AI Prediction Scorecard: Tracking Who Got It Right

Predictions are easy to make and hard to track. The AI industry is saturated with bold claims about timelines, capabilities, and consequences — claims that shape policy, investment, and public perception. Yet remarkably few of these predictions are ever systematically evaluated against reality.

This scorecard changes that. We track specific, verifiable predictions made by AI leaders, researchers, and public intellectuals. We record what was predicted, when it was said, what deadline was given (explicit or implicit), and whether the prediction turned out to be correct. Where predictions remain unresolved, we track them until they can be evaluated.

The goal is not to humiliate people who get things wrong. Prediction is hard, and the honest acknowledgment of uncertainty is a virtue. The goal is to build a public record that distinguishes careful forecasters from habitual hype merchants, and to create accountability for claims that influence billions of dollars in investment and the lives of billions of people.

For related data, see our AI Statistics 2026, AI Doomsday Clock, and AI Safety Complete Guide.

Methodology

Inclusion Criteria

A prediction is included if it meets all of the following:

• Made publicly (interview, publication, social media, testimony)
• Attributable to a specific individual
• Contains a verifiable claim about a future event or state
• Includes an explicit or strongly implied deadline

We do not include vague aspirational statements (“AI will transform everything”), hedged probabilities without specific outcomes, or private communications.

Evaluation Standards

Scoring Notes

We evaluate predictions charitably where reasonable. If someone predicted “AGI by 2025” and we are assessing in February 2026, we treat the prediction as wrong — but note if progress toward the predicted outcome was substantial. Conversely, if someone predicted something would not happen and it did, we do not penalize them for cautious underestimates unless the prediction was clearly directionally wrong.

AGI Timeline Predictions

The most consequential predictions in AI are about when — if ever — artificial general intelligence will be achieved. These predictions directly influence investment decisions, regulatory urgency, and public preparedness. See our glossary for the definition of AGI.

Capability Predictions

Market & Industry Predictions

Risk & Safety Predictions

Regulation & Policy Predictions

INHUMAIN.AI Predictions

We believe accountability should apply to us as well. These are our own predictions, with explicit deadlines:

We will evaluate these predictions publicly on December 31, 2026, and update this page accordingly. If we are wrong, we will say so clearly and analyze where our reasoning failed.

Predictor Track Records

Cumulative Accuracy (Evaluated Predictions Only)

Context: Most predictions are still pending because they concern events in the 2027-2035 range. Accuracy percentages on small samples should be interpreted cautiously. What the data does show is a pattern: those who predict rapid AGI timelines have consistently been wrong, while those who predict specific harms from current systems have been more accurate.

What This Tells Us

Three patterns emerge from the scorecard:

1. AGI timelines are consistently overestimated by those with financial or reputational incentives to do so. Elon Musk, Sam Altman, and other figures with significant investments in AI companies have repeatedly predicted AGI timelines that have not materialized. This does not mean AGI is impossible or distant — it means that the people making the loudest claims have the least reliable track records.

2. Harm predictions are consistently underestimated by the same people. The individuals who are most optimistic about AGI timelines are often the same individuals who are most dismissive of current AI harms. The data shows the opposite pattern: specific harms are materializing faster than predicted, while AGI remains elusive.

3. Uncertainty is the honest position. The predictors with the best track records are those who express genuine uncertainty, provide wide ranges, or focus on specific near-term developments rather than sweeping timeline claims. Geoffrey Hinton’s “5-20 years” range is more honest than Elon Musk’s “by next year.”

For how these predictions relate to our risk assessment, see the AI Doomsday Clock.

This scorecard is updated quarterly. New predictions are added as they are made publicly. Evaluation status is updated as deadlines pass. Corrections and additional documented predictions can be submitted through our contact page.