The AI Prediction Ledger: Who Was Right, Who Was Wrong, and What Comes Next

A 10% chance that AI causes human extinction within the next decade.

This is a probabilistic forecast, not a claim that extinction will occur. It cannot be marked wrong merely because the outcome does not happen; calibration requires a reference class of comparable forecasts.

More than a 10% chance that AI kills all humans within the next decade.

The estimate is unusually explicit but still lacks a published model showing how the number was derived. Keep it as a forecast, not a fact about current systems.

“Powerful AI” matching or exceeding Nobel-level experts across most disciplines could arrive in late 2026 or early 2027.

The definition is demanding and more useful than a bare AGI label. A fair resolution still needs agreed tests for expert-level performance, autonomous work and reliability across disciplines.

AI could eliminate half of entry-level white-collar jobs within one to five years.

This needs a fixed occupation list and baseline job count. Job openings, employment levels and task automation are different measures and should not be swapped after the deadline.

AI-driven displacement could push US unemployment to 10–20% within one to five years.

This is measurable, but causation matters. A recession that raises unemployment without demonstrated AI displacement would not satisfy the forecast.

AI would write 90% of code within three to six months and nearly all code within a year.

Coding-agent use rose quickly, but no credible industry-wide measurement approached 90% by September 2025. Amodei later said it happened “at least at some places,” which narrows the scope after the fact.

Powerful AI could compress 50–100 years of biological progress into 5–10 years and help cure or prevent most major diseases.

The clock begins only after Amodei’s “powerful AI” threshold, so the medical claim cannot honestly be assigned a fixed deadline until that trigger is resolved.

The first AI agents could “join the workforce” in 2025 and materially change company output.

Working agents arrived and enterprise adoption followed. The undefined words “join” and “materially” prevent a clean win, and OpenAI’s strongest quantitative evidence comes from its own users and staff.

Systems that can produce novel insights would likely arrive in 2026.

Resolution requires a disclosed, independently validated discovery, not a model generating a plausible new sentence or helping a human researcher work faster.

Robots capable of doing tasks in the real world may arrive in 2027.

Robots already perform constrained real-world tasks. The claim needs a stronger pre-registered bar, such as generality, autonomy and reliability across unfamiliar environments.

AGI with human-level cognitive versatility is five to ten years away.

Hassabis supplied a capability concept and a window, but no public test suite. The ledger will not let a company announcement settle the question on its own.

A 50% chance of “minimal AGI” by 2028.

This updates a forecast Legg made in 2011. Scoring requires the operational definition discussed in the source, not whatever system receives an AGI label in 2028.

Human-level AI would take several years, if not a decade, and would not arrive in the next year or two.

The near-term exclusion can be evaluated first. LeCun also predicted that scaling current language models alone would not produce human-level intelligence, a separate architectural claim.

AI would probably become smarter than any single human by the end of 2025.

No agreed scalar measure of “smarter” exists. Frontier systems exceeded every human on some tests and remained brittle on basic tasks. The deadline passed, but the wording prevents a defensible verdict.

Tesla would have more than one million robotaxis on the road in 2020.

Tesla had no public robotaxi service in 2020. Its September 2026 Texas fleet numbered about 420 registered robotaxis, roughly 99.96% below the promised million and almost six years late.

Hospitals should stop training radiologists because deep learning would outperform them within five years.

AI became valuable in image analysis, but radiologists were not made obsolete. Demand remained high, and Hinton later acknowledged that his forecast was wrong.

A 10–20% chance that AI causes human extinction within 30 years.

As with the Coxon and Hubinger forecasts, one binary outcome cannot establish individual calibration. The value comes from preserving the probability, horizon and subsequent revisions.

Within five years, AI could pass every test a human takes, including specialized professional exams.

The claim concerns test performance, not general intelligence or reliable professional practice. Resolution should use a frozen test set and human comparison rule.

An AI coding agent with roughly mid-level-engineer capability would become possible in 2025.

Coding agents became capable of substantial bounded work, but “mid-level engineer” bundles judgment, ownership and reliability that benchmark scores do not establish. Meta later shifted part of the timeline into 2026.

AI could make a two-day workweek possible within a decade as machines handle most routine work.

Technical productivity does not automatically shorten paid work. The resolution rule should track the median standard workweek across major economies, not isolated four-day-work pilots or individual choice.