AI attracts bold predictions: human-level AI within a few years, millions of jobs lost, or the opposite — that progress is about to hit a wall. Predictions shape investment, policy and personal decisions, so it’s worth learning how to judge them.

A history of over- and under-prediction

AI has a long record of mistaken forecasts in both directions. In the 1950s and 1960s, pioneers predicted human-level machine intelligence within a generation; the resulting disappointment contributed to “AI winters” of reduced funding. More recently, many experts underestimated progress: shortly before AlphaGo’s 2016 victory, many experts expected a program beating top Go professionals to be roughly a decade away, and few predicted how quickly large language models would improve. History suggests humility about any confident timeline.

What forecasting research teaches

Research by Philip Tetlock and colleagues, summarized in the book Superforecasting, found that some people consistently forecast world events better than others — including experts. Good forecasters tend to:

  • make specific, measurable predictions with deadlines;
  • express probabilities rather than certainties;
  • start from base rates (how often similar things happened before);
  • update frequently as evidence arrives;
  • consider multiple perspectives;
  • keep track records and learn from errors.

Forecasting platforms such as Metaculus aggregate many forecasts on AI questions and record accuracy over time.

Questions to ask about any AI prediction

  1. Is it specific? “AI will transform work” can’t be wrong. “AI will autonomously complete a week-long software project by 2030 with 90% reliability” can be checked.
  2. Is there a probability? Confident statements hide uncertainty.
  3. What is the evidence? Trend data, benchmarks, expert surveys, or only intuition?
  4. What are the incentives? Companies raising money, critics building audiences and researchers seeking funding all have reasons to emphasize particular views.
  5. What’s the track record? Has this person or organization predicted well before?
  6. What would change their mind? Good forecasters can name evidence that would update them.
  7. Does it confuse capability with impact? A model being able to do something doesn’t mean it will be widely adopted quickly; adoption depends on costs, regulation, organizations and trust.

Common warning signs

  • Extrapolating a single demo into sweeping claims.
  • Exponential curves forever: trends often slow as constraints bind.
  • “This time is different” without explaining why.
  • Vague terms like “AGI” without definitions.
  • All-or-nothing framing: either utopia or doom.

Useful sources of evidence

  • Trend data on compute, costs and capabilities, from organizations such as Epoch AI and the Stanford AI Index.
  • Expert surveys, which show wide disagreement — itself informative.
  • Forecasting aggregates, which combine many views and update continuously.
  • Real-world adoption data from businesses and governments.

How to use predictions personally

Rather than betting on a single forecast, consider scenarios: what would you do if AI progress is fast, moderate or slow? Choices that work reasonably well across scenarios — learning to use AI tools, building durable skills, staying adaptable — are usually wiser than all-in bets.

Example: rewriting a vague prediction

Vague: “AI will replace programmers soon.”

Better: “By the end of 2028, there is a 30% chance that a majority of new code at large software companies will be written by autonomous AI agents with only review by humans, as measured by company disclosures or credible surveys.”

The second version is specific, has a deadline, expresses uncertainty and names how it will be checked. It can be wrong — which is exactly what makes it useful. Applying this rewrite to predictions you read is a good test: if a claim can’t be made specific, it probably isn’t saying much.

Keeping your own track record

Write down a few predictions about AI each year, with probabilities, and check them later. It’s a humbling and effective way to calibrate your judgment — and to notice which sources helped you predict well.

What to watch next

  • Forecasting questions on AI capabilities and their resolution.
  • Expert surveys on AI timelines and risks.
  • Independent trend data on compute and costs.
  • Benchmarks that measure long-horizon autonomous tasks.

Key takeaways

  • AI history includes major over- and under-predictions.
  • Good forecasts are specific, probabilistic, evidence-based and updated.
  • Check incentives, track records and what would change a forecaster’s mind.
  • Distinguish capability from adoption and impact.
  • Plan with scenarios rather than single predictions.

Frequently asked questions

When will we reach AGI?

Nobody knows. Expert estimates vary enormously, partly because AGI lacks an agreed definition.

Are AI company leaders’ predictions reliable?

They have deep knowledge but also incentives. Weigh their claims alongside independent evidence.

What is a superforecaster?

A person who has consistently made accurate probabilistic forecasts, a concept from research by Philip Tetlock and colleagues.