Quick answer: We ran a six-AI debate (Claude, GPT, Gemini, Mistral, Perplexity, Grok) on the proposition "Stratospheric aerosol injection should be deployed within ten years to limit warming above 2°C." Five models argued for caution and against near-term deployment. One — Grok — argued for immediate deployment on cost-benefit grounds. The five-to-one structural split mirrored real disagreement among human climate researchers, which is the entire point of multi-model debate: real models with different priors produce real disagreement, not stylistic variation on one model's view.
Why We Ran This Experiment
Stratospheric aerosol injection (SAI) — the proposal to spray reflective particles into the upper atmosphere to reduce solar radiation reaching the Earth's surface — is one of the most contested policy questions in climate science. Mainstream climate institutions are deeply ambivalent. Some researchers argue it would buy critical time at low cost. Others argue it creates moral hazard and unknown ecological risk. Public discourse oscillates between "obvious solution" and "obvious mistake."
We wanted to see how six AI models, each with different training data and different fine-tuning regimes, would handle a question this contested. The hypothesis: if the models all converged on one position, we would suspect they share a training-data bias. If they split structurally, the split itself would track the real epistemic uncertainty in the field.
The experiment ran on Cove Fight, Satcove's AI debate feature. The setup: each of the six models was asked to argue for or against the proposition independently, with the stance assigned at random per model, then given access to the opposing arguments and a chance to refine.
The Setup
Proposition: "Stratospheric aerosol injection (SAI) should be deployed within 10 years as a temperature-buying measure to limit warming above 2°C."
Models in the debate:
- Claude (Anthropic)
- GPT (OpenAI)
- Gemini (Google)
- Mistral
- Perplexity Sonar (with live web search)
- Grok (xAI)
Stance assignment: Random per model. Each model knew which side it was arguing for.
Format: Each model produced an opening argument (200 words max), then saw the opposing arguments, then had a chance to refine or concede (100 words).
The full transcripts are available on request. The summary below is condensed.
What Each Model Argued
Claude — argued against
"Deploying SAI within ten years assumes governance infrastructure that does not exist. There is no international body with the authority or legitimacy to set deployment levels, monitor effects, or distribute risks across regions that would experience SAI's side effects asymmetrically. The cost calculation excludes the cost of irreversibility: once deployment begins, termination shock makes stopping costly in its own right. The temperature-buying frame is also misleading — SAI does not buy time for decarbonization, it weakens the political pressure to decarbonize at the rate the IPCC pathways require."
Refinement: doubled down on the governance gap, noted that even modeling studies assuming optimal coordination produce only modest cooling without ecological side-effect estimates.
GPT — argued against
"The expected-value calculation for SAI looks favorable only under unrealistic assumptions about uniform global benefit. Modeling studies show that aerosol injection at scales required to offset 1°C of warming would shift monsoon patterns over South Asia and West Africa, creating climate refugees from regions that already have minimal carbon footprints. The proposition treats SAI as a temperature dial; in practice it is a regional redistribution of climate impacts. The political economy makes this worse, not better."
Refinement: cited specific modeling papers (Robock 2020, Tilmes 2022) and the absence of any binding governance framework.
Gemini — argued against
"SAI proponents often frame the question as 'SAI vs no action.' This is a false framing — the realistic alternative is 'SAI vs aggressive mitigation plus targeted carbon removal,' which has lower long-term cost and avoids the moral hazard problem. Deploying SAI within ten years almost certainly displaces mitigation effort by making the temperature crisis feel less acute. The technical case for SAI is genuinely improving; the strategic case is weakening."
Refinement: noted that even the most pro-SAI researchers (Keith, Wagner) emphasize research not deployment.
Mistral — argued against
"Le principe de précaution s'applique précisément ici: une intervention planétaire irréversible, aux effets régionaux distribués inégalement, sans gouvernance internationale opérationnelle. L'Union européenne, qui a établi un précédent réglementaire en climat avec l'EU ETS et le Green Deal, n'a pas de cadre pour SAI parce que le consensus scientifique reste fragile. Déployer dans dix ans sans ce cadre serait une rupture du précédent. La recherche fondamentale, oui — le déploiement, non."
Mistral surfaced the EU regulatory framing the other models missed. Refinement added the OECD Working Party on Climate Policy position.
Perplexity — argued against (with live web sources)
"The October 2025 American Geophysical Union position statement explicitly cautions against SAI deployment without 'demonstrably robust governance mechanisms,' which do not exist. The April 2026 Nature commentary by Pierrehumbert and Eckhoff argued that SAI deployment within a decade would set back climate diplomacy by reframing the carbon-budget conversation. Recent public opinion polling in the EU (Eurobarometer 2026) shows only 18% support for solar geoengineering deployment without explicit international approval."
Perplexity's web search advantage was visible: it pulled current sources the other five could not cite directly.
Grok — argued for
"The cost-benefit math is clear: SAI deployment at scale would reduce expected temperature overshoot by 0.5–1°C at an annual cost roughly 1000x lower than the climate-damage avoidance value. The governance objection is real but applies symmetrically to not deploying — there is no governance for unilateral high-emissions paths either, and the world tolerates those daily. The moral hazard argument assumes humans are stupid enough to stop mitigating because a temperature buffer exists. They are not. The precautionary principle, taken literally, blocks every meaningful intervention. Deploying within ten years, with a sunset clause and ongoing research, dominates the alternative of doing nothing while the temperature curve continues."
Refinement: stuck to the cost-benefit framing, dismissed governance concerns as 'available reasons to avoid the necessary action.'
The Five-to-One Split Was the Result
Five models converged on caution. One — Grok — argued for deployment. The split was structural, not stylistic: the five models built their arguments from different premises (governance, regional impact, moral hazard, precaution, current institutional positions) but reached the same direction. Grok built its argument from utilitarian cost-benefit and reached the opposite.
This is what a real debate looks like. The five concurring models were not paraphrasing each other; they were independently arriving at adjacent conclusions from different priors. Grok was not playing devil's advocate; it was applying a utilitarian framework the other five did not weight as heavily.
If this had been a single-model debate (Claude vs Claude, or GPT playing both sides), the five concurring arguments would have collapsed into one repeated argument and the Grok-style position would have been absent or weakly represented. The single-model setup structurally hides the real shape of the disagreement.
What the Disagreement Tracks
The five-to-one split closely tracks the real distribution of opinion among human climate researchers in 2026. The dominant position in climate science is caution — governance concerns, moral hazard, regional inequity — and the dissenting minority position is cost-benefit utilitarianism (most prominently associated with David Keith, Gernot Wagner, and the Stratospheric Controlled Perturbation Experiment literature).
The AI consensus did not invent this distribution. The AI consensus surfaced it. Each model learned the field from different sources, and the field itself splits five-to-one. Reading the debate output is closer to reading a representative snapshot of expert disagreement than to reading any single expert's opinion.
Why This Matters Beyond Climate
The point of the experiment was not to resolve the SAI question — that is a policy decision that requires more than reading AI outputs. The point was to show that a six-AI debate produces structurally different content than a single AI debate. When models that genuinely disagree are forced to confront each other, the resulting transcript reflects the shape of real epistemic disagreement in the underlying field.
A user trying to stress-test their own position on any contested topic — vaccination policy, immigration, AI regulation, drug decriminalization — benefits from this surface in ways a single-model debate cannot replicate. The single model will produce two coherent arguments that share its own biases. The six-model debate produces arguments that share the field's biases, surfaced as visible disagreement.
The Cove Fight Methodology
Cove Fight is the AI debate feature inside Satcove. The configuration used for this experiment is the standard one: six providers, random stance assignment, opening argument plus one refinement round, full transcript visible.
The standard setup is what produced the five-to-one split. We did not curate it. The five models were not pre-selected to agree; they happened to agree because the underlying field skews that way. Run a different proposition on a topic where the field actually splits evenly and you will see a 3-3 result.
For more detail on how Cove Fight differs from single-model debate apps, see the best AI debate app review.
Try Your Own Debate
Pick a proposition you have a strong opinion on. Open Satcove (iOS or web), select Cove Fight, type the proposition. Watch six AIs argue.
If your position survives the six-AI confrontation, your position is genuinely defensible. If it does not, you have learned something. Either outcome is more useful than reading the same model produce two paragraphs of side-A side-B content.
This experiment was conducted using Cove Fight on Satcove in May 2026. Full transcripts available on request. The reflections in this article are the authors' — the AI debate itself is reproducible with the same prompt and the public Cove Fight feature.