Hype Over Time
AI moves fast, and so do the promises. Hype Over Time goes back to famous AI claims, the ones that moved markets, careers and headlines, and checks them against what actually happened. Each check uses the same four statuses as our Hype Ledger, and every fact links to its source.
The scorecard
5 promises checked. 5 did not hold up. 0 supported. That is not a verdict on AI itself: in most of these cases, real products and real progress arrived. What missed was the date, the scale, or the framing.
| Promise | Who | Made | Due | What happened | Status |
|---|---|---|---|---|---|
| Over a million robotaxis on the road | Elon Musk, Tesla | Apr 2019 | 2020 | First rides without a safety driver in Jan 2026; small fleets in a handful of cities, size undisclosed | Did not hold up |
| Stop training radiologists; AI will beat them in 5 to 10 years | Geoffrey Hinton | 2016 | 2021 to 2026 | Radiologist jobs, pay and residency slots all rose; AI became a tool inside the job | Did not hold up |
| GPT-4 scored around the top 10% on the bar exam | OpenAI | Mar 2023 | At launch | Likely passed; peer-reviewed recheck puts it near the 48th percentile among people who passed | Did not hold up |
| Devin, the first AI software engineer | Cognition | Mar 2024 | At launch | Outside test: 3 of 20 tasks done; now sold as a supervised helper, with fast revenue growth | Did not hold up |
| A wearable AI Pin to move us past screens | Humane | Apr 2024 (shipped) | Ongoing | About 10,000 sold against a 100,000 goal; shut down Feb 28, 2025 | Did not hold up |
Patterns so far
- Dates slip, then reset. When a deadline passes, a new and larger promise usually replaces it rather than an explanation.
- Percentiles need a “compared with whom.” Benchmark wins often depend on the comparison group or the test setup.
- Tasks are not jobs. Automating one task can raise demand for the people who do the rest of the job.
- Revenue is not the same as the launch claim. A product can sell well while the original framing still fails.
How we score
We judge the claim as it was made, at the time it was made, against evidence available now. Supported means the main sources back it up. Promising but unverified means there is evidence, but mostly from the people behind the claim. Needs follow-up means key details still depend on future events. Did not hold up means the evidence goes against it. Full definitions are on How We Rate Claims.
Know an old AI promise that is due for a check? Send it to us with a link to where it was made.
