Why Not Just Ask AI

People ask this constantly now, in slightly different words each time. Can't I just get AI to write my self assessment? Can't I paste in my case and let it draft the email to my manager? Why build a record by hand when a model can produce something in nine seconds?

The honest answer is: for part of this, yes. For the part that actually gets read, no. And the line between those two has nothing to do with being against AI. It has to do with understanding what a language model is actually doing when it writes for you, and what a review, a promotion case, or a disciplinary response is actually testing.

The part that's fine

Gathering the raw material is legitimate work for a tool. Dates, timelines, the order things happened in, a first pass at pulling scattered detail into one place — AI is genuinely useful here, in the same way a spreadsheet or a search function is useful. You feed it what actually happened. It helps you organise what actually happened. Nothing about that undermines your case, because nothing about that is standing in for your judgement.

This is the part people underuse AI for, if anything. Most people's record keeping problem isn't that they need help writing beautifully. It's that they haven't written anything down at all, and by the time review season arrives, six months of contribution has evaporated into "I think I did some good work on that project." A tool that helps you reconstruct a timeline from old emails and calendar entries is solving a real problem.

The part that isn't

Writing the actual case is a different task, and it's the one people keep trying to hand over anyway. A promotion case, a review response, a rebuttal to a disciplinary finding — these all need the same thing underneath: specific, defensible detail that only exists in your head, surfaced through genuine back and forth about your particular situation. Not a template with your name dropped in. Not a paragraph generated from three bullet points you typed in under a minute.

Here's the uncomfortable part. When a lot of people ask the same kind of model to solve the same kind of problem — "write me a paragraph about my leadership impact" — the outputs start to look like siblings. This isn't speculation; it's been measured directly. Studies on writing assisted by AI have found that when different people write with the same model, their finished pieces become measurably more similar to each other — the model smooths out the particular phrasing, structure and emphasis that would otherwise mark a piece of writing as unmistakably yours. A separate review synthesising more than 130 studies on this went as far as describing the collective effect on written and spoken language as a genuine risk to individual expression. That's not a minor stylistic quibble when the entire point of the exercise is proving that your contribution was distinct.

What "falls apart under scrutiny" looks like

This shows up most starkly in disciplinary and grievance work, which is where I spend a lot of my time. Legal commentary internationally has noted a sharp shift over the past couple of years: submissions that used to be short and factual are now arriving longer, more "legalised," and — this is the part that matters — often thinner on the actual detail that makes a case hold up. One industry survey found that the large majority of grievance and disciplinary submissions assisted by AI that employers had dealt with contained some degree of inaccurate or outright incorrect information. Generic language reads as generic because it is: it was built to sound plausible on the first pass, not to survive a further question about a specific date, a specific conversation, a specific number.

The same failure mode applies to a review or promotion case, just with lower immediate stakes and the same underlying problem. If your case can't survive your manager asking "can you give me an example of that," it wasn't ready to submit.

What this looks like from the employer side

Advising employers through these processes, I keep seeing the same three patterns, and they're worth telling apart.

The first is the exaggerated claim. Not necessarily dishonest — usually built on confident, fluent advice that came from a tool with no knowledge of the actual jurisdiction, the actual policy, or the actual relationship in front of it. It reads well and applies to nowhere in particular. A fluent case built on the wrong premise doesn't survive contact with someone who knows the real rules.

The second is the case that had genuine leverage and lost it immediately. A real concern, presented in a tone or structure that's obviously been borrowed rather than earned, spends whatever leverage it had proving it's genuine before it ever gets to the substance.

The third is the one worth sitting with longest: cases that could have been genuine — and probably were — but arrived so smoothed into confident, generic phrasing that the actual person underneath disappeared. Not dishonest. Just unrecognisable as belonging to someone with an actual, specific problem.

What's changed, watching this from the employer side, is where attention now goes. When a good share of what lands in an inbox reads like it came from the same drafting tool, the submissions that read as genuinely, specifically written — imperfect phrasing and all — stand out immediately, and they get read differently as a result. Nobody assessing a claim is grading for polish. They're matching what's in front of them against actual criteria: what happened, when, who was involved, what the pattern has been, what the policy says. An account that sounds like a real person describing a real situation, rough edges included, maps onto that far more cleanly than confident generic language that has to be untangled first to find out if there's anything underneath it. Imperfection reads as ownership. Polish, on its own, increasingly reads as absence.

That's worth separating from voice, because it isn't really a writing style problem. A distinctive voice built with AI, developed over genuine, sustained back and forth about how you actually think and actually speak, doesn't produce this effect — but that's a slower and more deliberate thing to build than typing three bullet points into a chat window. What produces the effect above is the default: fast, unedited, first draft, which is what most people are actually sending.

The credibility cost

Here's the part that should give anyone pause before pasting a self assessment straight from a chatbot: the research on how people respond to text written with AI is now extensive, and it isn't ambiguous. Large studies on what's been termed the "AI disclosure penalty" have found that once a reader believes something was written with AI help, they rate it as less authentic and trust it less — even when the underlying words are exactly identical to something written entirely by a person. The effect isn't about quality. It's about authorship. And in employment contexts specifically, that perception has been linked to lower ratings of competence and warmth, and to being seen as less appointable.

Meanwhile, the people reading your case are getting better at spotting it, not worse. AI use inside HR and management has grown fast enough that a meaningful share of managers now openly admit to having used it themselves to draft reviews, which means they're reading everyone else's submissions with a trained eye for the telltale phrasing. Separate research has found that a substantial share of employees already suspect their own review was generated by AI, purely from how it reads. The tools that would supposedly make your case for you are being written and read by an audience that increasingly recognises their own fingerprints.

Accuracy is not arrogance — but generic language reads as neither accurate nor considered. It reads as absent.

If you're working outside a closed system

There's a second problem that has nothing to do with writing quality: what happens to the information you type in. If you're pasting real project detail, client names, figures, or the substance of a difficult conversation into a public AI tool that isn't part of your employer's managed environment, you're sending that material somewhere outside your control. Security research on this has found that the large majority of employees who use generative AI at work admit to pasting sensitive material into it, and a meaningful share of that material includes personally identifying or otherwise confidential information. Some of the material people are documenting for a review or a case — performance figures, client names, details of a workplace dispute — is exactly the category of thing that shouldn't leave a closed system in the first place. That's before you even get to whether the writing is any good.

If you're working inside one

Using your employer's own enterprise AI account solves the confidentiality problem and creates a different one. Your prompts and drafts typically sit inside logs the organisation itself can access — which is worth knowing if what you're drafting is a rebuttal to that same organisation. And because everyone on that licence is prompting the same model with structurally similar requests — "summarise my contributions this quarter," "help me write my self assessment" — the outputs tend to converge in exactly the way the research above describes. Nobody copied anyone. But a reviewer reading through a stack of self assessments that all land on the same three adjectives and the same sentence rhythm doesn't experience that as originality. It reads as template, even when it technically isn't.

What actually holds up

None of this is an argument against using AI. It's an argument for using it in the one place it earns its keep — pulling your own record together — and staying out of the place it can't do the job, which is making the case that only you can make, in language that actually sounds like you, built from detail no model has access to.

That second part needs something a chatbot with no context can't provide: real back and forth, trained on your specific situation, that surfaces the detail generic prompting never will. That's the gap our own tools are built to close — not by writing your case for you, but by asking the questions that get the real material out.

---

FILED.