Publications

Across dozens of departures from OpenAI, Google, xAI, Anthropic, Meta, and Stability AI, the same concerns surface again and again across their writings. Safety teams are dissolved or absorbed into product work. Deployment timelines are compressed past the point where meaningful evaluation is possible. Researchers who raise objections internally find their concerns deprioritized, and sometimes face retaliation for speaking up.

These are not abstract worries about a distant future. Many of the people who left were senior scientists, alignment leads, and ethics researchers — some of them architects of the safety frameworks at their former employers. Their departures span several major AI organizations and accelerated through 2024 and into 2026, coinciding with the industry's pivot from cautious research to aggressive commercialization.

The writings below — peer-reviewed papers, policy reports, public resignation letters, and long-form essays — represent the intellectual foundation behind these warnings. They document what these researchers saw, what they tried to build, and why they ultimately decided they could no longer stay.

Forecasts are future-facing statements with observable outcomes. Warnings and claims capture important insider assessments that are not suitable for an accuracy score. Resolution criteria added after an original statement are dated and disclosed.

Under editorial review: this tracker's resolution criteria and adjudications are being re-reviewed. Individual records remain visible; aggregate confirmation scores are hidden until the review is complete.

3 open forecasts · 4 warnings and contemporaneous claims · 6 records under review

Partially ResolvedUnder review

AI companies will prioritize deployment speed over safety evaluation

Daniel KokotajloPredicted 2024

I believe the leaders of the top labs are making reckless decisions in how much autonomy and capability they're giving to their models.

Resolution— Jun 2025

The Washington Post and multiple outlets reported on compressed safety testing timelines at major labs through 2024-2025. Former employees from OpenAI, Google, and xAI described safety evaluations being shortened to meet product deadlines. OpenAI shipped GPT-4o with abbreviated red-teaming.

View evidence

OpenUnder review

AGI could plausibly be built in as little as three years

William SaundersPredicted 2024

I believe it is plausible that an AGI system could be built in as little as three years.

OpenUnder review

Advertising in ChatGPT will create manipulation incentives OpenAI cannot govern

Zoe HitzigPredicted 2026

Advertising built on that archive creates a potential for manipulating users in ways we don't have the tools to understand, let alone prevent.

Open

Geoffrey Hinton estimated a 10-20% chance of AI-caused human extinction within 30 years

Geoffrey HintonPredicted 2024

Estimated a 10-20% chance that AI causes human extinction within the next three decades (paraphrase of BBC Radio 4 remarks as reported by The Guardian, Dec. 27, 2024).
Paraphrase — not a verbatim quotation.Source

OpenUnder review

AI capabilities will advance faster than mainstream expert predictions

Daniel KokotajloPredicted 2021

AI timelines have been consistently shorter than most experts predicted. The pace of progress has surprised nearly everyone.

Open

Current human-feedback alignment methods will not scale to superintelligence

Ilya SutskeverPredicted 2023

Our current alignment techniques will not scale to superintelligence.
Verbatim excerpt.Source

Open

AI systems will generate persuasive disinformation at scale

Geoffrey HintonPredicted 2023

They want to use them for winning wars or manipulating electorates.
Verbatim excerpt.Source

OpenUnder review

AGI could plausibly arrive by 2027

Daniel KokotajloPredicted 2024

I left OpenAI because I lost confidence that it would behave responsibly around the time of AGI. I think AGI is coming and that we are not on track to handle it responsibly.

DisprovenUnder review

AI hype will fade as unrealistic expectations fail to materialize

Daniel KokotajloPredicted 2021, contradicted 2025

But the hype begins to fade as the unrealistic expectations from 2022-2023 fail to materialize.

Resolution— Jan 2025

AI hype accelerated through 2024-2025 rather than fading. Investment in AI companies surged to record levels, major tech companies increased AI spending dramatically, and public attention intensified with the mainstream adoption of ChatGPT, Gemini, and Claude.

View evidence

Contemporaneous claim

OpenAI safety culture has taken a back seat to product launches

Jan LeikePredicted 2024

Over the past years, safety culture and processes have taken a back seat to shiny products.
Verbatim excerpt.Source

Editorial synthesis

Without adequate oversight, alignment failures will occur in deployed systems

Jan LeikePredicted 2024

Building smarter-than-human machines is an inherently dangerous endeavor. OpenAI is shouldering an enormous responsibility on behalf of all of humanity. But over the past years, safety culture and processes have taken a back seat.
Verbatim excerpt.Source

Contemporaneous claim

Leading AI lab security is inadequate to protect model weights

Leopold AschenbrennerPredicted 2024

Argued that leading labs' security was far from sufficient to protect model weights from state actors (paraphrase of "Situational Awareness: The Decade Ahead," June 2024).
Paraphrase — not a verbatim quotation.Source

Contemporaneous claim

US government is dangerously unprepared for transformative AI

Leopold AschenbrennerPredicted 2024

Argued the U.S. government was dangerously unprepared for transformative AI, with national-security implications rivaling nuclear weapons (paraphrase of "Situational Awareness: The Decade Ahead," June 2024).
Paraphrase — not a verbatim quotation.Source