AI Predictions, Scored
AI forecasts are quoted constantly and checked almost never. This scoreboard logs the forecasts that appear across Report AI — who made them, what exactly they said, when they resolve and what would count as a hit — and scores them as their dates arrive, including the misses.
Thank you for reading this post, don't forget to subscribe!
The problem with AI forecasts is not that they are wrong. It is that nobody checks.
AI forecasting has a peculiar economy. A projection is published, quoted for years, and quietly forgotten when its date arrives — or worse, quoted as the outcome. Cybersecurity Ventures’ projection of $10.5 trillion in cybercrime costs for 2025, made in 2020, is now widely cited as what 2025 cost, though nobody measured it. A forecast that is never scored costs its author nothing when it misses, which is exactly why so many are bold.
Logging forecasts precisely, before they resolve, exposes a second problem: many cannot be scored at all. A prediction that half of AI-driven layoffs will be reversed “in some form” depends entirely on what counts as reversal. A market-size forecast resolves only against another estimate built on another definition. And a probability for a one-off catastrophe cannot be checked by any single outcome. We say so explicitly, using a verdict of unfalsifiable, because that is information too: a forecast that cannot be wrong should carry less weight than one that can.
So for each entry we record, at the time of logging, what would count as a hit. Writing the test down before the answer is known is what stops a scoreboard from becoming a list of generous reinterpretations. Where the forecaster’s wording leaves room, we state the reading we will score against.
The first round lands within three months. Forrester’s layoff-reversal forecast and Gartner’s prediction that 40% of enterprise applications will include task-specific AI agents both resolve at the end of 2026, and our own July 2026 odds on a federal AI “kill switch” resolve when the current Congress ends in January 2027. We will publish the results — including our own — in January.
- Three forecasts resolve by January 2027, including one of our own.
- Two are unfalsifiable as published — a $12.2T cybercrime projection with undisclosed method, and a 70% probability of catastrophe.
- Several have no measurement to resolve against. Nothing currently measures “generative-AI-enabled fraud”, which Deloitte forecasts at $40B by 2027.
- The test is written before the answer. Every entry states what would count as a hit.
The scoreboard
| Forecast | By | Resolves | What counts as a hit | Status · evidence so far |
|---|---|---|---|---|
| About half of AI-attributed layoffs will be reversed in some form by the end of 2026 | Forrester | 31 Dec 2026 | Credible survey or tracker data showing roughly 50% of AI-attributed cuts refilled or reversed | Due Jan 2027 32% of US hiring managers say an AI-eliminated role has been refilled; 55% of employers regret AI layoffs. Source page |
| 40% of enterprise applications will feature task-specific AI agents by end-2026, up from under 5% in 2025 | Gartner (Aug 2025) | 31 Dec 2026 | Gartner or independent survey measurement at or near 40% | Due Jan 2027 No published measurement of progress found; claims that it is “ahead of schedule” cite no data. Source page |
| Federal AI “kill switch” outcomes this Congress: House bill as written ~5%; Senate S.2938 intact ~8%; diluted shutdown power ~20%; narrower guardrails ~45%; nothing federal ~35% | Report AI (Jul 2026) | End of 119th Congress, Jan 2027 | Scored as probabilities: did higher-odds outcomes occur more often than lower ones | Due Jan 2027 Our own forecast; it will be scored the same way as the others. Source page |
| Whether the September 2026 frontier slowdown pact holds: an antitrust safe harbour passes; Buist v. Anthropic survives a motion to dismiss; evaluator access materialises; a paused capability is named; SpaceXAI commits beyond endorsement | Report AI watch list (Sep 2026) | Rolling | Each of the five scored yes or no as it is decided | Open Last checked at publication, 24 Sep 2026. Source page |
| Over 40% of agentic AI projects will be cancelled by the end of 2027 | Gartner (2025) | 31 Dec 2027 | Gartner or independent data showing more than 40% of agentic projects cancelled | Open No interim measure found. |
| US generative-AI-enabled fraud losses will reach $40B by 2027, from $12.3B in 2023 | Deloitte (2024) | 2027 data, reported 2028 | A published estimate of the same quantity at or above ~$40B | Open No measurement of this quantity exists. The FBI’s AI-linked complaint losses ($893M, 2025) measure something narrower. Source page |
| 170M jobs created and 92M displaced by 2030 — net +78M | World Economic Forum (Jan 2025) | 2030 | WEF’s own later editions and labour statistics broadly consistent with net job growth | Open WEF revises the projection in each two-yearly edition; we will log revisions. Source page |
| Synthetic-identity fraud losses of at least $23B by 2030 | Deloitte | 2030 | A published estimate of the same quantity at or above $23B | Open No annual tracker on the same definition found. |
| Half of AI-agent deployment failures will stem from governance gaps and broken interoperability by 2030 | Gartner | 2030 | A published attribution of agent failures by cause | Open Hard to score: no body counts agent failures by cause. |
| The AI video market will grow from $6.2B (2025) to $47.8B by 2034, a 25.4% CAGR | Grand View Research | 2034 | Grand View’s own later estimate on the same definition, checked against interim years | Open Different firms’ definitions give $0.9B to $6.2B for 2025. Source page |
| A “superhuman coder” milestone in early 2027, in the AI 2027 scenario | Daniel Kokotajlo et al. (Apr 2025) | Early 2027 | Published as a scenario, not a point forecast; scored on whether frontier labs report AI systems doing most of their own coding research by mid-2027 | Open Source page |
| Cybercrime will cost the world $12.2T a year by 2031, from $10.5T in 2025 | Cybersecurity Ventures | 2031 | — | Unfalsifiable Method not disclosed, and the 2025 “baseline” is itself a 2020 projection. Fact-check |
| Roughly 70% chance that AI catastrophically harms humanity | Daniel Kokotajlo (2024) | — | — | Unfalsifiable A probability for a one-off event cannot be scored by any outcome short of the event. |
How forecasts are scored
| Verdict | Meaning |
|---|---|
| Open | Not yet at its resolution date. |
| Due | Resolves within the next scoring round. |
| On track / Off track | Interim evidence points clearly one way, before resolution. |
| Hit / Miss / Partial | Scored against the test recorded at logging. |
| Superseded | Withdrawn or revised by the forecaster; the original and the revision are both kept. |
| Unfalsifiable | No outcome could show the forecast wrong as published. |
Probability forecasts, including our own, are scored as probabilities: one outcome cannot prove a 20% estimate wrong, but a record of many forecasts shows whether events rated likely happened more often than events rated unlikely. We record each forecast’s wording as published and link to the page where we first reported it.
Frequently asked
Do AI forecasts usually come true?
There is too little scoring to say, which is the problem this page addresses. Many widely quoted AI forecasts are never checked against outcomes, and some cannot be, because their definitions or methods are not published.
Will 40% of enterprise apps have AI agents by the end of 2026?
That is Gartner’s August 2025 forecast, up from under 5% in 2025. We found no published measurement of progress toward it as of October 2026. It resolves at the end of 2026 and we will score it in January 2027.
Will AI-driven layoffs be reversed?
Forrester forecast that about half of AI-attributed layoffs would be reversed in some form by the end of 2026. Interim data shows 32% of US hiring managers have refilled a role eliminated because of AI. The forecast resolves at the end of 2026.
What makes a forecast unfalsifiable?
A forecast is unfalsifiable when no possible outcome could show it to be wrong — because its method is undisclosed, its terms are undefined, or it is a probability for a single event. Such forecasts are not necessarily mistaken, but they deserve less weight.
Does Report AI score its own forecasts?
Yes. Our July 2026 probability estimates for a federal AI kill-switch law are on this scoreboard and will be scored in January 2027 on the same terms as everyone else’s.
Forecasts are logged as first reported on the Report AI page linked in each row, where their sourcing is set out. The “what counts as a hit” test is written by Report AI at logging and will not be revised after the fact; where a forecaster’s wording is ambiguous, the test states the reading we will score. Interim evidence is rated on the linked pages, mostly MEDIUM. Our own forecasts are marked “Report AI” and are editorial judgements. Scoring rounds follow resolution dates, starting January 2027. See also our editorial standards.
AI statistics that are wrong · Corrections log · Will AI replace my job? · The Dark Side of AI in Numbers · Regulating AI