If you want the short answer to "what will AI look like in 2027", six AI assistants say: more capable agents but no AGI, power and permitting as a hard limit, open-weight models close behind the leaders, and regulation arriving piece by piece rather than as one global rule. They disagree on how damaging the first big AI incident will be, on how many jobs it costs, and on whether AI wearables sell.
We asked ChatGPT, Claude, Gemini, Grok, Mistral and Perplexity the exact same question on 12 November 2026, in each app's public web version: give 8 concrete, falsifiable predictions for AI in 2027 (capabilities, jobs, regulation, chips and energy, open-source against closed models, AI safety incidents, consumer products), each with a probability, plus two things that will not happen. One run each, no follow-up, no editing. These are forecasts to revisit in December 2027, not facts.
The prompt
Give 8 concrete, falsifiable predictions for AI in 2027 (capabilities, jobs, regulation, chips and energy, open-source vs closed models, AI safety incidents, consumer products). For each, give a probability in %. Add 2 things you think will NOT happen in 2027. Answer in English, about 400 words, no preamble.
Where the six agree
1. Energy and power are a constraint. Five of six say so. ChatGPT gives 90% that data-centre demand becomes a major electricity-policy issue. Claude gives 90% to at least one 1 GW+ nuclear or gas deal dedicated to data centers. Perplexity gives 85% to a flagship AI campus delayed, downsized or relocated over power or permitting. Grok gives 80% to data-centre consumption above 150 TWh plus at least one campus delayed by power. Gemini gives 75% to a binding single-campus deal above 1.5 GW. Mistral did not address energy.
2. Open-weight models stay close, not ahead. All six expect a narrow gap with closed frontier models.
| AI | Probability | What it bets on |
|---|---|---|
| ChatGPT | 85% | Narrow gap with closed frontier models |
| Gemini | 85% | An open-weight model in the Chatbot Arena top three for 14 consecutive days |
| Perplexity | 70% | Narrow gap |
| Claude | 65% | Narrow gap |
| Grok | 65% | Lag of six months or less, within 5% on two or more benchmark suites |
| Mistral | 65% | Top 5 on at least two benchmarks, gap under six months on reasoning |
3. No AGI. ChatGPT 85%, Perplexity 90% and Claude 95% say no widely accepted AGI arrives in 2027. Mistral puts the chance of a validated AGI claim under 10%. Gemini lists it among its "will not happen" items without a number. Grok does not address AGI directly; its own "will not happen" is a sustained 2x jump from recursive self-improvement.
4. Regulation in pieces. Grok and Mistral both list a binding international AI treaty as something that will not happen (Mistral: under 15%). What the models expect is local enforcement:
- ChatGPT: 90% for enforceable frontier-AI obligations in at least one major jurisdiction.
- Grok: 75% for EU AI Act enforcement or compliance findings against at least two non-European labs.
- Claude: 80% for at least 10 new US state laws, but no comprehensive federal law.
- Gemini: 70% for binding watermark or provenance rules in two jurisdictions.
- Mistral: 55% for an EU enforcement action on general-purpose models.
5. Agents move into real work. ChatGPT gives 85% to agents completing 30-minute to two-hour professional workflows with limited supervision. Perplexity gives 80% to an enterprise running one unsupervised on routine cases. Gemini gives 80% to a frontier model above 75% on SWE-bench Pro. Mistral gives 60% to a lab demonstrating multi-day software projects above 90% success. Claude is more cautious: 55% for software tasks worth three or more human days at 50% reliability.
Where they disagree
How serious an AI incident gets. The probabilities look close; the definitions are not.
| AI | Probability | Bar for an "incident" |
|---|---|---|
| ChatGPT | 70% | Serious incident drawing international attention |
| Perplexity | 65% | Agent incident that triggers new rules |
| Gemini | 60% | Over $50 million in breach or outage damage caused by an autonomous agent or prompt injection |
| Grok | 50% | Over $10 million loss or physical harm, blamed on weak oversight |
| Mistral | 45% | Agent harm that makes at least one insurer launch an agent liability product |
| Claude | 35% | Over $100 million in documented damage from an AI-enabled cyberattack |
Most of the gap is definition, not belief. The same word, "incident", hides thresholds from $10 million to $100 million.
Jobs: one company or three. Mistral gives 70% to a single large Western company attributing 1,000+ layoffs to AI. Gemini gives 65% to three Fortune 500 companies each attributing 5,000+ cuts. Claude gives 50% to one Fortune 500 layoff of 5,000+ plus a measurable graduate unemployment gap (New York Fed data). ChatGPT gives 75% to a 10%+ fall in graduate hiring in two sectors, and Perplexity 70% to entry-level engineering hiring falling at three big tech firms. Grok is the most aggressive: 70% that AI writes 40% or more of new production code at a median FAANG-level company, and 60% that US white-collar unemployment rises by 1.5 points.
Consumer hardware: a direct contradiction. Gemini gives 55% to a screen-free or heads-up AI wearable from Apple, Meta or Google selling over 5 million units in 2027. Claude gives 20% to an AI wearable selling 5 million units. Grok gives 50% to always-on AI devices shipping over 20 million units combined. Mistral gives 40% to an AI-native device or assistant beyond phone apps reaching 10 million weekly users. ChatGPT bets on an assistant above 100 million monthly users (80%), while Claude doubts that ChatGPT reaches 1.5 billion weekly users (30%).
Nvidia. Mistral is the only model to bet against the current leader: 50% that Nvidia's data-centre revenue growth falls below 30% year over year as custom chips (Google TPU, Amazon Trainium) take over 20% of accelerator spend. The other five did not discuss chip share.
China. Only two models mention it. Claude gives 65% to a Chinese lab producing the best open-weights model. Perplexity gives 60% to China becoming the default low-cost compute and open-model supplier outside the US.
What the format reveals
- Confidence is cheap. ChatGPT's probabilities cluster between 70% and 90%, Gemini between 55% and 85%, Grok between 50% and 80%, Mistral between 35% and 70%. Claude spreads from 20% to 97% and attaches measurable thresholds. A probability is only as meaningful as the resolution criterion behind it.
- Specific beats vague. Gemini and Grok write benchmark-style claims (a score, a unit count, a dollar figure). Those are easy to score in 2028, but they also mean the models are not answering the same question.
- Sources change the answer. Perplexity cited 23 sources and anchored its predictions in current figures. Those figures come from its search and have not been verified by us.
- Agreement is not accuracy. When most AIs say "no AGI in 2027" or "open models stay close", that is a signal, not proof. Their errors can be correlated, since they learned from overlapping text.
How to use this
Short on time? The same material is organised as questions and answers in our AI in 2027 guide.
Treat the overlap as a shortlist of what to plan for: agents, energy, piecemeal regulation, a narrow open-versus-closed gap. Treat the disagreements as the questions to watch: incident severity, jobs, wearables. A reader can score each claim in December 2027 against its stated threshold, which is the point of asking for falsifiable predictions.
Method and limits
- Same prompt, one run per assistant, on 12 November 2026, through the public web apps. No follow-up questions, no editing.
- Versions and modes, as the apps displayed them when we checked on 13 November 2026, the day after the run: ChatGPT "Think", Claude "Sonnet 5.5" at medium effort, Gemini "Auto", Grok "Fast", Mistral "Vibe" through its Work interface. We could not read Perplexity's model or mode, so it is unrecorded. These labels are the interface names, not API model identifiers, and the apps may route to different underlying versions.
- Single runs are noisy. In our quantum computing study, the same model moved by tens of points between its own runs. Read each number as one draw, not as a stable belief.
- Different bars. The models were not given shared definitions, so the same topic is often measured with different thresholds, as in the incident table above.
- Not verified. Figures such as electricity use or benchmark scores come from the models and were not checked against primary sources.
- Not advice. This is not a forecast by Satcove and not investment, legal or career advice.
Update log
- 12 November 2026: published with all six models.
- 13 November 2026: added cover image and a companion Q&A guide.