Artificial intelligence holds up a mirror to whoever holds it

On 15 March 2016, in the Four Seasons Hotel, Seoul, a Go board was set up for the final game of a five-match series. Across the board sat Lee Sedol, the world's most decorated player of the 2,500-year-old game, and a computer programme named AlphaGo, developed by DeepMind in London. AlphaGo won the series 4-1. The commentators said the machine had learnt to play Go. The truer description is that the machine had learnt to play Go like the humans who had built it, trained on 30 million moves from human master games. Artificial intelligence is not an alien intelligence parachuted into our civilisation; it is an amplifier of the dataset, the designer, and the deploying institution. It holds up a mirror to whoever holds it. Bias in, bias out; vision in, vision out.

The mirror first reflects our biases. Joy Buolamwini, then a graduate student at the MIT Media Lab, ran an audit of commercial facial-recognition systems and published Gender Shades in February 2018. The findings were stark. On lighter-skinned men, the error rate of classifying gender was 0.8 per cent; on darker-skinned women, it was 34.7 per cent. The training datasets were overwhelmingly white and male. Amazon's experimental recruiting AI, scrapped in 2018, penalised resumes that contained the word 'women's', having learnt that the company's past hires were predominantly male. India's Aadhaar-linked welfare distribution, when piloted aggressively in Jharkhand and Rajasthan, produced documented exclusions of marginal Dalit and Adivasi populations whose fingerprints were worn by labour or whose Internet connectivity faltered; Reetika Khera and Jean Dreze's 2019 study estimated 2.5 to 6 per cent ration-card exclusion. Algorithms governing welfare reproduce the social hierarchies they were trained to assist.

The mirror then reflects power. The large language models now reshaping knowledge work, GPT-4 and 5 from OpenAI, Claude from Anthropic, Gemini from Google DeepMind, were trained primarily on English-language Internet text scraped before 2024. Their default responses, their grammatical preferences, their cultural defaults are Anglophone and Northern. The Centre for Strategic and International Studies estimated in 2023 that 96 per cent of training data for the largest models was in English; Hindi, Bengali, Marathi, Tamil together accounted for under 0.3 per cent. This is linguistic colonialism by another name. India has, in response, built local infrastructure. Bhashini (Digital India BHASHINI), launched in 2022, is a national language-translation mission spanning 22 scheduled languages. AI4Bharat, hosted at IIT Madras, has open-sourced the IndicTrans2 translation suite and the IndicBERT family of models. Sarvam-1, the first Indian-built large language model with 2 billion parameters trained on Indian languages, was launched in October 2024. The IndiaAI Mission was approved on 7 March 2024 with an outlay of Rs 10,372 crore over five years to fund foundation models, compute capacity (10,000 GPUs), and a national dataset platform.

The mirror also reflects accountability. The European Union AI Act, passed on 13 March 2024, is the world's first horizontal AI law. It tiers AI by risk: 'unacceptable' systems like social-scoring are banned; 'high-risk' applications in employment, education, and critical infrastructure face strict obligations; 'general-purpose' models above certain compute thresholds must publish a training-data summary. India's response combines NITI Aayog's Responsible AI Principles of 2021, MeitY's AI advisory of March 2024, and the IT Rules amendments of 2025 addressing deepfakes. The Supreme Court's 2023 caution in Jaswinder Singh v State of Punjab, that 'AI cannot replace human judgment in judicial reasoning', set a quietly important precedent.

The deepest concern is not bias in any single system but compounding across systems. AI is used today to triage CT scans, screen resumes, allocate welfare, route traffic, generate news copy, propose tax assessments, and target advertisements. Each application may individually be benign; their interaction is a mosaic of micro-decisions that shapes a life. ProPublica's 2016 investigation of the COMPAS recidivism algorithm in Florida found black defendants nearly twice as likely as white to be incorrectly labelled high-risk. The 2024 Delhi High Court order on Anil Kapoor's personality rights is an early skirmish of a much larger jurisprudential war.

An honest counter-argument should be heard. AI, sceptics say, is just another tool, like a hammer; tools do not have ethics, users do. The objection misses a category. Hammers do not learn in real time from millions of users; they do not adapt output to maximise engagement; they do not produce false but plausible content at scale. AI's distinctive properties, adaptivity, scale, opacity, and emergent capability, demand new categories of governance.

The Indian opportunity is real. India has the world's largest pool of English-speaking technical talent, the second-largest mobile-data market, and a digital public infrastructure that other countries are now studying. The DPI architecture treats data as a public utility, not a corporate asset; the AI architecture can be built on the same paradigm. The IndiaAI Mission's AIKosh dataset platform and the AI4Bharat Bhasha datasets are infrastructure-level interventions that can democratise access. The challenge is to insist that 'made in India' AI is a commitment to Indian-language inclusion, Indian-data dignity, and Indian-context safety testing.

What might full implementation look like? The Digital India Act should specify a risk-tiered framework akin to the EU AI Act: cap deepfake political content, mandate auditability of welfare-allocation algorithms, license high-risk systems for healthcare and judiciary, and create a statutory AI Authority of India. Public procurement should mandate documentation of training data, model-card disclosure, and bias audits. The Bhashini project must extend to all 22 scheduled languages and tribal tongues. Mission Karmayogi should add an AI-literacy module for civil servants.

Vivekananda, in his 1899 talk on 'The Education That India Needs', said, 'We are what our thoughts have made us; so take care about what you think.' AI is what our data has made it. To build trustworthy AI in India, we must first be a country worthy of being mirrored: less biased in our hiring, less colonised in our language, more careful with our public archives, more transparent with our algorithms. Algorithmic justice is downstream of social justice; an AI built atop an unjust dataset will faithfully reproduce that injustice at scale and speed. The choice before the Indian republic is therefore not whether to embrace AI but how to ensure that the mirror, when finally polished, reflects the constitutional values of equality, liberty, and fraternity rather than the worst of our inherited prejudices. To get this right is one of the defining tasks of the coming decade. The mirror is being held up. What we will see in it is, finally, ourselves.