About PrognostiBench
We are being inundated with claims about how AI is going to impact us every day. Quantitative measures such as GDP, productivity, and stock market performance exist to help us track those impacts, but are lagging indicators and sometimes require multiple years to be revealed. So many claims I see each day are forward-looking with implied timeframes, or are captured only in Substack posts, tweets, or podcast quotes. I thought someone should start keeping track of these claims, if nothing else, to try and identify converging and diverging views over time.
PrognostiBench is an attempt to change that. I (and my helpful AI assistants) extract testable, time-bounded predictions about AI from public sources, catalog them with structured metadata, and maintain them as a public record. As time passes and evidence accumulates, I will revisit these claims and update their status. I don't have subscriptions to every publication or Substack, and I don't read Twitter 24 hours a day. I also didn't build a web crawler to exhaustively source out claims. That being said, you're welcome to submit an article and I'll happily add it to the queue if I can access it.
This is not a prediction market, or a scorecard, or an attempt to identify superforecasters. Instead, it's more of a vibe check of the take landscape, while also taking seriously the claims that are being made and offering suggestions for how we might start measuring or evaluating these claims over time. The Pulse page offers a visual snapshot of aggregate trends across all tracked claims.
The goal is not to debunk or endorse any particular claim, but to create a shared, transparent record of what was predicted, by whom, and when, so that the public conversation about AI can be grounded in evidence.
How It Works
Articles are analyzed using a structured extraction process, then manually reviewed before publication. For details, see our methodology.
Built with AI
PrognostiBench was built in cooperation with Claude Code, and not by a single model. The application code, architecture, design, and testing were developed collaboratively across several frontier models: Claude Fable 5, Claude Opus 5, and Claude Sonnet 5 by Anthropic, together with GPT 5.6 Sol by OpenAI. Different models did different work — drafting, review, refactoring, and design — and the choice of which to reach for was itself part of the build.
The running pipeline is narrower than the workshop that produced it: Claude Sonnet 4.6 performs the structured claim extraction and the author biography generation on every article the site ingests.
Contact
Built and maintained by Trey Causey.