AI Debate & Evidence
How these pages differ: every claim carries the date it was made, and every number is labeled as one of three things — a checkable fact, a third-party estimate, or a judgment about an uncertain future. Where credible sources disagree, the disagreement is shown instead of averaged away.
Most coverage of AI risk asks you to pick a side before showing you any data. These pages do the opposite: they separate what the public record supports from what remains genuinely unresolved, and they say which is which. That approach is slower than a hot take, but it is the only one that survives the next news cycle.
Current topics
Should AI Development Slow Down? What "Pacing the Frontier" Actually Proposes
The September 2026 argument, tracked day by day: what Anthropic's CEO asked for on September 12, who supported it within 24 hours, who rejected it, why AI equities fell on September 14, and which claims in the debate are checkable versus judgments. Includes an original cost comparison sizing the proposed safeguard against a single training run.
Is AI Actually Dangerous? What 1,663 Recorded Incidents Show
What the AI Incident Database records about harm that has already occurred, broken down by year, risk domain, system type, and severity. The finding that gets least attention: misuse is a larger recorded category than malfunction, and software causes about twice as much documented harm as all physical systems combined.
What a Frontier Model Costs to Train, and Who Profits From Slowing Down
Estimated training costs from GPT-3 to the 2026 frontier tier, where the money actually goes, the DeepSeek counterexample and what it does not prove, and why the companies selling compute and the companies buying it want opposite things from a slowdown.
Is AI Taking Entry-Level Jobs? Where the Data Agrees and Conflicts
Payroll records show young workers losing ground in AI-exposed occupations while employer surveys show hiring intentions rising. Both measure something different, and both are reported here with their methods, alongside the contested question of causation.
What we will not do here
- No forecast of existential risk. There is no data that settles it, and pretending otherwise is the failure mode of this entire genre.
- No invented monthly costs or capability rankings. Where a figure is an estimate from cluster size and hardware pricing, it is labeled as one.
- No vendor-published numbers presented as independent results. Benchmark claims keep their source and their caveat.
- No averaging over disagreements. If two credible sources differ by a factor of three on the same model, both numbers appear.
Related on this site
- AI Agent Security — threat-modeling tool access, secrets, and approval boundaries
- Editorial & data methodology — how sources are checked and corrections handled
- AI API cost calculator — the published rates referenced throughout these pages
- Reviewed AI agent guides — the full engineering guide index
Bottom line
These pages exist so that a reader can disagree with a conclusion while still trusting the numbers underneath it. Every claim is dated, every figure is labeled, and corrections are accepted with a source through the contact page.