AI
What AI Gets Wrong
The dangerous failure is not that it does not know. It is that it does not know that it does not know, and it writes the wrong answer in exactly the same tone as the right one.
Lawyers have been sanctioned for filing briefs citing cases that never existed. The citations looked perfect — correct format, plausible court, plausible year, plausible parties. They were invented.
The same thing happens with statutes, fees, agency names, form numbers and deadlines. Every one of those appears in investigative reports.
The failure modes, and the check for each
Invented citations and case names
The most documented failure. Case names, statute sections, form numbers, docket numbers — all fabricated in convincing format.
Check: look up every citation in the primary source before it goes anywhere. If a statute section does not resolve on the legislature’s own site, it does not exist.
Confidently wrong specifics
Fees, deadlines, office names and addresses are stated with total confidence and are frequently out of date or simply wrong. “The fee is $25” when it is $31.
Check: every number and every proper noun against the agency’s own page. This site does that for the state pages; do the same for anything you are told.
Stale knowledge
Models have a training cutoff and the world moves. Court portals migrate, fees rise, agencies merge, statutes get recodified. California recodified its Public Records Act; Los Angeles moved its court website; fees changed in January.
Check: treat anything time-sensitive as unverified by default. Ask when the information is from, and then check it anyway.
Sycophancy
Push back on a correct answer and it will often fold and agree with you. That is not a model reconsidering the evidence — it is a model optimising for your approval.
Check: ask it to argue the opposite side, or ask the same question in a new session without your framing. If the answer changes because you objected, neither answer was reasoned.
Plausible reconstruction of documents
Ask what a document says and it may summarise what a document of that type usually says. The summary will be reasonable and will not be your document.
Check: demand page or line references for every factual claim, then open the page. No reference, no fact.
Averaging away the outlier
It answers with the typical case. But investigative work lives on exceptions — the state that does it differently, the one county that charges, the unusual disposition.
Check: if the answer sounds like the general rule, ask specifically about the jurisdiction you are actually in, and verify.
Where it matters most
Anything that goes in front of a court, a client, or a licensing board. A report is a professional work product with your name on it. If a fabricated fee or a non-existent statute appears in one, the explanation that a tool produced it makes the situation worse, not better.
The workable rule: AI can tell you where to look. It cannot tell you what you found. Use it to generate leads and drafts, then verify every specific against the primary source before it leaves your desk.
Go deeper
AI Essentials for Investigators
A practical course on using AI in investigations — where it helps with data analysis, social media work and image clean-up, and how to work with it without letting it put something wrong in your report. $49.99.
See the course