AI Predictions for 2027: A Skeptic's Scorecard (Sexbots and Super Intelligence Included)

The forecast everyone cites

The most-quoted 2027 forecast is AI 2027, a scenario published in April 2025 by the AI Futures Project, in which AGI arrives in 2027 and superintelligence follows months later. It is a scenario, not a promise. The authors never claimed everything in it would happen that year.

What is interesting is what happened to the authors. FutureSearch’s six-month review collects the revisions:

• Daniel Kokotajlo moved his median for AGI from 2028 to 2029, citing “improved timelines models & slightly slower than expected progress in general”.
• Eli Lifland’s median is roughly 2032, with about a 15-20% chance of AGI by 2027.
• Andrej Karpathy argued against a coding-agent takeoff as soon as 2027, naming the bottleneck as “actually making it work”. He also said his timelines are “5-10X pessimistic” relative to San Francisco AI circles, but still optimistic relative to AI skeptics.

So: a scenario named 2027 now has authors whose own medians are 2029 to 2032. The only thing that arrived on schedule was the title.

The same review points at the labs. xAI is focused on Grok for the X algorithm and “sexy companions”. Meta’s public focus is companions, ads and “superintelligence” messaging. OpenAI has shifted toward consumer products like the Sora app and shopping. The review’s heading for this section is the best line in the genre: commercial success may trump the race to AGI.

Hold that thought. We will come back to the companions.

What the State of AI Report predicts for 2027

The State of AI Report’s 2027 predictions are more concrete and, helpfully, more boring. A selection, with our scepticism grade:

That last row matters. The report gives no definition of AGI, so the prediction cannot fail. In its fairness, the same report scored its own previous year: 2 hits, 5 partial outcomes and 3 misses, with its agentic-commerce call among the misses. That is a respectable batting average for a field where the standard method is to predict the future and hope nobody scrolls back.

For the agent-payments angle, see our guide on agentic commerce and AI shopping agents.

AI vs SI: a field guide for 2027

Since 29 September 2026, US executive-branch agencies are told to say “Super Intelligence” or “SI” and to no longer acknowledge the term “AI”. The stated reason, as quoted by PolitiFact: “It’s not artificial. We all agree on that.” The science adviser must propose a federal definition, so by 2027 SI will have one.

Here is how the vocabulary now stacks up.

Field notes for the cautious reader:

• If a model is wrong 3% of the time, call it SI. If it is wrong 30% of the time, call it SI Preview.
• Superintelligence, per the industry, is smarter than every human at nearly everything. Super Intelligence, per the decree, is whatever the model was last week. Both need more power.
• Expect at least one vendor to put “SI-ready” on a chatbot that cannot reliably count the r’s in “strawberry”.
• Forecast for 2028: if SI proves embarrassing, the next rename is already waiting. “Ultra Intelligence” is available at the time of writing.

Our serious reading is simple. A renamed technology is the same technology. It still hallucinates, as KPMG’s retracted agentic-AI report demonstrated with 5 of 45 citations accurate. It still needs verification. Whichever name your vendor uses, ask for the error rate.

Our ten predictions for 2027

Each is dated, falsifiable and carries a probability. Revisit this page in January 2028 and mark us down.

1. A federal definition of SI is published and is broader than any industry definition (70%). The order requires a proposal. Committees round up.
2. A second frontier lab files publicly and “run-rate” disappears from its headline numbers (65%). A listed company has to report audited revenue, not slide decks—Anthropic’s S-1 is the first example, with OpenAI reportedly eyeing 2027. Audited does not mean profitable: the same filing still leaves partner gross versus net and a $498 billion compute bill on the table. See the $50 billion versus $68 billion gap for why the ruler still matters.
3. Another Big Four-style AI report is retracted for fake citations (80%). Deloitte (2025), EY and KPMG (2026) set the base rate.
4. A consulting or government contract adds an explicit “AI-use disclosure and citation audit” clause (75%). Procurement is how boredom wins.
5. A second mega-cap offers mass voluntary exits framed as capital efficiency (55%). Microsoft’s 60,000-person offer is the template, and its $45 billion quarter of capex is the motive.
6. The first ruling or settlement against an AI vendor over generated deepfakes (60%). The Grok cases have no verdicts or public settlements as of writing.
7. A card network publishes a liability rule for AI-agent purchases (50%). It is State of AI’s number one, and we only half believe it.
8. A federal SI Force produces a logo before a charter (90%). We will update if wrong.
9. Someone claims AGI and no one agrees on what that means (95%). Achieved by non-definition.
10. The sexbot prediction (see below).

The sexbot prediction

Here is the part you scrolled for.

The prediction: the first AI product category to deliver on its promises at scale is not superintelligence. It is companionship. The evidence is more reliable than most forecasts, because it is already shipping.

• OpenAI said in December 2025 that ChatGPT would add an “adult mode” in the first quarter of 2026. Age verification slipped, the launch was pushed back again, and age prediction accuracy remained a problem. The race to superintelligence may be decided by whichever lab first tells a 40-year-old from a 14-year-old.
• xAI and Meta are both pursuing companions, per the FutureSearch review above. Commercial success may trump the race to AGI.
• On hardware, Realbotix exhibited at CES 2026 with a range of AI-powered humanoids, including a male companion robot called David, and planned 19 humanoid deliveries between March and May 2026. Those are company press releases, so treat the claims as marketing. The delivery schedule, though, is a real date with a real number.

The joke writes itself, so here it is plainly: AGI has been five years away since roughly 2020. The companion robot has a delivery schedule. One of these industries has a working go-to-market, and it is not the one with the mission statement.

Our probability: 85% that at least one major lab ships or relaunches an adult or companion mode in 2027, and 0% that anyone calls it “SI”. We will not be taking questions about the supply chain.

Planning around forecasts, not adjectives

If you are making 2027 decisions, treat forecasts like vendor demos:

• Prefer dated, falsifiable claims. “AGI 2027” cannot fail. “A card network publishes a liability rule by December 2027” can. Plan around the second kind.
• Budget for verification. Whatever the label, outputs need checks. Our durable AI value versus speculation piece covers where value survives the hype.
• Price vendor risk. If your model provider lists in 2027, its pricing logic changes with quarterly targets. Keep prompts, evals and data portable, as outlined in the bubble survival playbook.
• Do not buy the noun. A rename adds no capability. If it does not appear in the contract’s performance clause, it does not exist.
• Keep a scorecard. Write your own predictions in a doc, with a date, and revisit them. It is the cheapest strategy tool there is.

Bottom line

By 2027, the forecasters will have moved their medians again, the government will have a definition of SI, at least one frontier lab will know what a quarterly earnings call feels like—and those calls will also disclose what the roadshow slid past, from net losses to noncancelable compute—and someone will have sold a very expensive chatbot with an extra adjective.

The useful forecast is the boring one. Verification becomes the product, liability moves toward whoever ships the output, and the first category to prove its business model is the one that sells companionship, not superintelligence. Plan for that, ignore the adjectives, and revisit this page in January 2028.