Generative AI - ‘Generations’ Of Biases
Author: LawSphereX
Category: Technology, Cyber & Privacy
Published:
Updated:
GENERATIVE AI: ‘GENERATIONS’ OF BIASES
Anisha Mukesh Kedia, Jindal Global Law School, JGU
ABSTRACT
Generative AI has established a stronghold across sectors in the fast moving and ever developing age of technology. It possesses immense transformative power, such that its use and deployment by individuals, corporations etc, seems almost indispensable. However, Generative AI systems, like any other breakthrough technology, after the initial fervour now sees a period of budding pessimism and caution from its users and governments across the globe. The recent years have seen governments seeking to regulate and form workable governance structures for Artificial Intelligence tools. The technology undermines various human rights and other rights such as privacy, non-discrimination, right to work etc. This paper focusses on the right to non-discrimination and how the use of Generative AI impacts the same.
It seeks to enquire how bias, especially stemming from gender and racial stereotypes, is reinforced through the use of Generative AI, and what are its larger implications in society considering its use and deployment for decision making across various sectors and industries- both public and private. The paper explores the current national and international legal frameworks, governance strategies and developer/ corporation-led reforms and their effectiveness in tackling the issue of biases, with special focus on India. It argues for international cooperation and reform to be embedded in a human rights taxonomy for effective redressal of the violations. Given the fast-moving nature of such technology, it is evident that governments and legal frameworks may find it difficult to keep pace with it.
Thus, a principlebased approach to legislation, with multi-level stakeholder corporation and developer led reforms, focussed on transparency and risk management, may be best suited until comprehensive legislations may come into place. The paper also discusses various suggestions made by experts and academia, for addressing the issues of bias by adopting measures across the lifecycle of the Generative AI products.
Introduction
The new age digital technology and its pervasiveness in everyday life of the people has most definitely ushered the world into an age of immense innovation and progress. The advent of Artificial Intelligence (AI) has not only enhanced ways of working across different sectors but has enhanced the way of life. More specifically, the dependence and extensive use of Generative AI, although having simplified various tasks, has not been without the pitfalls of human rights issues. The objectivity benefit of AI has been overstated, in that it has failed to keep human biases at bay, leading to perpetuation and entrenchment of existing stereotypes, and furthermore impacting the spread of information and decision making that employs such technology. Despite the technological advancement this issue has led to a reproduction of the injustices of the past.
The paper seeks to contextualise these looming issues of bias and discrimination posed by Generative AI, in a human rights taxonomy and proposes for a coordinated multi-level, multi-stakeholder approach, with more focus on developer lead reforms and governance. It discusses current legal reforms and adoptions at the intergovernmental and national levels, with emphasis on the regulatory landscape in India to tackle the issue and their potential effectiveness, arguing for strong implementation of transparency and accountability requirements for effective and safe use of Generative AI. It addresses the need for international co-operation in this arena and the solutions and regulatory techniques that have been suggested by academia for tackling the issue of bias. I.
The Bias Problem: Evidence and Its Human Rights Implication. Generative AI is a kind of technology that is based on a training model, where the technology is trained on large datasets to produce a certain output by finding patterns and relationships. A primary kind is Large Language Models (LLMs), which respond to text prompts. Developed iterations of the same are multi-modal models capable of producing and responding to audio and image outputs and inputs respectively. The very fundamental ways of training, development and deployment of the AI has the potential to exacerbate the issue of bias creation, leading to discriminatory output. Such AI systems are used across various sectors in automated decision making such as hiring (Human Resource management), education and research, healthcare, law enforcement etc. and can have potentially far-reaching impacts.
Evidence and research points to how bias generated by AI systems has percolated in these sectors.
An experimental study conducted, revealed how Stable Diffusion, when responding to text-toimage prompts, produced output, relating higher paying jobs with the fairer skin tone and men whereas lower paying jobs were associated with darker skin tone and women. 1 On comparison with actual US demographics, it was found that the AI results did not reveal the true picture. ChatGPT outputs too have sexist underpinnings where the old non-inclusive ideas pertaining to gender are perpetuated. 2 The feature of changing personas of the ChatGPT bot has highlighted the stereotypes that the system has imbibed and the toxicity it produces- where assigning personas such as dictator, Muhammad Ali etc have produced negative, injurious and biased outputs. 3 Google Photos’ facial recognition system, identified two Black persons as gorillas.
4 Google’s search results showcased sexually explicit pictures when ‘Black Girls’ was given as the search input. 5 Algorithmic systems as such which have been used to analyse data for decision making, have produced results that are unfair and discriminatory towards black people, in the medical field6 as well as in law enforcement7. ChatGPT has amplified sexual biases, in its inability to translate from low resource languages to higher ones, unable to effectively translate the use of gender-neutral pronouns and designated default ones to certain occupations. 8 The under-representation of low resource languages also leads to underperformance of the Generative AI systems, deepening the digital divide and data poverty issues9. Beyond racial and sexual discrimination, the use of such AI impacts the Global North- 1 L. Nicoletti and D. Bass, ‘Humans are biased.
Generative AI is even worse,” (Bloomberg, 9 June 2023) accessed 19 November 2024 2 N.
Gross, ‘What chatgpt tells us about gender: A cautionary tale about performativity and gender biases in ai’ (2023) 12 (8) Social Sciences 435; Morgan Briggs and Miranda Cross, ‘Generative AI: Threatening Established Human Rights Instruments at Scale’ (2024) ICAPAI accessed 19 November 2024 3 Ameet Deshpande, ‘Toxicity in ChatGPT: Analyzing Persona-assigned Language Models’ (2023) Findings of the Association for Computational Linguistics: EMNLP 1236 4 Alex Hern, ‘Google's solution to accidental algorithmic racism: ban gorillas’ (The Guardian, 12 January 2018) accessed 19 November 2024 5 Sahajveer Baweja and Swapnil Singh, ‘Beginning of Artificial Intelligence, End of Human Rights’ (LSE Blogs, 16 July 2020) accessed 19 November 2024 6 Carolyn Johnson ‘Racial bias in a medical algorithm favors white patients over sicker black patients’ (The Washington Post, 24 October 2019)
7 Lindsey Andersen, ‘Human Rights in The Age of Artificial Intelligence’ (Access Now, 2018)
accessed 19 November 8 S. Ghosh and A. Caliskan, ‘Chatgpt perpetuates gender bias in machine translation and ignores non-gendered pronouns: Findings across Bengali and five other low-resource languages’ (2023) Proceedings of AAAI/ACM Conference on AI, Ethics, and Society accessed 19 November 2024 9 United Nations Human Rights Office of the High Commissioner (B-Tech), ‘Taxonomy of Human Rights Risks Connected to Generative AI’ (2 November 2023) South divide, disadvantaging the latter, as the training data sets for the systems are reflective of the Western cultural hegemony. The impact of AI and its structural biases can be felt greatly on vulnerable groups such as the poor, children, refugees, etc. A study reveals how a tool to predict child neglect and abuse, inherently is disadvantages the poor families.
10 The underlying cause of this issue is the training methodology and design of the system. The training data sets which are amassed from the vast amount of data available across the internet is problematic because it is rife with immense problems of fake news, misinformation and opinions falsely represented as facts. Thus, from a quality perspective this data is far from perfect. 11 The data on the internet is also largely English- language dominated thus the AI systems in themselves become inaccessible to large non-English speaking communities, unfairly discriminating against the Global South majorly.
When the model itself is trained on data that is biased, rife with stereotypes and is under-representative of certain cultures, communities and groups, the outputs produced by them further entrench the bias, disadvantaging certain classes based on their gender, race, ethnicity, language, social status, sexuality etc. The information scraped off the internet is used without any filtering and monitoring of the data, creating hurdles at the model level itself. The use of the biased output data when used in further training of the models, creates a feedback loop12, where bias spreads to such an unprecedented scale that its far from elimination. Even in cases where explicit biases may be absent in the training data sets, the AI models, based on subtle, remote and unjustified correlations can produce biased output.
13 Additionally, overreliance on such data, where no explanation of decision making by the AI, or public 10 Rowena Rodrigues, ‘Legal and human rights issues of AI: Gaps, challenges and vulnerabilities’ (2020) 4 Journal of Responsible Technology accessed 19 November 2024 11 Philipp Hacker and others, ‘Regulating ChatGPT and other Large Generative AI Models’ (2023) Fairness, Accountability, and Transparency accessed 19 November 2024 12 Axel Schwanke, ‘The EU AI Act: Best Practices for Transparency and Explainability’ (Medium, 8 September 2024) accessed 19 November 2024; ‘As gen AI advances, regulators—and risk functions—rush to keep pace’ (McKinsey & Company, December 2023) accessed 19 November 2024 13 Axel Schwanke, ‘The EU AI Act: Best Practices for Transparency and Explainability’ (Medium, 8 September 2024) accessed 19 November 2024 information regarding the data sets is provided, is problematic in that, the end user is unaware the biases and underlying stereotypes of the output.
This can lead to manipulation of public opinions, perceptions and conduct. Thus, the objectivity aims with which such AI tools are deployed, to achieve equality, are not really achieved and is a disgrace to the transformative goal it sought. 14 II.
Foregrounding Of The Discourse In A Human Rights Taxonomy And The Multi- Stakeholder Approach The current discourse and initiatives do acknowledge the risks of Generative AI and its ethical implications, but it still needs to be foregrounded in the language of human rights, for reforms to materialise and reach their full potential. They lack the kind of due diligence mechanisms that are proposed under the United Nations Guiding Principles (UNGP)15 and Organisation for Economic Cooperation and Development (OECD) Principles16, which can serve as foundations to future regulatory mechanisms. 17 The United Nations’ High Commissioner for Human Rights, emphasized on incorporating “the common language of human rights” into the governance, deployment and regulation of new age technologies such as Generative AI. 18.
The office of the United Nations High Commissioner for Human Rights (OHCHR), through its Btech project initiative has released papers discussing the potential risks, implications and solutions pertaining to the deployment of Generative AI. To use the language of international human rights is particularly beneficial, in that it reinforces existing obligations of states under international agreements and businesses under their corporate responsibility commitments. 19 It provides a more specific language to address the 14 Sahajveer Baweja and Swapnil Singh, ‘Beginning of Artificial Intelligence, End of Human Rights’ (LSE Blogs, July 2020) accessed 19 November 2024 15 Office of the High Commissioner for Human Rights (OHCHR), ‘Guiding principles on business and human rights: Implementing the United Nations “protect, respect, and remedy framework” ’ 2011
16 OECD, ‘Recommendation of the Council on Artificial Intelligence’ (22 May 2019) OECD/LEGAL/0449
17 Morgan Briggs and Miranda Cross, ‘Generative AI
Threatening Established Human Rights Instruments at Scale’ (2024) ICAPAI accessed 19 November 2024; United Nations Human Rights Office of the High Commissioner (B-Tech), ‘Advancing Responsible Development and Deployment of Generative AI’ (November 2023) 18Ines Najeh, ‘Digital Rights: The Impact of AI and Emerging Technologies on Human Rights’ (Geneva International Centre for Justice, 17 June 2024) accessed 19 November 19 United Nations Human Rights Office of the High Commissioner (B-Tech), ‘Taxonomy of Human Rights Risks Connected to Generative AI’ (2 November 2023) violations in a common language of rights understood across the diverse set of stakeholders, thus helping garner maximum momentum in regulating the Generative AI technology.
20 The larger human right which the bias problem of Generative AI violates is that of right to equality and non-discrimination, enshrined in Article 2 and Article 7 of the Universal Declaration of Human Rights (UDHR) and Article 26 of the International Covenant on Civil and Political Rights (ICCPR). The solutions that the organisation proposes is that of a multi-lateral, multi stakeholder approach where responsibility to address the concerns are placed on the state, the corporations and end users. It argues for implementation of a “smart-mix” of regulations such as- risk management and assessment policies; periodical review mechanisms; transparency, accountability and explainability obligations; implementing best practices standards; redressal and enforcement mechanisms and AI literacy.
21 Additionally, governance of AI requires to be done throughout the AI lifecycle- from its training through collected data sets, algorithmic modelling and deployment, with measures tailored to each part tackling the various layers of bias generated. 22 III.
Regulatory Guidance And Frameworks At Play- Effectiveness And Need For International Cooperation. Multiple nations such as the EU, America, China etc. have formulated legislations and policies to regulate AI, but it is yet to be seen how effective they are and whether they align with the suggestions and guidelines of the United Nations Guiding Principles on Business and Human Rights (UNGP), OECD and other organisations. There have been multiple efforts at the intergovernmental level as well by organisations such as the OECD, G20, G7, United Nations etc. The regulations although fragmented, making it difficult for operations of globally present 20 United Nations Human Rights Office of the High Commissioner (B-Tech), ‘Taxonomy of Human Rights Risks Connected to Generative AI’ (2 November 2023)
21 United Nations Human Rights Office of the High Commissioner (B-Tech), ‘Advancing Responsible
Development and Deployment of Generative AI’ (November 2023) 22 Shuchen Tang and Haoming Zhu, ‘Mitigating bias in generative AI: a comprehensive framework for governance and accountability’ (2024) 3 Law Ethics Technol. ELSP accessed 19 November 2024 companies, have common themes of regulation. 23 The common themes include mainly transparency, explainability and accountability. Another theme that emerges is the importance of human oversight over these mechanisms. A. Intergovernmental regulations & recommendations At the intergovernmental level, the OECD pioneered the guidelines for AI in its ‘Recommendation on Artificial Intelligence’24.
It urges that stakeholders of AI implement five large principles for responsible AI namely- a) sustainable development and wellbeing; b) human-centred values and fairness; c) transparency and explainability; d) robustness, security and safety; and e) accountability. It urges the development of national policies and is open to be adhered by non-OECD countries as well. United Nations too adopted the UNESCO Recommendation on Artificial Intelligence in response to the concerns raised by the OHCHR on new and emerging digital technologies. Of the 10 principles it envisages for a human rights approach to the use and deployment of AI, some are transparency, explainability, accountability, human oversight, fairness and nondiscrimination, multi stakeholder governance and AI literacy.
25 Very recently, the UNGA has also adopted its resolution on AI Intelligence, which calls for techniques to mitigate bias among other governance measures. 26 At the international level, there has also been a policy statement in the form of the Toronto Declaration27, from various non-profits and non-governmental organisations which advocate for responsible AI, focussing especially on bias and potential discrimination mitigation. 23 ‘As gen AI advances, regulators—and risk functions—rush to keep pace’ (McKinsey & Company, 21 December 2023) accessed 19 November 2024
24 OECD, ‘Recommendation of the Council on Artificial Intelligence’ (22 May 2019) OECD/LEGAL/0449
25 Tomiris Amirova, ‘Comparing Models of Artificial Intelligence Governance
The Role of International Cooperation on Responsible AI and the EU AI Act in the Age of Generative AI’ (Master Thesis, European University Institute 2023) accessed 19 November 2024
26 UNGA Res 78/265 (21 March 2024) art 6 (h)
27 ‘The Toronto Declaration: Protecting the rights to equality and non-discrimination in machine learning systems’ (AccessNow, 16 May 2018) B. European Union The EU AI Act28, which came into force in August 2024, is the first AI tailored horizontal regulation in the world. It adopts a risk based-approach across 4 categories- (a) unacceptable risk (b) high risk (c) limited risk (d) minimal risk, with varying degrees of applicable regulation. 29 Largely the regulation places transparency, explainability and accountability requirements on the AI systems. Beyond the risk classification, there exists a separate category called General Purpose AI Models, which include the Generative AI systems. 30 The GPAI models’ providers are subject to general transparency requirements under Article 5031 (for user interaction) and other specific obligations under Article 5332.
Article 50 requires that users be informed that they are interacting with AI systems and that output is AI generated. The obligations of providers of GPAI systems under Article 53, requires that they maintain documentation regarding the training and testing process and data used for the same. It requires information of training methodologies, the rational and assumptions made for its key choices, and details on the training data, how it was collected, the cure methodologies employed etc to gauge the suitability of the data and presence of any biases. These details are also required to be shared with the deployers. Additionally, for foundation models such as those which can be bases for other AI systems, stricter regulation requirements have been placed, such as data governance and risk management.
Moreover, when such models are applied in high-risk situations such as law, healthcare, employment etc. or that they perform certain profiling functions of natural persons they must clearly provide information on the capabilities and limitations of the same, explaining how the decisions are made and how to interpret the data. 33 For models categorised as high risk, the Act also contains a provision requiring human oversight over the technology. 34
28 The EU Artificial Intelligence Act [2024] OJ 1689/2024
29 The EU Artificial Intelligence Act [2024] OJ 1689/2024
30Tomiris Amirova, ‘Comparing Models of Artificial Intelligence Governance: The Role of International Cooperation on Responsible AI and the EU AI Act in the Age of Generative AI’ (Master Thesis, European University Institute 2023) accessed 19 November 2024
31 The EU Artificial Intelligence Act [2024] OJ 1689/2024 a.50
32 The EU Artificial Intelligence Act [2024] OJ 1689/2024 a.53
33 Axel Schwanke, ‘The EU AI Act
Best Practices for Transparency and Explainability’ (Medium, 8 September 2024) accessed 19 November 2024
34 Article 14, EU AI Act
A notable win in these regulations is that they differentiate between AI systems and models, systems being the applications, and regulate the models through which the obligations are then for the technology used. This helps in addressing the root cause of the problem. Although these regulations are trailblazing in the structured approach that they offer, it fails to address how generative AI not only in high-risk systems can pose bias issues but have an impact when such systems are consumer-facing as well. 35 The consumers, relying on AI data as objective given the speed and sophistication of the output, may very well imbibe the biases it produces. Even when not applied in high-risk situations, it impacts general human behaviour based on what they learn and read as information. C.
United States of America The country lacks a common federal law regulating AI technology, however multiple legislative policies discussions are underway. The current regulation in the US differs from state to state and across sectors. Some of the states that have legislations in place are- California, New York, Colorado, Columbia, Illinois etc. These legislations largely require impact assessment tools for AI. 36 For example, Illinois has a specific act regulating use of AI in video interviews, which places transparency and explainability notice requirements before employing the AI tool. 37 Additionally, of the first state led legislations within the USA, is Colorado’s ‘Colorado AI Act’, which delineates obligations and requirements for developers and deployers with focus on bias and discrimination mitigation.
38 There also exist multiple guideline documents and frameworks which are not binding but identify potential issues. One of them is the Blueprint for AI Bill of Rights, which is based on 35 Shuchen Tang and Haoming Zhu, ‘Mitigating bias in generative AI: a comprehensive framework for governance and accountability’ (2024) 3 Law Ethics Technol. ELSP accessed 19 November 2024 36 Tomiris Amirova, ‘Comparing Models of Artificial Intelligence Governance: The Role of International Cooperation on Responsible AI and the EU AI Act in the Age of Generative AI’ (Master Thesis, European University Institute 2023) accessed 19 November 2024
37 Arsen Kourinian and Mayer Brown, ‘Addressing Transparency & Explainability
When Using AI Under Global Standards’ (2024) Bloomberg Law accessed 19 November 2024
38 ‘AI Watch: Global regulatory tracker - United States’ (White&Case LLP, 13 May 2024)
accessed 19 November 2024 five principles namely- Safe and Effective Systems, Algorithmic Discrimination Protections, Data Privacy, Notice and Explanation, Alternative Options. 39 In October 2023, President Biden issued an executive order, on “Safe, Secure and Trustworthy Development of AI” through which it addressed regulation on generative AI, with a specific statement on mitigation of biases. The Biden administration committed to deploying AI as per standards that protect from unlawful discrimination, especially when employed in hiring, housing, healthcare. It commits to promote regulation, oversight and technical re-evaluation mechanisms for the same. 40 D. India While USA awaits any overarching federal law for regulating AI, specifically generative AI, India seems to have taken a middle path between regulation and innovation.
41 India too lacks a specific law regulating AI, however certain advisories have been issued by the Government departments and guidelines by the NITI Aayog. Despite various statements, guidelines, declarations etc. the position of the Indian government and its various agencies on the issue of regulation of AI has largely been polarised and confused. An advisory issued by the MeitY in March 2024, called for self-regulation by corporations requiring them to label their output as AI generated or as unreliable AI generated if not well tested. 42 This guideline is in line with the EU AI Act, which seeks to impose similar labelling requirements for high-risk systems. This advisory superseded and clarified the position of an erstwhile advisory, which had stipulated that intermediaries procure prior permission of the Indian Government for deploying any “untested” or “unreliable” AI models.
43 The later advisory did away with this requirement. The advisory is only binding in nature and not issued under any legal statute. Moreover, the stipulated obligation is on the intermediaries employing AI Models, LLMs, Generative AI etc 39 Ningxin Su and others, ‘Generative AI: Application and Regulation in Asia Pacific’ (Centre for Regulatory Strategy Deloitte, 2024) accessed 19 November 2024 40Marcin Szczepanski, ‘United States approach to artificial intelligence’ (European Parliamentary Research Service, January 2024) accessed 19 November 2024 41 Soibam Rocky Singh, ‘Stringent regulations could hinder growth of AI in India: experts’ The Hindu (New Delhi, 23 June 2024) accessed 19 November 2024 42 Amlan Mohanty and Shatakratu Sahu, ‘India’s Advance on AI Regulation’ (Carnegie India, 21 November 2024) accessed 28 February 2025
43 ibid
on their computer resources to ensure that there is no threat of bias or discrimination etc44 and not on the AI companies themselves. The relevant government agencies, rather than seeking to adopt an entirely new legislation to regulate AI, have considered amending existing regulations such as the Information Technology Act, 2000 to include regulations on AI, Generative AI Models. 45 Although nonbinding and mere guidelines, the NITI Aayog has published various documents46 that address concerns for safe design, deployment and use of AI, with increasing focus on principles of transparency, accountability, inclusion and non-discrimination etc. In the interim, the government has been taking steps to address the challenges of AI deployment across various sectors, the complex regulatory landscape for the same and formulate a workable framework of AI governance in the country.
A sub-committee on ‘AI Governance and Guidelines Development’ was formed under an Advisory Group set up by the Office of Principal Scientific Advisor which released its report on January 6, 2025. 47 As per the report, the IndiaAI Mission gained approval which seeks to catalyse AI innovation and establish a framework around AI governance, working with public and private players. The report calls for a “whole-of government” approach to AI Governance and for building a consensus on the approach to regulation. 48 The overall sentiment expressed in the report is that India should adopt a principle-based approach, where the governance framework is a combination of voluntary commitments, standards and risk-based regulations applicable to entities involved across the lifecycle of AI software and models.
Not only does the report express the impending need for regulation of AI having identified the gaps in existing laws in tackling issues of bias 44 Ministry of Electronics and Information Technology, ‘Due diligence by Intermediaries / Platforms under the Information Technology Act, 2000 and Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021. ’ (2024) No. eNo. 2(4)/2023-CyberLaws–3
45 Aashish Aryan, ‘Govt May Amend IT Act to Add New Rules for AI, GenAI models’ Economic Times (04
January 2024) accessed 28 February 2025
46 Ankit Agrawal, ‘Ensuring Responsible and Trustworthy AI for All’ (Invest India, 02 January 2024)
accessed November 2024 47 The Ministry of Electronics and Information Technology, Government of India, Report on AI Governance Guidelines Development, accessed 28 February 2024 48 The Ministry of Electronics and Information Technology, Government of India, Report on AI Governance Guidelines Development, accessed 28 February 2024 and non-discrimination but cautions against an overzealous prospective regulation regime that may be restrictive, without an appreciation of risk of harm being real and specific. The focus is more on industry led reforms such that the requirement of restrictive regulation doesn’t arise or stifle innovation and development of the transformative AI technology. The report must be hailed as a budding effort to take a balanced approach in regulation of AI and creating an environment for further innovation in the field.
It not only addresses regulation from the perspective of the lifecycle of the AI, need for a multi-stakeholder approach and a holistic moderately restrictive regime but also upholds and understands the need for AI innovation and technology for a developing nation such as India. 49 Given that there can be no one size fits all regulation to govern the ever evolving and extremely transformative potential of the AI technology, a principles-based-approach best suits the need of the hour. The principles recommended in the report align with the OECD principles and recommendations made by various think tanks. A principle-based approach would not only be a starting point for forthcoming regulations but also instrumental in helping build a global consensus on AI regulation given its transnational presence and cross-border uses. E.
Need for international cooperation The regulations across jurisdiction being sparse and fragmented make it difficult for regulations of larger, globally present multinational corporations. In this light it must be understood that given these regulations are in their nascent stage, and there exists a leading corpus of regulatory guidelines framework helmed by the OECD, international collaboration and cooperation is not unattainable. International cooperation in this regard, would help establish a common framework and understanding of AI, its associated problems and would allow for a global governance in this regard. There are multiple ways in which such cooperation can be achieved, not to say that there aren’t efforts underway to achieve the same already.
Countries are already concluding Free Trade Agreements and Bilateral Treaty Agreement on mutual use and access of AI, however the mutual cooperation in governance strategies and cooperation in regulation 49 The Ministry of Electronics and Information Technology, Government of India, Report on AI Governance Guidelines Development, accessed 28 February 2024 of the same is at an infant stage. 50 These agreements provide a good opportunity to bind one another into uniform regulatory mechanisms of the AI systems, requiring sharing of cross-data needed to help develop and train these models, such that they become more diverse and representative. Additionally, it would provide an opportunity for sharing AI governance experiences, learnings and best practices to effectively develop and common successful framework.
An example of such agreements and forums already in place is EU’s TTC with USA, which has launched the Joint Roadmap on Evaluation and Measurement Tools for Trustworthy AI and Risk Management. 51 This joint effort suggests adoption of collaborative measures to govern, regulate and develop AI that is safe for end users. The Brussels Effect52 is another phenomenon if materializes, that could help garner collaboration or setting up global standards to guide domestic legislations. Brazil has somewhat emulated EU’s approach in AI regulation, with its legislation yet to come out. 53 IV.
Tackling The Bias Problem Every Step Of The Way Academics, international officials and stakeholders such as NGOs have made various recommendations to help mitigate these biases as well, which include risk mitigation and data governance strategies. 54 To counter issues these solutions need to applied across the lifecycle of the data as has been suggested.
55 At the data layer, techniques such as adversarial training and counterfactual data augmentation can be used, alongside the efforts of amassing and verifying data to ensure its not under-representative of certain groups especially from low- 50 Joshua Meltzer, ‘Toward International Cooperation on Foundational AI Models: An Expanded Role for Trade Agreements and International Economic Policy’ (HDDSR, May 2024) accessed 19 November 2024 51 Tomiris Amirova, ‘Comparing Models of Artificial Intelligence Governance: The Role of International Cooperation on Responsible AI and the EU AI Act in the Age of Generative AI’ (Master Thesis, European University Institute 2023) accessed 19 November 2024 52 The Brussels Effect is the process by which the European Union unilaterally leads regulatory globalisation, where multinational corporations adopt EU standards when operating in non-EU countries as well.
53 Tomiris Amirova, ‘Comparing Models of Artificial Intelligence Governance: The Role of International Cooperation on Responsible AI and the EU AI Act in the Age of Generative AI’ (Master Thesis, European University Institute 2023) accessed 19 November 2024
54 Ines Najeh, ‘Digital Rights
The Impact of AI and Emerging Technologies on Human Rights’ (Geneva International Centre for Justice, 17 June 2024) accessed November 2024
54 Organisation for Economic Co-operation and Development (OECD), Advancing Accountability in AI:
Governing and Managing Risks Throughout the Lifecycle for Trustworthy AI, OECD Digital Economy Papers, No. 349 (OECD Publishing, February 2023).
55 Organisation for Economic Co-operation and Development (OECD), Advancing Accountability in AI:
Governing and Managing Risks Throughout the Lifecycle for Trustworthy AI, OECD Digital Economy Papers, No. 349 (OECD Publishing, February 2023). income and middle-income nations. 56 At the model layer and the algorithm layer, the goal of should be to prioritise context and not attain maximum accuracy at the cost of perpetuating biases. 57 Tools may include fairness constraints, regular audits, evaluations and human oversight at this stage. Hiring an AI ethicist as part of the development team, can help detect and mitigates such issues at an earlier stage. To ensure that human oversight is effectively able to mitigate these issues, it also important that the development and oversight teams themselves are diverse, such that there are minimal chances of the human biases in AI, re-entering or being confirmed by the individuals.
58 Transparency and explainability, a mechanism adopted in most legislations is characteristic here in that they provide the rationale behind decisions of the AI. This coupled with the solutions that can be applied at the application layer can help create a safe and regulated environment to use AI. At the application level or the user level, it is important to ensure AI literacy. This entails educating user and stakeholders of the various impacts of AI and its social and ethical considerations. It requires educating and informing the end user of the various kinds of biases inherent in AI decision making, such that the user can be wary of the output and use with limitations. The age-old adage, “garbage in, garbage out” is well applicable to the issue of bias in Generative AI models and systems.
Corporations have undertaken methods and measures to deal with this fundamental issue that arises in the inception and training stage of the LLM itself. Given that a biased LLM, cannot itself detect and fix the biases, corporations have are using newer datasets that are not only EU law complaint, but scrubbed of Personal Identifiable Information and fine tuned such that biases are kept at bay as much as possible. 59 Corporations also have resorted to certain best practices to fine tune and train the model such as Data quality Reviews, Reinforcement Learning from Human Feedback (RLHF) and Reinforcement Learning from AI Feedback (RLAIF). 60 However, these methods are not perfect, as human
56 Organisation for Economic Co-operation and Development (OECD), Advancing Accountability in AI:
Governing and Managing Risks Throughout the Lifecycle for Trustworthy AI, OECD Digital Economy Papers, No. 349 (OECD Publishing, February 2023).
57 Organisation for Economic Co-operation and Development (OECD), Advancing Accountability in AI:
Governing and Managing Risks Throughout the Lifecycle for Trustworthy AI, OECD Digital Economy Papers, No. 349 (OECD Publishing, February 2023). 58 Shuchen Tang and Haoming Zhu, ‘Mitigating bias in generative AI: a comprehensive framework for governance and accountability’ (2024) 3 Law Ethics Technol. ELSP accessed 19 November 2024 59 The Atlantic Council of United States, ‘Scaling Trust on the Web’ accessed 28 February 2024 60 Meta, ‘Meta Llama Responsible Use Guide’ (April 2024) accessed 28 February 2024 feedback could still be rife with biases which only results in the continuation of the “garbage in- garbage out” cycle. More than regulation, corporation led reforms can best prevent the generation of biased outputs through the use and deployment of Generative AI models. Companies have varying commitments for responsible AI embedded in differing kinds of taxonomies.
Certain companies like, Google, Salesforce have AI principles in their policy that explicitly refer to broad human rights protections. 61 Whereas companies such as Meta, Microsoft make commitments to protect more specific and identified rights. Emerging tech companies and startups lack such an approach. 62 Their commitments to responsible AI Governance are more company policy and ethics oriented. Corporations, as per public standards and best practices develop their own risk assessment methods at the development stage of the model and across other stages of its lifecycle. 63 Microsoft is one such example, where the company has required conducting of product reviews on the lines of “Responsible AI”. 64 Other methods include fairness testing and Algorithmic Audits and Impact Assessments.
Fairness testing involves checking whether output generated by prompts is under-representative, discriminatory, biased etc and further mending the training datasets accordingly. 65 Algorithmic Impact Assessments, include testing the models for any adverse social impacts and Audits are a more root-cause analysing tool. 66 These methods too have their limitations as they are not conducting assessments on the anvil of human rights, but against their own AI principles or in the taxonomy of risks or harms. They do closely resemble human rights assessments, but are not aligned with UNGPs, as their approach is bound within pre-identified risks and limited within the geographical context of where they operate rather than using internationally recognised human rights.
67 Companies which have conducted 61 United Nations Human Rights Office of the High Commissioner (B-Tech), ‘Responsible AI and Human Rights: An Overview of Company Practices’ (November 2023) 62 United Nations Human Rights Office of the High Commissioner (B-Tech), ‘Responsible AI and Human Rights: An Overview of Company Practices’ (November 2023) 63 United Nations Human Rights Office of the High Commissioner (B-Tech), ‘Responsible AI and Human Rights: An Overview of Company Practices’ (November 2023) 64 United Nations Human Rights Office of the High Commissioner (B-Tech), ‘Responsible AI and Human Rights: An Overview of Company Practices’ (November 2023) 65 United Nations Human Rights Office of the High Commissioner (B-Tech), ‘Responsible AI and Human Rights: An Overview of Company Practices’ (November 2023)
66 ibid
67 United Nations Human Rights Office of the High Commissioner (B-Tech), ‘Responsible AI and Human Rights: An Overview of Company Practices’ (November 2023) Human Rights Assessments have unfortunately not made them public.68 V.
Conclusion
Generative AI has permeated to the very beings of individuals in today’s world, with increasing dependence on the technology and its deployment in various aspects of life. The technology is not only a privilege for the developed nations but for everyone across the globe. However, the problem is that the development of the technology has been in the hands of the first world, developed nations. Taking a top-down approach in this regard, the technology has failed to incorporate and include the experiences and data from the other, minority or underrepresented nations such that the AI has imbibed in it the biases flagrant in the viewpoint of the West and the dominant. The regulations and guiding principles need to address the issue from a collaborative approach to avoid any arbitrage in policies that could be leveraged by AI companies to move to weakly regulated jurisdictions.
This paper attempted to lay bare the bias problem in the use of Generative AI, used for personal as well as public purposes especially for decision making. It briefly outlines existing relevant legislations and practices around the globe with special focus on the developing law in India. It has addressed the pressing need for a human rights-based advocacy for AI regulation across the globe, one which involves a multistakeholder approach and in the light of the existing regulatory framework in the world, echoed the need for international cooperation for AI governance. Lastly it has addressed the various methods suggested by academia in tackling the issue of bias, by attending to each part of the AI lifecycle from its design to deployment.
68 United Nations Human Rights Office of the High Commissioner (B-Tech), ‘Responsible AI and Human Rights: An Overview of Company Practices’ (November 2023)