The most expensive habit around generative AI right now is hitting send on its first draft. The tool isn't bad; it's confident in a way that reads as competent, and competence is what people trust when they skim.
A polished paragraph with one invented statistic still looks like a polished paragraph. A summary that gets the tone right but flips a number still gets forwarded. The mistake almost never announces itself, which is exactly why it lands in a client email, a board deck, or the group chat where five people quote it back at you by lunch.
The second read isn't a nicety — it's where the actual work happens.
The Mistakes Don't Look Like Mistakes
AI errors rarely look like typos. They look like sentences. A fabricated case citation reads with the same rhythm as a real one. An invented quarterly figure sits inside a paragraph that's otherwise correct.
A summary of a long document drops the one clause that changes the meaning. According to MIT Sloan, judges around the world issued hundreds of decisions addressing hallucinated content in court filings between 2023 and 2025, with the volume climbing sharply in the most recent year on record.
That climb is the story. The tools got better, and the mistakes kept slipping through anyway because they don't wave a flag. They wear the same coat as the accurate sentences around them.
There's a useful guide on how to minimize errors when you use AI at work that makes this point plainly: the check has to be aimed at substance, not polish.
The errors are structural, not clerical. A model predicts likely next words. It isn't checking a database of truths before it writes. When it doesn't know, it produces the shape of an answer instead of the answer, and the shape is usually convincing.
Why Proofreading Isn't the Fix
The intuitive reaction is to proofread harder — read it twice, send it to a colleague, run it through spellcheck and hope something jumps out.
None of that catches the real problem, because proofreading is tuned to surface: grammar, tone, transitions, awkward phrasing. Hallucinations sit underneath the surface. The sentence is grammatical. The tone is right.
The citation exists in every respect except the one that matters, which is whether the case, study, or number is real.
Reading it twice with the same eyes also produces the same blind spots. You already know what the paragraph is supposed to say, so you read what you expect. This is the same reason writers miss their own typos for years.
A second read that isn't structurally different from the first is really just a slower first one.
Build a Verification Pass, Not a Proofread
What works is a second pass built to look for the specific things AI gets wrong. Harvard Business Review's framework for generative AI argues that the higher the cost of an error, the more the workflow needs a human standing between the model and the reader. That's the frame.
The second read isn't a favor to the draft — it's the control that makes the draft usable.
A workable version has four moves:
- Check the checkable. Every number, name, date, quote, statute, case, and URL gets clicked or looked up—not skimmed, opened. If a source doesn't resolve on the first try, treat that as a warning, not a formatting issue.
- Read against the original. For summaries, put the source document next to the output and confirm the load-bearing sentences match. Watch for dropped qualifiers, flipped conditions, and softened caveats.
- Ask what's missing. Hallucinations of omission slip past most readers. If the draft doesn't mention a risk, exception, or counterargument you'd expect a competent human to include, that absence is the finding.
- Change the eyes. Read it aloud, paste it into a different app, or wait a few hours. Anything that breaks the pattern of the first read helps you see what your brain glossed over.
None of this takes long once it's a habit. The whole pass usually runs shorter than the time you'd spend apologizing for one bad citation.
Make the Habit Boring
The second read holds up when it stops depending on discipline. Put the checklist at the bottom of the template. Add a verification step to the workflow before the send button, not after. If a document is going to a client, a regulator, or anyone whose trust you'd like to keep, no AI-assisted draft leaves without a human whose job it is to confirm the facts.
The tools will keep getting faster. The reader on the other end will keep assuming a person stood behind the words.
The habit that closes the distance between those two things is small, repeatable, and almost aggressively unglamorous. It's also what separates the people who ship clean work from the ones stuck explaining a fabricated citation to their boss.
Leave a Reply
You must be logged in to post a comment.