Beyond AI Hype: Reading India’s Economic Survey 2025–26

The Economic Survey 2025–26, in its Chapter 14 on AI, marks an important inflection point in how India’s economic establishment talks about artificial intelligence – but it also illustrates how quickly AI hype can seep into otherwise sober policy thinking. The chapter recognises that “greater visibility” has brought “greater clarity on the nature of AI,” yet some of its citations and framings betray an overreliance on volatile benchmarks and fashionable narratives that may not age well, as in the case of the 2024 open‑model benchmarking graph from epoch.ai.

Three Layers of AI Hype: Concept, Economy, Institutions

For several years, I have argued that AI hype operates in three layers: conceptual overreach about what AI is, economic overstatement about what AI can do for growth and jobs, and institutional under-specification about how AI systems will actually be governed in practice. Chapter 14 tries, commendably, to move beyond the first two layers by grounding AI in sector-specific contexts and by admitting that employment adjustment in the professional, business and information services sector is “more nuanced” than the polarised debates on AI-driven job losses suggest. This is a welcome departure from the binary “doom or boom” discourse that often accompanies AI, and it shows a recognition that labour markets adjust through messy, path-dependent processes rather than clean graphs in consulting reports.

Open, Interoperable Systems: Commons or Branding?

Where the Survey is strongest is in its embrace of a bottom-up, multiple-sector-specific approach anchored in open and interoperable systems. This is not merely a technical design choice; it is a political-economic stance against enclosure, concentration and opacity in AI infrastructure. The idea of an India-specific repository, analogous to GitHub, for code, datasets and possibly models, is “fine” in itself, but its value will depend on whether it becomes a true commons that communities can meaningfully contribute to and fork from, or whether it ossifies into yet another government-branded platform with limited epistemic diversity.

Rethinking Expertise: Beyond Degrees and Titles

I also agree with the chapter’s emphasis that India needs talent capable of understanding “the algorithmic issues involved in building models” along with software engineering knowledge and hands-on skills. However, the more radical insight – which the Survey only briefly gestures towards – is the need to re-evaluate what counts as “work experience” and “formal education” in this ecosystem. If AI systems are increasingly shaped by iterative experimentation, community-led tooling and domain-specific embeddings, then rigid credential hierarchies will fail to recognise the actual sites where expertise is produced and maintained. A serious AI strategy for India must therefore problematise the university-centric model of expertise and create pathways for practitioners, tinkerers and domain specialists to participate in high-stakes AI work without being treated as second-class actors.

Sequencing Over Speed: An AI Economic Council Worth Testing

The proposed AI Economic Council is an interesting institutional experiment, especially given its stated commitment to “human primacy,” economic purpose, labour-market sensitivity by design and a preference for sequencing over speed. These principles, if taken seriously, amount to a rebuttal of the “move fast and break things” ethos that has defined much of global AI deployment. The idea that AI adoption should be phased in line with institutional readiness and skill pipelines acknowledges that capacity – legal, bureaucratic, technical and social – is not a side constraint but a core variable in determining the legitimacy and sustainability of AI interventions. In other words, the chapter implicitly recognises that AI is not a magic external shock to productivity, but a co-evolutionary process between technology and human capital.

Data as Asset, Data as Bargaining Power

On data, the Survey’s position is more ambivalent and therefore more revealing of the tensions in India’s AI imagination. It rightly notes that rigid data localisation mandates have been avoided and that policy must remain “cognizant of the potential value embedded in India’s data.” This framing avoids the simplistic “data is the new oil” metaphor, yet it continues to treat data primarily as a latent economic asset rather than as a site of contestation over rights, bargaining power and value-sharing. The discussion on data categorisation based on the DPDP Act, with separate treatment for large-scale behavioural, transactional and inferred datasets, is a step towards risk-sensitive governance – particularly when coupled with the idea of “value-retention” for higher-risk categories.

From Data Mirroring to Data Contracts

However, the pivot from this nuanced categorisation into a discussion on “incentivising localisation” through data mirroring exposes unresolved tensions. The justification offered is that regulatory oversight should not become ineffective simply because processing occurs offshore, which is a fair concern. Yet, instead of clearly articulating economic and contractual mechanisms for value-sharing when foreign or domestic firms exploit Indian data, the chapter retreats into infrastructural fixes like mirroring and proportional, threshold-specific oversight structures. I had hoped for more explicit thinking on how Indian individuals, communities and institutions could negotiate their interests in data-driven value chains, beyond being abstract “sources” of data whose value is realised elsewhere.

Who Really Benefits From Indian Data?

The Survey does, to its credit, acknowledge the “heterogeneity of actors” and proposes proportional regulation keyed to thresholds, which makes sense in a landscape ranging from startups to large multinationals. It also recognises the economic value of Indian data and lists contributions such as local training of models for sector-specific applications, building or supporting research labs and contributing to datasets. These are valuable starting points, but they still operate within a model where firms and labs are the primary sites of value creation, and citizens appear mainly as inputs. A fuller political economy of AI in India would ask how communities, public institutions and smaller entities can co-own, steward and govern the data and models that shape them.

Epistemic Accountability: Calling Out Evaluation Theatre

On accountability and safety, the chapter is refreshingly candid about the behaviour of large AI firms. It notes how big tech obfuscates its evaluation methods, hides its reasoning and provides dubious interpretations of results, with little publicly available evidence on implemented safeguards. This is a rare acknowledgment in an official document that informational asymmetry and corporate secrecy are not incidental bugs but core features of the current AI industry. The emphasis on transparency, responsible use, clear records of dataset provenance, standardised model documentation and detailed accounts of training limitations is, therefore, more than a procedural checklist; it is a demand for epistemic accountability.

Red-Teaming as Civic Practice, Not Compliance Ritual

The suggestion to institutionalise red-teaming of AI models, including through analogies to CRISPR and safety governance in biotechnology, is another important move. If implemented meaningfully, such red-teaming could help shift AI evaluation from marketing-led benchmarks to adversarial, context-aware stress tests that take social harms seriously. Yet, the success of these measures will hinge on who gets to define the threat models, who sits at the table in these red-teaming exercises, and whether dissenting voices – from civil society, labour, marginalised communities and independent researchers – have real influence.

Against Benchmarks and Buzzwords as Policy

Ultimately, Chapter 14’s most valuable contribution lies in its implicit rejection of uncritical AI hype while still insisting that AI matters for India’s economic trajectory. It treats AI neither as a silver bullet nor as a purely existential risk, but as a field where infrastructures, incentives, skills and governance must be designed together. To carry this agenda forward, India will need to deepen the economic and contractual imagination around data, democratise expertise in AI development and evaluation, and resist the temptation to measure progress through benchmarks and buzzwords alone. The real test will be whether the principles outlined – human primacy, labour-market sensitivity, sequencing over speed, and co-evolution of technology and human capital – can survive contact with political pressures, corporate lobbying and the enduring seduction of AI hype.