Where AI-Drafted Resume Bullets Drift From the Truth

AI-generated resume bullets rarely contain outright fabrications you would notice at a glance. They drift, in a small number of predictable directions, and each direction makes you sound slightly more senior, slightly more central, and slightly more measurable than you were. Learning the four or five recurring patterns is faster than reading every line with fresh suspicion, because once you know the shapes you can find them in seconds.

The reason the drift is directional rather than random is worth understanding. A language model is completing a pattern learned from resumes that were written to impress. Given a modest input, the most probable continuation is an impressive sentence. Nothing in the process pulls towards accuracy, because accuracy is not something the model has access to.

Scope creep

You supported one product line; the bullet says you supported the product portfolio. You worked on the onboarding email; the bullet says you owned the onboarding experience. The nouns get bigger.

Scope creep is easy to catch once you look for the widest noun in each sentence and ask whether you could name its boundaries. “Improved the checkout flow” is checkable — you can say which steps and which pages. “Improved the customer journey” usually is not, and if you cannot draw the edges of the thing you improved, you have inherited a phrase rather than described your work.

Ownership creep

This is the one to watch hardest. The input said you were on a team that migrated a database; the output says you led the migration. The input said you contributed to a report; the output says you authored it. Verbs like led, owned, drove, spearheaded and headed are all claims about your position relative to other people, and they are the claims most likely to unravel in a reference conversation.

The fix is not modesty for its own sake. If you led it, say so. The point is that the verb is a commitment you will be asked to explain, and an AI tool has no way of knowing which verb is true, so it picks the strongest one that fits the sentence.

Invented specifics

Ask for a bullet about improving a process and you will often get a bullet about improving a process by a percentage. The percentage came from nowhere. Sometimes the tool inserts a placeholder like “by X%” and you fill it in, which is fine. Sometimes it inserts an actual figure, which is not, and the figure looks exactly as plausible as a real one because plausibility is what it was generated for.

The same applies to timeframes, team sizes, budget figures, and user counts. Any number in a generated draft should be treated as absent until you have replaced it with one you can source. Delete first, then decide whether you have a real figure to put back.

Invented specifics also show up as tools you did not use. If a draft mentions a system that was merely adjacent to your work, it will read as a skills claim and someone may interview you on it. Cut anything you would not want to be asked a follow-up question about, and be honest with yourself about how deep your exposure to each tool really was.

Borrowed vocabulary

Certain phrasings recur across generated resumes because they recur across the training data: cross-functional stakeholders, leveraged, spearheaded, seamless, robust, best-in-class, at scale. On their own they are only stylistically tired. The problem is that they are also imprecise, and imprecision hides drift. “Leveraged cross-functional relationships to drive alignment” could describe running a programme or attending a weekly meeting. A sentence that vague cannot be checked, which means you cannot check it either.

Replace the borrowed phrase with what you actually did in the plainest available words. The bullet gets shorter and more specific at the same time, and any remaining overstatement becomes visible.

Seniority inflation in the summary

The summary or headline is where drift is least constrained, because there is no specific event to anchor it. A three-year career becomes “seasoned professional”; a coordinator becomes “strategic leader”. This deserves its own attention, since an AI-written summary can quietly promote you in a way that contradicts the dates directly below it.

Reading a draft for drift

A practical pass, in order:

1. Circle every number. Delete any you cannot source. Do this before reading for sense, so a plausible figure does not become part of the story in your head.
2. Underline every verb of ownership. For each, ask who else would have to agree with it.
3. Find the widest noun in each bullet. Narrow it to something with edges.
4. Mark every named tool, system or method. Keep only what you could discuss for two minutes.
5. Read the summary against the dates. If the seniority claimed and the years shown disagree, the summary loses.

None of this requires you to distrust AI drafting in general. It requires you to treat the output as a proposal about your life written by something that has never met you. A proposal is a good starting point. It is not a record.

When a bullet cannot be rescued

Sometimes the underlying work was real but the generated sentence has nothing true in it, and editing turns into arguing. Delete it and write one plain sentence from memory instead. It will be less polished and considerably more defensible, and you can polish it afterwards. If you find you cannot write the plain sentence either, that is useful information about the claim, and there is a decision to make about what to do with it.