AI-generated job applications: how to spot them, and what works instead

Not reliably. There is no signature to find, because a language model leaves no mark in the words it produces. A detector measures how predictable the wording is and returns a probability.

A Stanford study of seven widely used detectors, read as the arXiv paper on 30 September 2026, ran them over 88 US 8th-grade essays and 91 essays by non-native English speakers sitting the TOEFL. On the school essays they were near perfect. On the TOEFL essays the average false positive rate was 61.22 percent, all seven agreed that 18 of the 91 were AI-written, and 89 of the 91 were flagged by at least one. Rewriting genuinely generated essays in more literary language then dropped detection from 100 percent to 13 percent.

Both errors at once, and both the wrong way round: it accuses the careful second-language writer and clears whoever ran one more rewrite. The vendors say so themselves. GPTZero's own FAQ, opened the same day, says "we don't believe that any AI detector is perfect", that edge cases exist "where AI is classified as human, and human is classified as AI", and that results "should not be used to punish students".

Writing confidence, which for most roles is not the job. Consider who a style filter catches first.

• Someone writing in their second or third language, taught to write correctly rather than distinctively.
• Someone applying for a first job, using the structure a careers service gave them.
• A careful person working from a template, which is what you reach for when the expectations are unstated.

Then consider who it misses: whoever has the most practice at prompting produces the least detectable text. So the filter runs against the diffident and in favour of the fluent, and it tracks who learned English at home, which is a question for your own counsel.

Polished prose stopped carrying information around 2023, and an hour grading it is an hour not spent asking a question. More in how to reduce bias in candidate screening.

Plenty, and none of it is style. Sort every line into three buckets.

Swipe the table sideways to see every column.

Read the document as plain text once: instructions aimed at an automated screener only show up when the styling is gone. Ignore the dash-and-adjective folklore.

The test for the middle bucket survives every change in writing tools: is this a claim the applicant should be able to expand on? Every first-person verb promises a week behind it, and if the week is missing the claim is thin whoever typed it.

Pick one claim and ask about one week. Not the role, not the project: a week.

• "Which part did you build, and who built the rest?" Real work has a boundary.
• "What broke, and how did you find out?" Invented work contains no failures.
• "Who disagreed with you, and what happened next?" A person, a decision, an outcome.
• "What did you decide not to do?" Trade-offs live in a memory, not a summary.
• "Walk me through the week it went live." A real week has an order and a surprise.
• "What would you do differently, and what did it cost?" Unreachable from the outcome.

A model can write about a project; it cannot remember one. What it produces is an outcome with the causal chain removed: no sequence, no surprise, no colleagues. So the person who did the work gets more specific under pressure, and the person working from a document gets vaguer, because abstraction is the only way out.

Ask everyone the same questions and write the answers down, which is what turns it into evidence. The format is in the structured interview scorecard.

Two applications read identically well. One belongs to an engineer who asked a chatbot to tighten her wording.

The other belongs to someone who has never run a migration. The documents cannot separate them. Ninety seconds of follow-up can.

The claim is dissolving when

• The answer broadens every time you narrow the question.
• No proper nouns: no tool, no team, no customer, no month.
• "We" never becomes "I", or "I" turns out to have been thirty people.
• A number arrives with no denominator.

It is only polish when

• A boring, specific answer arrives at once, sometimes stressing the wrong part.
• They correct your understanding of their own project.
• They volunteer what went wrong before you ask.

Hold it loosely. There are honest reasons to be vague: a customer nobody may discuss, a job five years ago, nerves, a second language. Ask again from another angle, and never make it an accusation.

Write the rule down, in the advert and in the invitation. Candidates using AI assistants in interviews are guessing at your policy, and the guess runs from "anything goes" to "a detector is watching".

1. Preparing with AI is fine, the application included.
2. The interview is in your own words, live. Say whether notes are allowed.
3. We will ask follow-up questions about specifics. That sentence moves the reward from writing well to having done the thing.

Your applicants are already reading the platforms' own policies, which contradict each other: can you use ChatGPT in an AI interview quotes each vendor's published rule. Do not ask for a declaration you cannot check, and never claim a detection ability you do not have.

Rarely, and the two need different responses. An honest applicant using ChatGPT to write more clearly is the ordinary case. Fraud is a smaller set: a stolen identity, employers that do not exist, a proxy sitting the interview, deepfaked video.

Those carry checks that belong before an offer, and they are covered in fake job applicants and candidate fraud, including how to identify fake candidates in recruitment without treating everyone as a suspect. The overlap is cheap: a follow-up question and a recording to compare with a later round help with both.

Nothing, deliberately. Career1 runs no detector on a resume, an application or a transcript, and publishes no score claiming to be the probability that a machine wrote something. That number could not be defended to a candidate who asked.

What happens instead is a conversation. Applying starts it: the AI recruiter speaks with each applicant in the browser for about eight minutes, in English. It follows up whenever an answer is vague, impressive or suspicious, digs into the projects on the resume by name, and asks how and why rather than for definitions.

What comes back is a report, not a verdict: a match score, an overall score, a recommendation from strong hire to no hire, scored sections, risks and questions for your own round. The part that matters here is the skill ledger.

Every skill claimed on the resume is marked demonstrated or not, with either a quote from the transcript or the words "not probed", and that second option is the honest one: eight minutes never reached the claim, rather than the person lied. A complete sample interview report is published, with the transcript and recording beside it.

There is no reject action. Career1 ranks and recommends; a person on your team decides. Plans are on the pricing page, from a free tier of two a month to $199 for 150.

• Eight minutes tests some claims, not all. It decides who you speak to next, not who you hire.
• The interview is English only, which is its own constraint on a page about second-language writers.
• Career1 verifies no identities: no document checks, no face or voice recognition, no deepfake detection, no screen monitoring.
• No published accuracy or bias measurement, ours or anyone's. A readable transcript instead.
• Want proctoring or a detector in round one? Wrong tool.

Both were opened on 30 September 2026. No figure here is second-hand.

• Liang, Yuksekgonul, Mao, Wu and Zou, "GPT detectors are biased against non-native English writers", arXiv:2304.02819v3, 10 July 2023: arxiv.org/abs/2304.02819. The paper states that its published version appears in Patterns.
• GPTZero, "What are the limitations of the classifier?": gptzero.me/faq, quoted for the vendor's own stated limitations only.

Not cited on purpose: the survey percentages and detector accuracy claims circulating in the pages that currently answer this search, none of them traceable to a method we could read.