All playbooks

Playbook · for everyone

AI research that doesn't hallucinate on you

For: anyone using AI to answer questions that matter. Time: the method costs minutes per claim. Cost: free; it's a discipline, not a tool.

AI research tools are genuinely good now, and they still make things up: confidently, fluently, with fake-plausible specifics. We publish research-backed work, so we run a verify-before-publish method on everything. It is simple enough to use tonight.

Do this

  1. Ask for claims with sources, then open the sources. The discipline is the clicking. An AI summary of a page nobody opened is a rumor with good grammar.
  2. Keep two piles: "the AI said it" and "I saw it in the source." Only the second pile is knowledge. We literally tag claims UNVERIFIED in our notes until a human or a second process has seen the source say it.
  3. Date everything. Stats rot. A "19% of businesses" figure from 2024 is a different fact in 2026. If the source has no date, that's a finding about the source.
  4. Get a second path to anything load-bearing. If a number drives a decision, find it twice, independently. One search result quoting another search result is one path wearing a disguise.
  5. Force live search for anything recent. AI models' built-in knowledge lags months behind. For current events, prices, laws, and people's jobs: make the tool search, then check what it found.

Skip that

  • Skip "studies show" without a study you can open. That phrase is where made-up facts go to sound respectable.
  • Skip using one AI to fact-check another without sources. That's two rumors agreeing.
  • Skip publishing the unverified pile, ever. If you can't back it, cut it; the credibility you keep is worth more than the sentence you lose.

Sharp edges

The failure mode is not obvious nonsense; it's the plausible specific: a real-sounding link that goes nowhere, a statistic with one digit drifted, a bill described as law when it died in committee. We have caught all three in our own work. Nuance flattens too: "the program was announced" and "the program is funded" read the same in a summary and are very different facts. When a claim matters, read the primary source's actual words.

We ran it ourselves: every fact in our published research carries a source a reader can open, and claims we couldn't verify got cut. This playbook is our literal production process.