TL;DR
When you notice that the AI might have gotten something wrong, don't read a ready-made explanation of "what's wrong" — first write your own. Then attach a short personal reminder phrase to that mistake, one that's yours and yours alone. These two simple tricks together make your vigilance toward AI errors stronger and longer-lasting.
People know AI can make mistakes and want to check its answers — but they still miss errors anyway. The problem isn't knowledge or laziness: the right information simply doesn't surface at the right moment. The classic example is the Moses illusion: everyone knows it was Noah who built the ark, but when asked "how many animals did Moses take on the ark," many people answer without noticing the switch, because the question gives no hook to pull the right fact out of memory. Smooth, confident AI text works the same way — it doesn't send a signal that says "double-check this."
The method targets two points: how you encode information about a mistake the first time (better to formulate the explanation yourself than read someone else's) and what reminds you of it later (a personal trigger phrase, not a generic warning). In the experiment, self-explanation added 10 percentage points to error-detection rates, and a personal trigger noticeably slowed the decline in vigilance during repeated work with AI.
Method Blueprint
STEP 1 (in the moment you spot a mistake): You notice something suspicious in the AI's answer →
YOU write the explanation of why it's wrong and how to check it →
only then do you compare it against the correct answer/fact
STEP 2 (one-time, right after the review):
Come up with a short personal trigger phrase tied specifically to this case →
save it (a note, a phone reminder)
STEP 3 (ongoing, before every new round of checking the AI):
You see/recall your trigger →
it switches on the same review logic, even without re-explaining anything
Steps 1 and 2 can be done in a single conversation with the LLM. Step 3 is a habit, not a prompt.
Example in Action
Task: Every week you ask ChatGPT to summarize sales figures and explain the trend. One time the AI smoothly explained "23% growth," when in fact growth was 9% — the model mixed up the periods.
Prompt (Step 1 — self-explanation instead of a ready-made answer):
I suspect there's an error in this AI answer: the "23% sales growth" figure.
Here's my explanation of why I think it's wrong and how I'd check it:
[your text, e.g., "based on the monthly data, growth was 9%; it looks like the AI
pulled a figure from a different period or added percentages incorrectly.
I can verify this against the source Excel table"].
Tell me whether I've understood the problem correctly, and add anything I might have missed.
Prompt (Step 2 — creating a personal trigger):
Help me come up with a short personal reminder phrase (2-4 words, doesn't need to
make sense to anyone else) that will remind me of this case with the fabricated
23% growth figure. I want to use it as a daily trigger before I review reports.
Result: In the first conversation, the AI won't hand you a ready-made explanation — it will confirm or build on your version, because you're the one formulating it first. In the second, it'll suggest a few short personal phrases, and you pick the one that's yours. After that, it's no longer a prompt: before every new report review, you recall that phrase, and it brings back the specific review logic — not a generic "stay alert."
Why This Works
Smooth, confident AI text doesn't send signals saying "there might be an error here, remember your check." Even if you're aware of AI's typical failure modes, that knowledge simply doesn't surface at the moment you need it — it's in your head, but it's not "retrievable."
Human memory is built so that things you created yourself stick better and are organized more effectively than things you merely read — because creating something engages the same mental operations you'll need later when checking it. And a personal trigger, even one that's meaningless to outsiders (like "the button" or "23% — check the source"), can switch on exactly the chain of thought that was active during your original review of that specific error — not generic anxiety, but concrete logic.
The method leverages both quirks: it forces you to first generate your own explanation instead of reading a ready-made one — this builds a durable memory. And it anchors that memory to a personal trigger — so the specific review logic kicks in exactly when needed, instead of dissolving into a vague "stay vigilant."
Control levers: - Detail in your explanation → the more specific you are (what's wrong, why, how to check it), the stronger the memory - How personal the trigger is → a generic phrase like "stay alert" works weakly; a personal one tied to the specific case works strongly - Reminder frequency → regular use of the trigger slows the decay of vigilance, but doesn't stop it entirely
Prompt Template
Step 1 — self-explanation:
I suspect there's an error in this AI answer: {what seemed wrong}.
Here's my explanation of why it's wrong and how I'd check it:
{your own explanation}.
Tell me whether I've understood the problem correctly, and add anything I might have missed.
Step 2 — personal trigger:
Help me come up with a short personal reminder phrase (2-4 words) tied to this case:
{brief description of the error and the lesson}.
The phrase doesn't need to make sense to anyone else — it just needs to remind
me personally of this specific check.
Fill in the specific error you noticed in the AI's answer ({what seemed wrong}), your own explanation ({your own explanation}), and a brief description of the case for the trigger.
🚀 Quick Start — paste into chat:
Here's a method using self-explanation and a personal trigger for checking AI answers.
Adapt it to my task: {your task — e.g., reviewing reports, code review, checking legal texts}.
Ask me questions to fill in the blanks.
[paste the template above]
The LLM will ask what specific mistake you noticed and in what context you're working — because the trigger only works if it's tied to your real case, not an abstract example.
Limitations
⚠️ Requires discipline: the effect only works if you actually write your own explanation every time, instead of asking the AI to hand you a ready-made answer right away — this is slower than just accepting the model's conclusion.
⚠️ Fades, just more slowly: even with a personal trigger, vigilance toward errors declines over time with continuous AI use — this method slows the decay, it doesn't stop it forever.
⚠️ The trigger has to be genuinely personal: a generic warning like "AI can make mistakes" works weaker than a personal phrase tied to a specific case. Standard warnings built into AI interfaces don't produce this effect.
⚠️ Tested on similar, recurring errors: participants trained on a particular type of error. How well the method transfers to a completely new, unfamiliar type of AI mistake hasn't been tested.
How It Was Studied
Researchers ran two field experiments at a real company — an enterprise software maker in China — with 640 customer-facing employees. In the first experiment (400 people), one group was shown a ready-made explanation of an AI error, while the other group was made to first write their own explanation of the same error — and only afterward were shown the correct version. When both groups went on to review new AI answers, those who'd written their own explanations caught 10 percentage points more errors.
In the second experiment (240 people, one week of daily work), everyone first wrote their own explanation, then received a personalized reminder every day. But one group's reminder contained a personal trigger tied to their past review of the error, while the other group got a phrase that was personalized but meaningless. The group with a genuine trigger saw vigilance decline noticeably more slowly.
An interesting design detail: the authors deliberately separated the effect of "just getting a personal message perks you up" from the effect of "a specific trigger restores specific review logic." The difference held up specifically in the second case — meaning it's not about pleasant personalization, but about a real associative link to the original review.
Resources
Fu, X., Ramasubbu, N., Galletta, D.F. Knowing Is Not Enough: Information Retrievability as a Precondition to Effective LLM Oversight. Preprint, forthcoming in Journal of Management Information Systems, 2026. Georgia State University, University of Pittsburgh.