[Image Generated by AI Image Generator - DALL-E]
In June 2023, India's Minister of State for Electronics and Information Technology hinted at the inclusion of provisions in the upcoming Digital India Act to prevent AI from harming users. Some of these may try to address AI's alignment problem, which includes biases in AI systems. While global standards are moving towards data quality norms to combat this, there are concerns:
- Strict data quality standards can limit data availability, impacting AI quality. China's struggle with AI, due to its data censorship, serves as a cautionary tale.
- Data quality standards might infringe on free speech rights.
- Even "clean" datasets can produce biased AI results, as seen with Amazon's recruitment AI.
The blog suggests that instead of just focusing on regulating AI development, safeguards should be in place during AI deployment, including informing users and allowing opt-outs from AI evaluations.
In June 2023, Rajeev Chandrashekhar, the Minister of State for Electronics and Information Technology, suggested that the upcoming Digital India Act[1] would include provisions to ensure that artificial intelligence (AI) platforms did not harm users.[2] If the Government is indeed developing such a framework for AI, it might, in part, be trying to address AI’s alignment problem. The alignment problem encompasses issues like the perpetuation of outdated stereotypes related to gender and race by AI systems. The data used to train AI can be inherently biased, leading to biased AI models.[3]
India might introduce data quality norms under the Digital India Act to reduce user harm from biased data. These norms emphasize the integrity of datasets, including accuracy, relevance, inclusivity, and bias hygiene. Such standards are becoming common in global legislation. For example, the European Union’s AI Act mandates that developers of high-risk AI systems adhere to data quality criteria, ensuring datasets are “relevant, representative, free of errors, and complete.” [4] While the intent behind regulatory efforts to address bias in AI systems is praiseworthy, policymakers should be wary of the drawbacks of overly emphasizing data quality.
Firstly, data quality standards might reduce data availability, affecting the reliability and quality of the AI system. China's experience with generative AI illustrates how data limitations can hinder a nation's AI aspirations. Although China is seen as a leader in AI, on-the-ground evidence suggests otherwise. A report highlighted that while China prioritizes AI dominance, it struggles to balance this with its tradition of information control.[5] In simpler terms, China's strict data censorship limits data availability. As a result, its AI systems, like the chatbot ERNIE by Baidu (China's Google equivalent), underperform. In one instance, ERNIE's poor performance led to a drop in Baidu's stock value.[6]
Secondly, data quality assurance standards, such as relevance and completeness, might infringe upon free speech rights. The right to free speech encompasses both sharing and receiving information.[7] Since data is a form of information, discarding training datasets that don't meet specific quality criteria could be seen as a violation of free speech rights, especially if these criteria exceed the reasonable speech restrictions outlined in Article 19(2) of the Constitution.[8]
Thirdly, there's no assurance that "clean" or "error-free" datasets will produce unbiased results. Data bias is just one facet of the alignment problem.[9] For instance, even with an inclusive dataset, the words themselves can carry biases.[10]
Amazon's attempt to use an AI system for recruitment serves as a case in point. In 2014, Amazon aimed to develop a program to sift through job applicant resumes.[11] By 2015, they found that the model was not “gender-neutral”.[12] It was trained on patterns from resumes submitted over a decade, most of which came from men, reflecting the tech industry's “male dominance”. [13] Even after adjusting the model for gender neutrality, there was no guarantee it wouldn't discriminate in other ways. [14]
In conclusion, AI alignment issues are hard to resolve, even for experts. Instead of focusing on regulating the AI development process to prevent harm, decision-makers may consider implementing safeguards around AI deployment. Specifically, in cases of potential AI-induced discrimination or bias, potentially affected parties should be informed and given the option to opt out of AI evaluations. As AI continues to shape our future, it is crucial that regulations prioritize transparency and user empowerment during deployment to ensure that technology augments human potential without unintended consequences.
——
[1] For those who are uninitiated, the Digital India Act is slated to replace India’s 23-year-old technology law, the Information Technology Act, 2000.
[3] Christian, Brian. The Alignment Problem: Machine Learning and Human Values. W.W. Norton & Company, 2020.
[4] Article 10(3), EU Artificial Intelligence Act, accessible at: https://eur-lex.europa.eu/resource.html?uri=cellar:e0649735-a372-11eb-9585-01aa75ed71a1.0001.02/DOC_1&format=PDF
[5] https://www.theatlantic.com/international/archive/2023/04/chatbot-ai-problem-china/673754/
[6] https://www.theatlantic.com/international/archive/2023/04/chatbot-ai-problem-china/673754/
[7] 1995 SCC (2) 161
[8] To be clear it could be argued that discriminatory AI systems violate other rights but these arguments go beyond the scope of this blog.
[9] Christian, Brian. The Alignment Problem: Machine Learning and Human Values.
[10] Christian, Brian. The Alignment Problem: Machine Learning and Human Values.
[11] Reuters. ‘Amazon Scraps Secret AI Recruiting Tool That Showed Bias against Women’. 10 October 2018, sec. Retail. https://www.reuters.com/article/us-amazon-com-jobs-automation-insight-idUSKCN1MK08G.
[12] Reuters. ‘Amazon Scraps Secret AI Recruiting Tool That Showed Bias against Women’.
[13] Reuters. ‘Amazon Scraps Secret AI Recruiting Tool That Showed Bias against Women’.
[14] Reuters. ‘Amazon Scraps Secret AI Recruiting Tool That Showed Bias against Women’.