Personalization vs. making things up
The invented details are always the most convincing ones
· 5 min read · Career
Ask an AI tool to make your CV "stronger and more specific," and it will. Here's a line from a real-looking CV, before and after:
Before: Led the migration of payment services to AWS.
After: Led the migration of 40+ payment services to AWS, cutting infrastructure costs 35% and managing a team of 12.
The second version is better in every way a recruiter would notice. It's specific, quantified, and shows scope. It is also, possibly, invented from top to bottom. Where did 40 come from? Or 35%? Or twelve people? Not from you. You never wrote them down.
That's the problem with AI personalization, and it's the most important thing to understand before you let any tool rewrite anything with your name on it.
Two things that look identical
Personalization reshapes what's true to fit a context. It reorders your experience, translates your vocabulary into the employer's, surfaces the relevant parts, cuts the irrelevant ones.
Making things up adds what isn't true. A number, a tool, a scope, a title.
From the outside, the output of both is indistinguishable: fluent, specific, tailored text that sounds like a stronger version of you. The only way to tell them apart is to check each claim against what actually happened — and that's precisely the step people skip, because it's their own CV and it sounds right.
Why the model does this
This isn't a malfunction. It's the model doing exactly what it was asked, in the way it's built to.
A language model produces the most plausible continuation of what it's given. Ask it to make a vague bullet "more specific," and the most statistically plausible way to be more specific is to add specifics. Numbers are what strong CV bullets contain, so it produces numbers. It has no way of knowing which ones are yours, and — crucially — no internal signal that a number is invented rather than recalled. Both come out with the same fluency.
So the more you ask for impact, the more impact it will manufacture.
The fabrications are the convincing parts
Here's what makes this dangerous rather than merely annoying: invented details are almost always the most persuasive parts of the rewrite.
A real, vague bullet is weak. A fabricated, specific one is strong. The tool's worst output is its most impressive-looking output, which means your instinct to keep the best-sounding version actively selects for fiction.
And the cost lands later, in the one place you can't edit: the interview. "Tell me about those 40 services — how did you sequence them?" There's no answer, because there were never 40 of anything.
What to look for
When any AI tool hands you a rewritten CV, cover letter, or interview answer, these are the things it's most likely to have invented:
- Numbers — percentages, counts, currency, team sizes, timelines
- Titles — "Senior" appearing, or "Lead," or "Head of"
- Tools and technologies — especially ones from the job description
- Scope — "global" for regional, "led" for contributed
- Dates — quietly adjusted, often to close a gap
- Credentials — certifications or degrees that sound right
- "We" becoming "I" — team outcomes absorbed as personal ones
Diff the output against your original, line by line. For every specific claim, ask one question: do I know where this came from? If you can't source it, it comes out — no matter how good it sounds.
Why "don't make things up" isn't a safeguard
Most tools handle this with an instruction in the prompt: do not invent facts. It helps. It doesn't solve it.
An instruction to a language model changes a probability, not a rule. It reduces how often fabrication happens; it can't make it impossible, because the model has no reliable way to detect its own invention in the first place. For creative writing, a lower rate is fine. For a document you'll sign your name to and be questioned about, "usually doesn't lie" is not a standard anyone should accept.
Some facts need a hard guarantee rather than a polite request.
The rule we build to
At labor.quest we treat certain parts of your career as immutable: company names, dates, your original job titles, and your credentials. Those don't get rewritten, full stop.
And we don't rely on asking the model nicely. The check happens in code, after generation — the output is compared against what you originally gave us, and any change to a protected field is caught as a violation rather than shipped as an improvement. The model is allowed to rephrase, reorder, and translate your experience. It isn't trusted to preserve the facts, so something else does.
That's the general principle, and it applies well beyond us: anything that must never be wrong should be enforced by a deterministic check, not by an instruction to a probabilistic system.
The test that still works
Whatever tool you use, the oldest test in CV writing still applies to every line of AI output: could you talk about it for five minutes, out loud, to someone who does this for a living?
If the answer is no, it doesn't matter who wrote it — you, a friend, or a model. It's not personalization. It's a claim you'll be asked to defend, about something that never happened.
All posts