A few days ago I got an interesting question from LinkedIn News Europe: How do you decide when AI is wrong, or missing something important?
I start by distrusting how right it sounds.
IT FAILS FLUENTLY.
AI doesn't fail like a calculator, throwing an obvious error you can't miss. It fails fluently. The output that's completely wrong reads exactly like the output that's completely right. That's what makes it dangerous, and it's why fluency is the first thing I refuse to trust. The more confident and well-formatted the answer, the more I slow down.
Last week made the point better than I could. EY pulled a report after the research group GPTZero found significant issues with its citations and sources, including a McKinsey study that doesn't exist. This isn't a competence problem. Whatever sat between the AI draft and the published report didn't catch it, and that gate is what most teams still don't have.
THREE PLACES I CHECK.
After fluency, judgment gets concrete. Three places I always look:
- Anything externally verifiable: numbers, names, citations, dates. This is where models invent most confidently, and it's exactly what busy people skip checking because checking is tedious.
- Anything that needs a single source of truth. If the answer depends on which version of "the truth" the model reached for, it will sound certain about a figure your own systems would contradict.
- Anything irreversible or public. The cost of being wrong, not the odds of being wrong, decides how hard I check.
This isn't theoretical for me. I read the sources. I check whether the studies an AI cites actually exist and actually say what it claims they say. It is tedious, and I do it anyway, because the cost of not doing it is simply too high. A fabricated number in a board deck or a fake citation in a client report doesn't cost you an hour. It costs you your credibility, and sometimes the client. Set against that, the hour it takes to verify is the cheapest insurance you can buy.
BUILD IT INTO THE PIPELINE.
But personal diligence is the floor, not the ceiling. You cannot read every output when there are ten thousand a day, so the real shift is structural: you build the judgment into the pipeline instead of leaving it to whoever happens to be paying attention. Confidence scoring, so low-certainty outputs flag themselves. Human approval gates on anything high-stakes or irreversible. Audit trails, so a wrong answer can be traced back and fixed. Grounding outputs in real data rather than the model's memory. Gartner expects more than 40% of agentic AI projects to be scrapped by the end of 2027, citing unclear business value and inadequate risk controls. That's not a verdict on the technology. It's a verdict on deploying it without any of this.
THE QUESTION FOR EVERY LEADER.
So the question I'd put to any leader: where in your company does AI output reach a customer, a regulator, or a board with no gate in between? Find that gap before it finds you.
Full disclosure: I used AI to help write this and it got a few things wrong, which the checks above caught before they reached you. Why I'd still use it for a piece about not trusting AI? That's for the next article.
SOURCES.
- GPTZero, "Chasing the Hallucinations" — gptzero.me/investigations/ey
- EY retracts study after researchers discover AI hallucinations — Financial Times, May 2026
- Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 — Gartner press release, June 25, 2025