The Situations Where AI Writing Actually Gets You in Trouble

There's a category error in how most people think about AI writing risk. They think about it in terms of detection — will someone run this through a detector and flag it? That's the wrong frame.

Detection is one risk. But the more immediate risk, and the one that trips people up more often, is context mismatch. AI text fails not just when a tool catches it but when a human reader notices something doesn't fit.

Job applications

Cover letters generated by AI have a particular problem: they're often better than the applicant could write themselves, but they're obviously not the applicant. The tone is polished in a way that doesn't match the resume or the phone screen. Hiring managers who read hundreds of these have started noticing the pattern — not necessarily because they're running detectors, but because the letter reads like it could have been written for anyone.

The cover letter is the one place where awkward specificity beats clean prose. Mentioning the exact product feature you've been following, or the specific talk by their CTO that changed how you think about something, is worth more than five paragraphs of polished enthusiasm. AI can't write that because it doesn't know what you actually noticed.

Medical and legal documents

AI produces confident text on medical and legal topics. The confidence is not backed by anything. Models hallucinate citations, misstate dosages, confuse jurisdiction-specific rules, and fill in uncertainty with plausible-sounding guesses that can be seriously wrong.

This is the one domain where "it sounds right" is actively dangerous. The cost of being wrong is high and the model has no way to signal its own uncertainty reliably.

Anything personal

Eulogies. Apology letters. Messages to someone going through a hard time. The people reading these are specifically attuned to whether someone actually meant what they wrote. They will know. Not from a detector — from the fact that AI doesn't generate grief or guilt or specific memory. It generates the shape of those things without the substance.

Where it works fine

First drafts of internal documents. Background research summaries. Boilerplate sections of longer pieces. Anything where "competent and clear" is the bar and the reader isn't specifically invested in whether a human wrote it.

The risk isn't AI writing. The risk is using AI writing in contexts that require the real thing — and not knowing which contexts those are until after it goes wrong.