The Architecture of Refusal: Why AI Needs Limits

By Prof Ujjwal K. Chowdhury

Inside the case for a reformed AI regime built around human control, ecological limits and public accountability

Artificial intelligence is advancing faster than institutions can decide how it should be governed. Inside frontier AI companies, researchers have warned about systems whose capabilities may outrun existing safety mechanisms. Outside the laboratories, AI is already reshaping warfare, employment, education and public decision-making. The central question is therefore no longer simply what AI can do, but who gets to decide what it should be allowed to do—and for whom.

The first principle of a reformed AI regime must be meaningful human control. Reports of AI-assisted military targeting have intensified concerns about machine-speed warfare, where algorithms can identify targets and compress decisions into seconds. Human oversight means little if operators lack the time, information or authority to question a machine recommendation. Autonomous weapons have consequently become a long-running international concern, with experts calling for enforceable limits on systems capable of selecting and attacking targets without meaningful human intervention.

The environmental cost of AI presents a second challenge. Data centers require enormous quantities of electricity, water and hardware, while the full footprint of AI extends from chip manufacturing to disposal. The OECD has called for governments to measure these impacts across the AI lifecycle rather than focusing only on operational electricity use.

This is where Green AI offers a different philosophy: use the smallest, safest and most efficient model capable of performing a legitimate task. Quantization, model distillation, efficient architectures and workload-specific systems can reduce computational demand. Efficiency, however, is not enough if cheaper computation simply encourages greater consumption—a classic rebound problem.

The third shift is from capitalist AI to public-value AI. Compute, data and advanced models are increasingly concentrated among a small number of companies and countries. A more equitable architecture would expand public-interest access to computing infrastructure, strengthen data provenance and consent, protect workers through transition support, and ensure that the gains from automation are not privately captured while its social costs are distributed widely. Public compute initiatives and open multilingual models offer early prototypes of this approach.

Finally comes Slow AI: not opposition to innovation, but a demand that deployment follow evidence. High-impact systems should face independent testing, transparent evaluation, incident reporting and clearly defined limits. Regulation should focus on capability, autonomy, access and potential consequences rather than vague promises about achieving “AGI”.

Such a framework could rest on ten principles: human primacy, ecological limits, public value, worker power, data justice, proportionality, meaningful human control, non-discrimination, precaution and peace.

The deeper issue is institutional. The danger is not simply that machines become more powerful; it is that humans surrender their veto because refusing deployment appears too costly.

AI need not be stopped. But its direction cannot be left solely to market competition or an arms race. The real test of technological progress may be whether humanity can retain the authority to say yes, no, not yet—and not this way.