The claim, under the stated reading, was fully realized by its deadline. The charts call this “On time.”
HOW WE JUDGE
How we judge a prediction.
We compare each forecast with what happened, using its deadline and stated scale. Each record explains our interpretation and links to the evidence.
What counts as success?
We read each claim in context, retaining its cost, capability, adoption, geography, mechanism and deadline requirements. A forecast of a trend does not require universal adoption.
Wireless keyboards and mice were practical and in ordinary use by 2009. He was right about the technology, but less accurate about consumer preference: wired devices remained in use. Our default judgment credits the move toward wireless; the stricter reading also tests how far cables disappeared. The household robot forecast set a higher bar: robots in nearly every household, performing a broad range of chores.
Consumer preference, cost and regulation can limit adoption even when a technology works. We weigh those factors case by case and explain the judgment. A working option may satisfy a capability forecast; a forecast of widespread use also needs adoption evidence. When the wording leaves room for disagreement, we retain alternative readings and identify what changes the result.
Each record separates the attributed forecast from its testable criterion, outcome evidence and editorial verdict. The displayed claims are paraphrases, not verified verbatim quotations. “Disputed” means a reasonable alternative reading changes the result.
Stricter and broader readings
Alternative readings can change the population or deadline as well as the standard of success. Broader readings may relax a mechanism or adoption requirement. Changed conditions are listed on the record; alternative readings never add extra predictions to the count.
How the reading changes the result
Same 67 predictions in every reading. One brain-modeling forecast is left out because its alternative deadline has not passed.
| Reading | Predictions | On time | Fully realized |
|---|---|---|---|
| Reference (stated reasonable reading) | 67 | 37.3%25 / 67 | 53.7%36 / 67 |
| Strict / alternative reading | 67 | 25.4%17 / 67 | 37.3%25 / 67 |
| Broader / directional reading | 67 | 55.2%37 / 67 | 92.5%62 / 67 |
The main comparison has 68 due predictions. This comparison uses the 67 eligible in all three readings. Broader success may mean meeting different conditions; open a record to see what changed.
Five outcomes.
Fully realized after the deadline. It counts as realized, but not on time.
Meaningful parts happened, but the full claim has not been met by the review date. Partial gets no fractional credit.
The deadline passed and the claim was not fulfilled under this reading. This is a judgment as of the review date, not a claim that it can never happen.
The deadline is still ahead. Progress can be recorded, but the prediction stays out of deadline-based scores.
Each record shows the outcome at the original deadline and at the review date.
Unreviewed, unscorable and excluded records
Unreviewed is a research state, not a failure. Unscorable means no defensible test or usable deadline is available. Excluded marks a duplicate, misattribution or another documented exclusion. These records remain accessible and receive no points.
Some undated records have assessable outcomes. Their verdicts remain visible, but they are excluded from deadline-based rates.
Late means it happened.
A forecast that has not happened is not “late” merely because its date has passed. A late verdict requires evidence of full realization after the deadline.
The supercomputing forecast had a 2009 target and reached the chosen measured-performance benchmark in 2013: four years late. The chess forecast succeeded in 1997 under either of its cited 1998 and 2000 deadlines. That leaves its outcome clear and its exact timing uncertain.
“Documented by 2008” means we have evidence from 2008. It does not mean 2008 was the first year it happened. We distinguish exact dates, documented-by bounds and approximate periods.
Whole-year deadlines close on December 31. At this October 15, 2026 snapshot, a 2026 deadline is still open. Approximate dates use the interpretation recorded on the detail page. A later revision never overwrites an earlier forecast.
What is being counted?
The collection has 84 distinct forecast versions. 68 have passed their deadlines and are eligible for the default comparison. Future and undated records stay out of both rates below.
On time25 / 68 = 36.8%Fully realized by the review date, including late36 / 68 = 52.9%
Both use the same predictions. Partial receives no fractional points; missing evidence does not automatically mean failure. The homepage comparison uses this group. The full ledger also includes future and undated records.
Identical repetitions count once. Materially different deadlines or thresholds remain separate versions, linked to their related forecasts. Each version has one primary category; extra tags do not multiply its weight. Several versions from one forecast family can therefore influence a category.
Inspect all 68 predictions in the comparison →
Counting and research details
Eligibility requires an assessed, due, canonical record with the supplied eligibility flag. Non-scoring comparators and conditional judgments on unresolved legacy wording stay out of every rate. Future claims stay out even if achieved early.
We have not implemented family-level weighting. Category groups have different deadlines, claim types and selection histories, so these descriptive counts are not estimates of future forecasting skill. Zero eligible records means no rate, not zero accuracy.
Research checks attribution, fixes a criterion, resolves duplicate mentions, assesses both dates and records the rationale. Attribution, outcome and timing confidence are stored separately. An unresolved original-edition detail need not erase an otherwise supported outcome.
The limits of this collection.
This is a selected, expanding collection, not a complete audit or a random sample of Kurzweil’s predictions. A one-record category cannot establish a general strength or weakness.
The research and editorial judgments are AI-assisted and have not had independent human review. Most book claims were checked through author retrospectives, reprints or article mirrors. The original book editions have not been independently checked against those passages.
Before making a lifetime accuracy claim, we would need systematic coverage of the books and interviews, including the full 147-item set of predictions for 2009, and consistent review of the outcomes. Precise first-achievement dates are also missing for many records.
All 88 original intake rows and the v0.1, v0.2 and v0.3 releases are preserved. Evidence additions carry their own dates; rebuilding the site does not update an assessment.
Sources, coverage and downloads →