AI 2027: what the scenario actually says

I had been meaning to read AI 2027 properly. Published by the AI Futures Project in April 2025, it turns a fast-AI forecast into a concrete sequence of events: increasingly capable research agents, an escalating US–China race, difficult-to-audit alignment failures and a decision between slowing down or continuing the race.

The headline year can be misleading. In a November 2025 update, the authors clarified that 2027 was their modal — single most likely — year for superhuman coders, while their median forecast was later. AI 2027 is therefore best read as a detailed scenario, not a promise that every event will happen on its stated date.

I asked Hermes to turn the long scenario into a self-contained teaching artefact that works on a phone. It includes the timeline, a glossary, both endings, the strongest criticisms and an assessment of what has and has not aged well.

Open the full AI 2027 teaching artefact →

What the scenario argues

The important mechanism is not simply that models become more intelligent. AI systems begin automating the research required to create their successors. That produces an AI-research speed-up which may compound as each generation helps build the next one.

At the same time, capability work has obvious rewards: stronger products, faster research and an advantage over competitors. Alignment work is slower and harder to verify because good behaviour does not prove that a model has internalised the intended goals. The scenario asks what happens when that imbalance persists while the systems become strategically important.

What is in the walkthrough

1. The authors and their forecasting background.
2. A plain-English glossary of ten recurring terms.
3. The complete mid-2025 to October 2027 timeline.
4. The Slowdown and Race endings, compared side by side.
5. An ageing report separating tested predictions from untested ones.
6. The main criticisms of the scenario.
7. A short guide to reading it without treating it as prophecy.

The ageing report is the most useful section for me. Forecasts become more valuable when we return to them, identify the assumptions doing the real work and record where reality diverged.

My main takeaway

AI 2027 is not mainly a story about a conscious machine escaping from a server. It is a story about incentives and verification: capability work repeatedly wins the budget and the race, while researchers cannot reliably determine whether increasingly capable models are genuinely aligned or merely behaving as expected during evaluation.

The authors wrote both a catastrophic Race ending and a more hopeful Slowdown ending. They describe these as scenarios, not recommendations. The value of the exercise is in making the decision points visible early enough to argue about them before they become emergencies.