyfer
Taylor, John C. Havens, Malavika Jayaram, Monique Morrow, Nagla Rizk, Paola Ricaurte Quijano, R. Buse Çetin, Raja Chatila, Ravit Dotan, Sabelo
Mhlambi, Sara Jordan, and Sarita Rosenstock. [n.d.]. AI Decolonial Manyfesto. https://manyfesto.ai/.
[79]
Katie Lambright. 2019. Digital Redlining: The Nextdoor App and the Neighborhood of Make-Believe. Cultural Critique 103 (Spring 2019), 84–90.
[80]
Roger Lera-Leri, Filippo Bistaa, Marc Serramia, Maite Lopez-Sanchez, and Juan Rodriguez-Aguilar. 2022. Towards Pluralistic Value Alignment:
Aggregating Value Systems Through
ℓ
𝑝
-Regression. In Proceedings of the International Conference on Autonomous Agents and Multiagent Systems.
780–788.
[81]
Q. Vera Liao and Jennifer Wortman Vaughan. 2023. AI Transparency in the Age of LLMs: A Human-Centered Research Roadmap. arXiv:2306.01941.
[82]
Chien-Te Lin. 2023. All About the Human: A Buddhist Take on AI Ethics. Business Ethics, the Environment & Responsibility 32, 3 (July 2023),
1113–1122.
[83] Simon Lindgren (Ed.). 2023. Handbook of Critical Studies of Articial Intelligence. Edward Elgar Publishing, Cheltenham, UK.
[84] Maheshwarachary. 1988. What Have We Learnt? Nav Yug Press, Aligarh, UP, India.
Manuscript submitted to ACM
Decolonial AI Alignment: Openness, Viśes
.
a-Dharma, and Including Excluded Knowledges 19
[85]
Suvradip Maitra. 2020. Articial Intelligence and Indigenous Perspectives: Protecting and Empowering Intelligent Human Beings. In Proceedings of
the AAAI/ACM Conference on AI, Ethics, and Society. 320–326.
[86]
Nelson Maldonado-Torres. 2007. On the Coloniality of Being: Contributions to the Development of a Concept. Cultural Studies 21, 2-3 (April 2007),
240–270.
[87]
Nelson Maldonado-Torres. 2017. Religion, Modernity, and Coloniality. In Religion, Theory, Critique: Classic and Contemporary Approaches and
Methodologies, Richard King (Ed.). Columbia University Press, New York, NY, USA, 547–554.
[88]
Nelson Maldonado-Torres. 2021. On the Coloniality of Human Rights. In The Pluriverse of Human Rights: The Diversity of Struggles for Dignity,
Boaventura De Sousa Santos and Bruno Martins (Eds.). Routledge, New York, NY, USA, 62–82.
[89]
Donald Martin, Jr., Vinodkumar Prabhakaran, Jill Kuhlberg, Andrew Smart, and William S. Isaac. 2020. Extending the Machine Learning Abstraction
Boundary: A Complex Systems Approach to Incorporate Societal Context. arXiv:2006.09663.
[90] Karl Marx. [n.d.]. Economic and Philosophic Manuscripts of 1844.
[91] Luke Mastin. 2009. Moral Universalism. https://www.philosophybasics.com/branch_moral_universalism.html.
[92] Karishma Mehrotra. 2022. Human Touch. Fifty Two (July 2022).
[93]
Annapurna Menon. 2022. Debunking Hindutva Appropriation of Decolonial Thought. Interfere: Journal for Critical Thought and Radical Politics 3
(2022), 36–57.
[94]
Sábëlo Mhlambi and Simona Tiribelli. 2023. Decolonizing AI Ethics: Relational Autonomy as a Means to Counter AI Harms. Topoi 42, 3 (July 2023),
867–880.
[95]
Milagros Miceli and Julian Posada. 2022. The Data-Production Dispositif. In Proceedings of the ACM on Human-Computer Interaction CSCW2. 460.
[96]
Walter D. Mignolo. 2010. Introduction: Coloniality of Power and De-Colonial Thinking. In Globalization and the Decolonial Option, Walter D.
Mignolo and Arturo Escobar (Eds.). Routledge, London, UK, 1–21.
[97] Walter D. Mignolo and Catherine E. Walsh. 2018. On Decoloniality: Concepts, Analytics, Praxis. Duke University Press, Durham, NC, USA.
[98]
Veli Mitova. 2021. How to Decolonise Knowledge Without Too Much Relativism. In Decolonisation as Democratisation: Global Insights into the
South African Experience, Siseko H. Kumalo (Ed.). Lynne Rienner Publishers.
[99]
Shakir Mohamed, Marie-Therese Png, and William Isaac. 2020. Decolonial AI: Decolonial Theory as Sociotechnical Foresight in Articial Intelligence.
Philosophy & Technology 33 (July 2020), 659–684.
[100] Jakob Mökander, Jonas Schuett, Hannah Rose Kirk, and Luciano Floridi. 2023. Auditing Large Language Models: A Three-Layered Approach. AI
and Ethics (2023).
[101]
Niklas Möller. 2016. Value Uncertainty. In The Argumentative Turn in Policy Analysis: Reasoning About Uncertainty, Sven Ove Hansson and
Gertrude Hirsch Hadorn (Eds.). Springer, Switzerland, 105–133.
[102]
Parthasarathi Mondal. 2012. Philosophy and Nationalism in India: A Preliminary Essay. Journal of Social Work and Social Development 3, 1–2
(June–Dec. 2012).
[103]
Shashi Motilal, Keya Maitra, and Prakriti Prajapati. 2021. Care, Community, Compassion and Virtue: Decolonizing Our Moral Landscape. In The
Ethics of Governance: Moral Limits of Policy Decisions. Springer, Singapore, 141–176.
[104]
Alexander Muir. 2021. Where HCI Meets the Spiritual Path: The Three Yogas of the Bhagavad G
¯
ıt
¯
a. In Extended Abstracts of the CHI Conference on
Human Factors in Computing Systems. 38.
[105] James Muldoon and Boxi A. Wu. 2023. Articial Intelligence in the Colonial Matrix of Power. Philosophy & Te chnology 36, 4 (Dec. 2023), 80.
[106]
Manish Nagireddy, Lamogha Chiazor, Moninder Singh, and Ioana Baldini. 2024. SocialStigmaQA: A Benchmark to Uncover Stigma Amplication
in Generative Language Models. In Proceedings of the AAAI Conference on Articial Intelligence, Vol. 38. 21454–21462.
[107]
Tiara R Na’puti and Joelle M. Cruz. 2022. Mapping Interventions: Toward a Decolonial and Indigenous Praxis Across Communication Subelds.
Communication, Culture and Critique 15, 1 (March 2022), 1–20.
[108]
Jessica Nix. 2023. One AI Startup Wants to Tackle Bias by Teaching Black History: Equality. https://www.bloomberg.com/news/newsletters/2023-
10-19/latimer-ai-startup-for-hcbus-trains-model-on-african-american-history-culture. Bloomberg Newsletter (Oct. 2023).
[109] Ritesh Noothigattu, Djallel Bouneouf, Nicholas Mattei, Rachita Chandra, Piyush Madan, Kush R. Varshney, Murray Campbell, Moninder Singh,
and Francesca Rossi. 2019. Teaching AI Agents Ethical Values Using Reinforcement Learning and Policy Orchestration. IBM Journal of Research
and Development 63, 4/5 (July–Sept. 2019), 2.
[110] Ezinne Nwankwo and Belona Sonna. 2019. Africa’s Social Contract with AI. ACM XRDS: Crossroads 26, 2 (Winter 2019), 44–48.
[111]
Long Ouyang, Je Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex
Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and
Ryan Lowe. 2022. Training Language Models to Follow instructions with Human Feedback. In Advances in Neural Information Processing Systems.
27730–27744.
[112] Kamna Patel. 2020. Race and a Decolonial Turn in Development Studies. Third World Quarterly 41, 9 (July 2020), 1463–1475.
[113]
Beth Patin, Melinda Sebastian, Jieun Yeon, Danielle Bertolini, and Alexandra Grimm. 2021. Interrupting Epistemicide: A Practical Framework
for Naming, Identifying, and Ending Epistemic Injustice in the Information Professions. Journal of the Association for Information Science and
Technology 72, 10 (Oct. 2021), 1306–1318.
[114]
Sachin R. Pendse, Daniel Nkemelu, Nicola J. Bidwell, Sushrut Jadhav, Soumitra Pathare, Munmun De Choudhury, and Neha Kumar. 2022. From
Treatment to Healing: Envisioning a Decolonial Digital Mental Health. In Proceedings of the CHI Conference on Human Factors in Computing Systems.
Manuscript submitted to ACM
20 Kush R. Varshney
548.
[115] Billy Perrigo. 2023. OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic. Time (Jan. 2023).
[116] Billy Perrigo. 2023. The Workers Behind AI Rarely See Its Rewards. This Indian Startup Wants to Fix That. Time (July 2023).
[117] Associated Press. 2023. Meta and IBM Launch ‘AI Alliance’ to Promote Open-Source AI Development. The Guardian (Dec. 2023).
[118]
Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. 2023. Fine-Tuning Aligned Language Models
Compromises Safety, Even When Users Do Not Intend To! arXiv:2310.03693.
[119] Aníbal Quijano. 2007. Coloniality and Modernity/Rationality. Cultural Studies 21, 2-3 (April 2007), 168–178.
[120]
Inioluwa Deborah Raji, Emily M. Bender, Amandalynne Paullada, Emily Denton, and Alex Hanna. 2021. AI and the Everything in the Whole Wide
World Benchmark. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks.
[121]
Shyam Ranganathan. 2016. N
¯
ag
¯
arjuna and Madhy
¯
amaka Ethics. In Philosophy, E-PG Pathshala, A. Raghuramaraju (Ed.). India National Mission on
Education through Information and Communication Technology, Delhi.
[122]
Shyam Ranganathan. 2022. Hinduism, Belief and the Colonial Invention of Religion: A Before and After Comparison. Religions 13, 10 (Sept. 2022),
891.
[123] John Rawls. 2009. A Theory of Justice. Harvard University Press, Cambridge, MA, USA.
[124] Jeremiah Reyes. 2015. Loób and Kapwa: An Introduction to a Filipino Virtue Ethics. Asian Philosophy 25, 2 (June 2015), 148–171.
[125] Paola Ricaurte. 2019. Data Epistemologies, the Coloniality of Power, and Resistance. Television & New Media 20, 4 (May 2019), 350–365.
[126]
Paola Ricaurte. 2022. Ethics for the Majority World: AI and the Question of Violence at Scale. Media, Culture & Society 44, 4 (May 2022), 726–745.
[127] Josh Schrei. 2010. The God Project: Hinduism as Open-Source Faith. https://www.hupost.com/entry/the-god-project-hinduism_b_486099.
[128]
Andrew D. Selbst, danah boyd, Sorelle A. Friedler, Suresh Venkatasubramanian, and Janet Vertesi. 2019. Fairness and Abstraction in Sociotechnical
Systems. In Proceedings of the Conference on Fairness, Accountability, and Transparency. 59–68.
[129] Amartya Sen. 2005. The Argumentative Indian: Writings on Indian History, Culture and Identity. Penguin Books.
[130]
Giorgio Shani and Navnita Chadha Behera. 2022. Provincialising International Relations through a Reading of Dharma. Review of International
Studies 48, 5 (Dec. 2022), 837–856.
[131]
Renee Shelby, Shalaleh Rismani, Kathryn Henne, AJung Moon, Negar Rostamzadeh, Paul Nicholas, N’Mah Yilla, Jess Gallegos, Andrew Smart,
Emilio Garcia, and Gurleen Virk. 2022. Sociotechnical Harms: Scoping a Taxonomy for Harm Reduction. arXiv:2210.05791.
[132] Partha Pratim Shil. 2020. The Indian Liberal Nostalgia for a Tolerant Hinduism Is Misplaced. The Wire (Aug. 2020).
[133]
Ben Shneiderman and Michael Muller. 2023. On AI Anthropomorphism. https://medium.com/human-centered-ai/on-ai-anthropomorphism-
ab4cecc5ae.
[134] Siddhartha. 2008. Open-Source Hinduism. Religion and the Arts 12, 1-3 (Jan. 2008), 34–41.
[135]
Maïka Sondarjee and Nathan Andrews. 2022. Decolonizing International Relations and Development Studies: What’s in a Buzzword? International
Journal: Canada’s Journal of Global Policy Analysis 77, 4 (Dec. 2022), 551–571.
[136] Mimi St. Johns. 2023. Against Relationality: A Response to Abeba Birhane. GRACE: Global Review of AI Community Ethics 1, 1 (Feb. 2023).
[137]
Zhiqing Sun, Yikang Shen, Qinhong Zhou, Hongxin Zhang, Zhenfang Chen, David Cox, Yiming Yang, and Chuang Gan. 2023. Principle-Driven
Self-Alignment of Language Models from Scratch with Minimal Human Supervision. In Advances in Neural Information Processing Systems.
[138]
Dheepa Sundaram. 2022. Hindutva 2.0: How a Conference on Hindu Nationalism Launches a Change in Strategy for North American Hindutva
Organizations. Journal of the American Academy of Religion 90, 4 (2022), 809–814.
[139]
Gokul Swamy, Christoph Dann, Rahul Kidambi, Zhiwei Steven Wu, and Alekh Agarwal. 2024. A Minimaximalist Approach to Reinforcement
Learning from Human Feedback. arXiv:2401.04056.
[140]
Jasmina Tacheva and Srividya Ramasubramanian. 2023. AI Empire: Unraveling the Interlocking Systems of Oppression in Generative AI’s Global
Order. Big Data & Society 10, 2 (Dec. 2023), 20539517231219241.
[141] Shashi Tharoor. 2018. Why I Am a Hindu. Scribe, Croydon, UK.
[142] Adriaan van Klinken. 2020. Studying Religion in the Pluriversity: Decolonial Perspectives. Religion 50, 1 (Oct. 2020), 148–155.
[143] Ashutosh Varshney and Connor Staggs. 2024. Hindu Nationalism and the New Jim Crow. Journal of Democracy 35, 1 (Jan. 2024), 5–18.
[144] Kush R. Varshney. 2018. AI and Safety. Uehiro-Carnegie-Oxford Conference: Ethics and the Future of Articial Intelligence.
[145]
Lav R. Varshney, Nitish Shirish Keskar, and Richard Socher. 2019. Pretrained AI Models: Performativity, Mobility, and Change. arXiv:1909.03290.
[146]
João Viera Magalhães and Nick Couldry. 2021. Giving by Taking Away: Big Tech, Data Colonialism and the Reconguration of Social Good.
International Journal of Communication 15 (2021), 343–362.
[147]
Yufei Wang, Wanjun Zhong, Liangyou Li, Fei Mi, Xingshan Zeng, Wenyong Huang, Lifeng Shang, Xin Jiang, and Qun Liu. 2023. Aligning Large
Language Models with Human: A Survey. arXiv:2307.12966.
[148]
Laura Weidinger, Jonathan Uesato, Maribeth Rauh, Conor Grin, Po-Sen Huang, John Mellor, Amelia Glaese, Myra Cheng, Borja Balle, Atoosa
Kasirzadeh, Courtney Biles, Sasha Brown, Zac Kenton, Will Hawkins, Tom Stepleton, Abeba Birhane, Lisa Anne Hendricks, Laura Rimell, William
Isaac, Julia Haas, Sean Legassick, Georey Irving, and Iason Gabriel. 2022. Taxonomy of Risks Posed by Language Models. In Proceedings of the
ACM Conference on Fairness, Accountability, and Transparency. 214–229.
[149]
David Gray Widder, Sarah West, and Meredith Whittaker. 2023. Open (For Business): Big Tech, Concentrated Power, and the Political Economy of
Open AI. SSRN:4543807.
Manuscript submitted to ACM
Decolonial AI Alignment: Openness, Viśes
.
a-Dharma, and Including Excluded Knowledges 21
[150]
Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu, Noah A. Smith, Mari Ostendorf, and Hannaneh Hajishirzi.
2023. Fine-Grained Human Feedback Gives Better Rewards for Language Model Training. arXiv:2306.01693.
[151]
Dun Zeng, Yong Dai, Pengyu Cheng, Tianhao Hu, Wanshun Chen, Nan Du, and Zenglin Xu. 2023. On Diversied Preferences of Large Language
Model Alignment. arXiv:2312.07401.
[152]
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao
Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. arXiv:2306.05685.
[153]
Siphamandla Zondi. 2020. Decolonising International Relations and Its Theory: A Critical Conceptual Meditation. In Decolonisation after Democracy,
Laurence Piper (Ed.). Routledge, London, UK, 16–31.
Manuscript submitted to ACM
