Editorial standards

How we evaluate AI predictions

A prediction should be specific enough to fail, early enough to matter and documented well enough to verify.

Method in one sentence

AI Predictions rewards lead time, specificity, falsifiability, correct mechanism, realised alignment and consistency — while separating documented facts from our editorial interpretation.

1. Lead time

We give more weight to a prediction made before the trend becomes obvious. A 2017 forecast about commercial generative content is more informative than the same claim made after consumer image generators become mainstream.

2. Specificity and falsifiability

“AI will change everything” is not a forecast. A date, probability, measurable capability or defined commercial outcome can be evaluated later.

3. Mechanism

Identifying the process behind an outcome matters. Scaling laws, learning systems, inference-time computation and organisational adoption are mechanisms; vague optimism is not.

4. Realised alignment

We compare the original claim with subsequent events. Where an outcome has not yet resolved — such as a future AGI date — we say so rather than treating the forecast as a success.

5. Consistency

A prediction that is repeatedly moved away from the present receives less weight than one held stable over time. Probabilistic forecasts are judged according to the probability originally assigned, not as binary promises.

6. Source hierarchy

Evidence typeHow we use it
Peer-reviewed paper / academic publisherHighest weight for technical and historical claims
Recorded interview / official transcriptStrong evidence for dated public predictions
Publisher / employer / personal siteUseful first-party evidence, clearly labelled when self-reported
Secondary journalismUsed for context and corroboration
Unverified repostsNot used as sole support for important claims

7. Category-sensitive ranking

We do not pretend that every prediction is the same. Timeline forecasting, mechanism forecasting, commercial foresight and risk forecasting are separate disciplines. The overall ranking therefore reflects the methodology used for a specific article, while category winners are identified explicitly.

Corrections

If a source is superseded, a date resolves, or new evidence changes a judgment, we update the article and its machine-readable data. Every major page includes a visible last-reviewed date.