Artificial intelligence might help improve energy systems, but building and running AI also leans on electricity grids, boosts demand for data centers, and adds to hardware waste. That trade-off is now right in the middle of the climate debate. Ahead of the United Nations climate summit COP31 in Antalya, Turkey, the Turkish government previewed a pledge aimed at cleaner energy, stronger environmental reporting, and more responsible AI development.
Those are sensible goals. The real question is whether a pledge like this can actually make them happen. The text describes the aims as voluntary and non-prescriptive, with no new obligations. That matters because a list of good intentions isn't the same thing as a mechanism that changes how companies design data centers or how governments plan electricity supply.
This article looks at what the pledge proposes, where voluntary commitments often fall short, and what a more credible approach could require. It also has to deal with a harder truth about the AI industry: AI can cut emissions in some settings while increasing them elsewhere - sometimes, for example, by accelerating fossil fuel production. The practical test isn't whether AI can be described as "green." It's whether its full costs and benefits are measured and acted on.
What does the COP31 AI climate pledge propose?
Turkey previewed its Antalya Artificial Intelligence Pledge ahead of COP31, inviting other countries to endorse it before its planned launch on November 16. The pledge brings AI infrastructure into climate discussions that have often treated digital services as weightless. They aren't. AI runs on physical data centers, computing hardware, and electricity networks.
The pledge covers a few areas that do matter: it calls for data centers to rely increasingly on clean energy, encourages internationally recognized standards for responsible AI development, and points to reporting requirements that could make environmental impacts harder to ignore. It also acknowledges the e-waste tied to the generative AI boom.
These categories map onto different parts of the problem. Clean electricity affects the power used to run facilities. Reporting affects whether governments and communities can see how demand and emissions are changing. And hardware management matters because chips, servers, and other equipment involve materials in the first place - and eventually need repair, reuse, or disposal.
Good goals need operational definitions
Where this gets thin is in the phrasing. Broad language leaves too much room for interpretation. "Increasingly powered by clean energy," for instance, doesn't say how "clean" is counted, which emissions are included, or what time horizon applies. A company could buy some renewables on paper while still drawing electricity from a fossil-heavy grid during the hours a data center is actually running.
The same problem shows up with transparency. Reporting helps only if it's consistent enough to compare and detailed enough to support decisions. A disclosure that bundles multiple facilities - or doesn't clearly state the electricity source - can technically satisfy a reporting exercise while still being too fuzzy for a local grid operator evaluating a new data center.
For e-waste, the pledge also needs clarity about what "responsible management" means in practice. Extending equipment life, repairing parts, and tracking where retired servers go are not the same as simply recording that equipment was disposed of.
In other words: naming the right subjects is a start. A pledge becomes usable when it sets a baseline, a measurement approach, a deadline, and consequences for missing the target. Without those pieces, different signatories can claim progress while doing very different things.
Why might a voluntary pledge fall short?
The pledge says its objectives are voluntary and non-prescriptive, creates no new obligations, and will be implemented according to national circumstances. That wording is straightforward. Countries can support the principles without agreeing to enforceable targets, shared reporting rules, or changes to domestic policy.
Voluntary pledges aren't automatically pointless. They can help align language, signal political interest, and make it easier for governments to coordinate later. A public commitment can also give civil society and investors a reference point for asking questions. But all of that depends on what comes next. Without a reporting cycle, independent review, or follow-through in policy, a pledge can end up as just another statement of intent.
The incentives do not change on their own
Data centers are expensive and long-lived. Companies choose locations based on electricity availability, cost, network connectivity, and permitting. Utilities and governments then get pressure to meet the demand - sometimes by expanding power infrastructure. A general pledge doesn't automatically rewrite those investment decisions.
What you end up with is a potential gap between what's promised publicly and how operations actually work. A company might endorse clean energy in principle while pushing rapid capacity growth in a region where additional low-carbon power isn't available. A government might support transparent reporting without requiring facilities to disclose consistent, site-level information. The incentive to build quickly stays in place, and the pledge may not add a real counterweight.
International participation matters here, too. The source reporting on the pledge notes that the United States has more data centers than any other country, but the Trump administration had withdrawn the U.S. from the global climate accord, and the U.S. had no official delegates at the summit in Turkey. A pledge without buy-in from a major market can still influence discussion - but it has less leverage over the infrastructure and companies that shape the sector.
A useful test for any pledge
When you evaluate a voluntary commitment, ask four questions:
- Who is accountable? A government, a company, a utility, or a data-center operator has to own the action.
- What is measured? The scope should cover the emissions, electricity use, and hardware impacts that matter, not just whatever is easiest to report.
- When is progress due? A target without a timeline is hard to separate from a preference or aspiration.
- Who checks the result? Independent verification and public reporting make claims more credible.
If the answers to these questions are fuzzy, the pledge may still have diplomatic value. But it won't yet function as a reliable way to reduce environmental impact. That's the difference between getting attention for an issue and changing the system that produces the impact.
Does AI reduce emissions or add to them?
It depends on the use case, the energy behind the computing, and what happens beyond the immediate application. AI can help forecast renewable generation, improve building controls, detect equipment faults, and optimize some industrial processes. Those could reduce waste or make existing energy systems more efficient. But a potential benefit is not the same as a confirmed net climate gain.
A useful assessment compares the emissions avoided with the emissions created by developing, training, and running the system. It also asks what would have happened without the AI tool. If a simpler software system, a process change, or better measurement could deliver the same result, then the added compute probably isn't worth it.
Count the whole system, not the model alone
A data center's environmental footprint doesn't start and end with electricity used during model training. Ongoing inference, facility operations, hardware manufacturing, and hardware replacement all add up. Depending on where the data center is and how it's designed, water use and local effects of electricity demand may also need attention. Looking only at a model's efficiency per task can miss the obvious reality that total demand rises when more tasks are added.
That's why efficiency claims need a clear denominator. A more efficient model may use less energy per request - but if the number of requests grows enough, total energy use can still rise. That doesn't make efficiency irrelevant. It means you track efficiency alongside total consumption and capacity growth, not instead of it.
This is also why indirect effects matter. Research published in a Nature journal in August found that AI can increase carbon emissions by boosting fossil fuel production more than it helps avoid through gains in renewable energy. That challenges the idea that "useful climate applications" automatically outweigh AI's broader consequences. It doesn't prove every AI project increases emissions. It does show why the net effect has to be examined rather than assumed.
Use a practical decision test
For any proposed AI climate application, ask:
- What measurable outcome is supposed to improve? Specify the emissions, energy, or resource use you expect to change.
- What is the baseline? Compare the AI system against the current process and plausible non-AI alternatives.
- What are the full operating costs? Include computing, hardware, electricity sources, and relevant local impacts.
- Is the benefit additional and durable? A short-term efficiency gain might disappear if demand rises or if the energy supply changes.
This is a business decision as much as an environmental one. If AI delivers a measurable improvement at a reasonable cost, it may be worth deploying. If the climate claim depends on assumptions that can't be verified, the organization shouldn't market the project as an emissions solution.
What would make an AI climate commitment credible?
A stronger commitment would connect the pledge's broad aims to rules that can actually be implemented and checked. You don't need every country to adopt identical energy policy, but you do need enough consistency for governments, companies, and communities to compare performance and spot where impacts are getting worse.
Set comparable reporting requirements
Governments could define a common minimum disclosure for large data centers, with reporting detailed enough to show local impacts. Helpful fields might include electricity consumption, the sources and timing of electricity supply, relevant emissions, water use where applicable, and what happens to hardware through reuse or disposal. The exact requirements should fit what data is available - but a shared baseline matters more than a long list of optional metrics.
Reporting should also distinguish between a company-wide renewable-energy claim and what electricity a specific facility actually draws. Those are not interchangeable. Clear definitions make it harder to present procurement choices as proof that a data center runs on clean power all the time.
Tie targets to infrastructure decisions
Targets need dates and accountable owners. A country could set conditions for new data-center permits, require facilities to disclose demand forecasts, or align expansion plans with grid capacity and clean-energy development. That will involve trade-offs: strict conditions can slow projects or raise costs, while weak conditions can shift the burden of new infrastructure onto communities and electricity customers.
The right policy depends on local grid conditions. A region with abundant low-carbon supply faces a different challenge than one where new demand risks locking in fossil infrastructure longer. "National circumstances" may justify different paths - but it shouldn't become an excuse to avoid measurement altogether.
Verify claims and track rebound effects
Independent review can test whether reported outcomes match the methods used. Public reporting gives regulators, researchers, and local communities a way to scrutinize plans. At the company level, teams should track total resource use as well as efficiency per unit of computation. Otherwise, lower costs per task can create enough additional demand to erase the savings.
A credible commitment also separates AI's climate applications from AI's own footprint. An organization shouldn't claim its AI supports renewable energy without disclosing the material costs of developing and operating the system. Both sides belong in the same assessment.
Conclusion: judge the pledge by what changes
The COP31 AI climate pledge points to real issues: data-center energy, environmental transparency, and hardware waste. The problem isn't that the goals are irrelevant. It's that voluntary principles - without shared measures, deadlines, or accountability - may leave the core decisions driving environmental impact largely unchanged.
A serious test starts after the announcement. Look for comparable disclosures, actual policy changes, verified progress, and evidence that AI applications deliver net climate benefits compared with realistic alternatives. Until those mechanisms exist, this pledge is best read as a starting point for coordination - not proof that AI is becoming environmentally sustainable.