How to use AI on a job application without sounding like AI
AI detectors are unreliable in both directions. Here is what a model is genuinely useful for on a job application, and what has to stay yours.
After four revisions of the same cover letter, it is hard to judge how it reads. The paragraph about the eleven months you were out of work reads as defensive or as breezy depending on which hour you read it. And somewhere behind that sits a second worry: that if you hand any of this to a chatbot, the person on the other end will be able to tell, and being caught will cost you more than the clumsy version ever would have.
That worry deserves a straight answer. Part of it is real and part of it rests on tools that do not work as advertised. What follows is where a model genuinely helps on an application, and where it is worse than useless. The detector question comes first, because it is the one that stops people writing.
The detector question, and the only numbers worth quoting
Start with the tools, because most advice on this topic quietly assumes they work. The clearest evidence that they do not comes from the companies and institutions with the least reason to say so.
OpenAI published a classifier for AI-written text on 31 January 2023 and was blunt about it in the announcement: "Our classifier is not fully reliable." On its own challenge set of English texts, the tool "correctly identifies 26% of AI-written text (true positives) as 'likely AI-written,' while incorrectly labeling human-written text as AI-written 9% of the time (false positives)." A note now sits on that same page: "As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy." The company's help centre still answers the question "Do AI detectors work?" with "In short, not in our experience," and adds that when OpenAI tried to train one, "we found that it labeled human-written text like Shakespeare and the Declaration of Independence as AI-generated."
The failure lands unevenly. Stanford HAI reported in May 2023 on a study of seven detectors that found them "near-perfect" on essays by US-born eighth-graders while classifying 61.22% of TOEFL essays by non-native English students as AI-generated, with 97% of the 91 essays tested flagged by at least one detector. Vanderbilt University disabled Turnitin's AI detector in August 2023 and did the arithmetic publicly: at the 1% false positive rate Turnitin claimed at launch, of the 75,000 papers Vanderbilt submitted in 2022, "around 750 student papers could have been incorrectly labeled as having some of it written by AI."
Now the other direction. The same OpenAI help page says that "even if these tools could accurately identify AI-generated content (which they cannot), students can make small edits to evade detection," and the Stanford write-up notes that detectors are easily subverted by asking a model to rewrite a passage in more sophisticated language. So the tools flag writing that a person laboured over and wave through writing that a model produced ninety seconds ago.
Two caveats, because those figures are from 2023 and vendors say their products have improved since. We looked for an independent, current accuracy figure worth printing and did not find one, so no percentage in this article claims to describe where detection stands in 2026, and any number published by the company selling the detector should be read as marketing. The research also sits in classrooms rather than in hiring, though the mechanism is the same: a statistical guess about the texture of a piece of prose.
What this means at your desk is small. Do not run your letter through a detector for reassurance, because a green score says nothing about whether a human reviewer will believe you, and do not rewrite a sentence you are proud of because a score went red. If an employer runs your application through one of these tools, a bad result is a coin flip weighted by how you learned English, and the defence against that is an application somebody can check. A plainer paragraph does nothing for you there.
Where a model earns its place in an application
The useful jobs are the ones where the model is reading rather than writing.
The first is decoding the advertisement. Job ads are written by committee and contain two layers: the requirements that actually decide the shortlist, and the boilerplate that got pasted in from the last vacancy. Paste the ad and ask which lines describe the work and which describe the template. Ask what the ad is anxious about, since a role that lists "comfortable with ambiguity" three ways is telling you something about the team. This costs one message and changes what you emphasise.
The second is structure. Write 600 unstructured words about what you have done, badly and quickly, with every date and number you can remember. Then ask for an order rather than for prose. A model is good at noticing that your third paragraph contains the strongest thing you have said and that it is buried.
The third is pressure-testing, and it is the highest-value move on this list. Ask: which claims in this letter would a sceptical interviewer ask me to evidence, and what evidence would satisfy them? Anything you cannot answer comes out now, at your desk, instead of in a room with two strangers and a printed copy of what you wrote.
The fourth is the dead sentence. You have read the opening line so many times it has stopped meaning anything. Ask for four plainer versions, keep the meaning fixed, and pick one or write a fifth yourself.
Part of the application | Worth asking a model | What has to stay yours |
|---|---|---|
The job advertisement | "Which requirements here are load-bearing, and which are template?" | Deciding whether you want the job it describes |
Your messy brain dump | "Propose an order for this. No prose yet." | Every fact in the dump |
The achievement paragraph | "Which claims would a sceptical interviewer ask me to evidence?" | The evidence, and the number |
One sentence you have killed | "Four plainer versions, same meaning." | Choosing the one you would say out loud |
The "why this company" paragraph | Nothing | All of it |
Your CV, as text | "What would a stranger think I do for a living?" | Titles, dates, employers |
Nobody can invent the reason you want the job
The model produces "I am passionate about your mission" because it has been handed nothing else to say. Human applicants write that sentence constantly, because it is what anyone produces with nothing specific in hand. The difference is that a model has nothing specific about you unless you supply it, so left alone it will reach for the generic register every single time.
Three kinds of content have to come from you. Start with the detail only you were there for: the week the system went down and you were the one still on the phone at seven, or the process you quietly fixed because it had annoyed you for a year. Then anything a recruiter can check, which means dates, titles, headcounts, revenue figures, tool names and languages. A model will supply plausible versions of all of these if you leave a gap, and plausible is exactly the problem, since fabricated detail reads perfectly until someone rings your old employer. (Our explainer on why models invent facts covers why this happens and why no setting turns it off.) Last, the reason you want this job in particular. Two true, specific sentences there outperform a paragraph of admiration, and that is the part an interviewer is most likely to read back to you.
So read the output line by line against your own facts before anything goes near a form. Every qualification, every certification, every claimed proficiency, every percentage. It takes four minutes and it is the difference between a letter you can defend and one you have to remember.
Read it out loud, then hear it in a second voice
The out-loud test does most of the work that detector tools pretend to do. Read the letter to the wall. Any sentence you would not say to a person goes, along with any word you would not use in the interview. Nobody has ever said "robust track record" aloud in a room.
After that, a second model is genuinely informative, because two of them given the same paragraph produce noticeably different registers, and hearing your content in another voice tells you which parts were ever yours. In MultiChats that action is called "Try Again" on the web and "Branch here" in the mobile app, and it opens a new thread with the conversation carried over, so the first answer stays where it was. Watch the mobile labels: "Try again" on a phone is regenerate, which deletes that answer and everything after it.
Here is what to listen for, in your writing and in the model's.
Pattern | Why it reads as generated | What to do instead |
|---|---|---|
"I am passionate about your mission" | Any of 400 applicants could send it | Name the specific thing, and when you met it |
"It is not just X, it is Y" | A rhetorical shape models reach for far more often than people do | Say the thing you mean, once |
Lists that always arrive in threes | Padding to a rhythm rather than counting real items | Keep only the items you can evidence |
Paragraphs of identical length | Nobody writes evenly while nervous | Let one paragraph be two lines long |
"Leverage", "spearheaded", "utilise" | A register that exists only in documents | The word you would use in the interview |
A closing paragraph that restates the letter | Models summarise by reflex | End on the ask |
If a form asks whether you used AI
Answer it. Employers are all over the place on this right now: some ask, some forbid it outright, some encourage it and want you to say how, and plenty have no policy at all. The question belongs to them and the answer belongs to you, and there is no version of this article that advises you to lie on an application form.
There is a whole industry selling "humaniser" and bypass tools whose entire promise is that a detector score will change, and the first section of this page is the reason that promise is empty. The defensible position costs nothing and survives contact with an interview: everything in the application is true, you can talk fluently about all of it, and the words are ones you would use. Keep your own first draft in the thread. If someone asks how the letter came together, "I wrote out everything I had done, had a model help me order it, then rewrote it in my own words" is a completely unembarrassing sentence to say.
Before you paste your CV in
A CV is a small dossier: home address, phone number, sometimes a date of birth, and the name of everyone who has employed you. Strip the contact block before pasting. It adds nothing to a rewrite and it is the part you would least like sitting in a chat log.
On the wider data question, MultiChats does not train on your conversations. What the provider behind a given model does with them varies by model, and the six models where we know of a difference carry a notice on their info card. Grok's puts it without decoration: "xAI may use your conversations to train Grok models." The other 54 cards say nothing, so look before you paste and do not read silence as reassurance.
If you would rather attach the file, 25 of our 60 models accept a PDF attachment, at one file per message up to 10MB and three per day. A CV you scanned or photographed into a PDF is transcribed first if it runs to 20 pages or fewer, and transcription can misread a name or a date. Pasting the text is the reliable route when you are not sure which kind of PDF you have.
The free plan runs 50 messages per 24 hours across 21 of the 60 models from 18 providers, which is more than one application needs. Web search, which is how you read what a company has actually published before writing the paragraph about why you want to work there, is 2 searches per 30 days on the free plan and sits behind Pro at the button in the mobile app.
A disclosure, since this page tells you what to spot
This article was held to the list in the table above. Every pattern in it was on the checklist before the first draft, which is why nothing here arrives in threes for the rhythm of it and why the page stops without a paragraph restating itself. Where a sentence here still reads as machine-made, that is a fair thing to hold against us.
Open the letter you have already rewritten four times, paste it in, and ask for an order instead of a rewrite. Free to start, Pro at $20.99/mo. Start a chat.