Commentary: Artificial intelligence is being sold as both a climate risk and a climate solution. The contradiction is real, but it is not a reason for paralysis. It is a reason to demand rules that measure the costs before celebrating the benefits.
That argument matters as governments, companies and international institutions try to place AI inside climate policy. A voluntary pledge associated with the COP31 summit in Türkiye highlights potential uses including emissions monitoring, efficiency improvements and climate modelling, while also raising the prospect of environmental-impact disclosure and cleaner power for data centres. The Japan Times reports that the initiative reflects an obvious tension: AI may help address warming even as the infrastructure supporting it increases energy demand and fossil-fuel pollution.
Our view is straightforward: climate claims about AI should be treated as unproven until they are accompanied by comparable, independently verifiable data. The industry should not be allowed to count hypothetical efficiencies while obscuring electricity use, water consumption, supply-chain emissions or the rebound effects that follow when cheaper computation increases demand.
Disclosure is the minimum
California offers a useful, if unsettled, example. Its climate-disclosure law is intended to require large companies doing business in the state to report direct, purchased-energy and broader value-chain emissions. The first deadlines and regulations remain subject to changes, including a proposed shift in timing, but the principle is clear: companies cannot manage what they refuse to measure. An industry analysis of California’s rules describes the particular difficulty of gathering emissions data across the AI supply chain.
That difficulty is not an excuse to weaken the requirement. It is an argument for better standards. Regulators should require data-centre operators and major AI developers to disclose energy use per service, the carbon intensity of the electricity consumed, water impacts where material, and the boundaries of any claimed emissions savings. Audits should distinguish reductions that would have happened anyway from those caused by an AI system.
The counterargument deserves respect. Smaller companies may lack the money and staff to satisfy complex reporting obligations. Rigid rules could also favour the largest technology firms, which can absorb compliance costs and lobby for standards written around their own systems. Policymakers should therefore use phased thresholds, shared reporting templates and public infrastructure for verification rather than imposing identical burdens on every startup.
Voluntary promises are not enough
Voluntary pledges can move faster than treaties and may encourage experimentation. They can also create political momentum: companies and governments that publicly promise cleaner power or greater transparency may become easier to hold accountable. But voluntary schemes tend to reward the participants most comfortable with publicity, not necessarily those with the strongest results.
The wider digital-rights debate offers a warning. Freedom House’s 2026 assessment says internet freedom declined for a 16th consecutive year and identifies censorship, surveillance and AI-related threats as part of a worsening global environment. Its report calls for stronger safeguards against the misuse of technology, including surveillance of journalists, dissidents and rights defenders. Climate policy cannot be separated from that governance question. A system that optimises energy use but expands unaccountable surveillance is not automatically socially beneficial.
Nor should climate regulation become a back door for censorship or industrial protectionism. Rules must be technology-neutral, transparent and contestable. Governments should publish the evidence behind public-sector AI deployments, protect researchers who test environmental claims, and avoid treating national champions as exempt from scrutiny.
Make the promise testable
The Inter-Parliamentary Union has urged lawmakers to establish risk assessments, safety requirements, liability rules and serious-incident reporting for advanced AI. Those recommendations point toward the missing ingredient in today’s climate conversation: democratic oversight.
AI may indeed improve electricity grids, detect methane leaks, forecast extreme weather and accelerate scientific discovery. Those are worthwhile goals. But optimism should be earned through results, not substituted for them. Governments should tie subsidies, procurement contracts and favourable regulation to independently audited performance. Companies should disclose failures as readily as successes. Citizens should be able to see whether a promised climate benefit is measurable, additional and durable.
The choice is not between embracing AI and rejecting it. The choice is between governing it and marketing it. If AI is to become part of the climate solution, its environmental footprint must be visible, its claims testable and its power accountable. Anything less turns a potentially useful tool into another source of avoidable emissions—and another promise made without consequences.