Pasting a Written Answer Costs You the Follow-Up

Check the Employer's Stated Rule First

A published rule for that round settles the question, so look for one on the posting, the invitation and the careers page before you write a word. Anthropic's candidate guidance asks for a first draft written yourself and refined with Claude on the written application, and rules out AI on take-home assessments and during live interviews unless it says otherwise ³. That is one employer's policy, dated July 2025, not an industry norm.

The UK Civil Service publishes the same boundary in two lists. Refining and clarifying your own ideas, researching the organisation, and checking spelling, grammar and clarity are acceptable. Inflating or inventing skills, pasting generic AI responses into an application, and using AI to complete personality and ability assessments are not ². It is recruitment guidance for one employer, it carries no date on the page, and it binds nobody else.

The Civil Service draws its line at the accuracy of your claims rather than at whether a tool touched the text, and Anthropic at who writes the first draft.

Sometimes the platform states the rule instead of the employer. Sapia's candidate explainer tells candidates not to copy answers, and says its chat interview checks for plagiarism from internet sources and can detect whether answers come from content generators such as ChatGPT ¹. The page gives no method, no accuracy figure, and no account of what happens to a flagged answer. Read it as the vendor's stated position rather than as a demonstrated capability.

A vendor's description of its own product is a claim about it, not an audit of it.

Whether using AI to apply for jobs is cheating turns on whether a rule was stated at all, and it is where to start when no rule is published anywhere you can find.

What Does the Written Round Do With Your Answers?

Where a later round follows, it can supply the material that conversation is built from. A written screen is not always the decision by itself, and when the transcript travels to whoever runs the next round, the specifics in your answers can become the questions you get asked. That is why a pasted paragraph is expensive even when nobody objects to it. You have handed over claims you did not make.

A chat interview with no camera and no timer reads like a form. It is a screening interview conducted in text.

Employers are told this plainly from their own side. The guidance on whether candidates are using AI during live interviews makes the point in the employer's own language. An assistant holds nothing about a decision the candidate personally made, so it fills the gap with invention, and the invention does not survive a follow-up. None of that reasoning depends on catching anybody. It depends only on someone asking a second question about something you wrote.

The written round is where those second questions come from.

A generated answer written with nothing of yours to work from is fluent and unspecific. It supplies a rubric-shaped situation, task, action and result, with a stand-in where your own detail belongs. A follow-up aimed at the stand-in finds nothing behind it, because the model had your prompt and the job posting and was not in the room. The thinness reads as a weak candidate long before it reads as a dishonest one.

That cost lands whether or not a single check runs, and no policy has to be broken for it to land.

Why Do Pasted Answers Collide With Each Other?

Because applicants prompting the same models with the same question can get back much the same phrasing. Vervoe's help documentation says its plagiarism check compares the first 500 words of every text answer across candidates and flags duplicate responses ⁵. A paid add-on also matches content from ChatGPT and websites and shows the source next to the flag. That is documentation for one platform, and several of those features are optional or paid.

Published prompt packs aimed at one platform hand the same prompts to everyone who uses them.

The flag is not the interesting part of that. What a duplicate check measures is that an answer belonged to nobody in particular, and a person reading the transcript can see that an answer holds no detail of yours without any tooling. An answer that could have come from any applicant tells the round nothing about you, which is the one thing it was asking for. The check only makes that visible earlier.

Text detectors are not the thing to plan around, in either direction. A study tested twelve public AI-text detection tools plus two commercial systems used in academic settings ⁴. It concluded that the available tools are neither accurate nor reliable, and that they lean toward calling text human-written. That was a 2023 test against the output of that era, in an academic-integrity setting.

A detector that does not work cannot repair an answer that is not yours.

Write It Yourself, Then Use the Model to Tighten It

Draft from your own memory first, then use a model to cut it down, fix the grammar and make the point land. That order keeps the claims yours and matches what both published policies above allow on a written application ³². Neither treats that as covering every stage, so a rule stated for the round itself comes first. It also produces the version you can defend out loud, which is the only property that carries into the next round.

The test to run before you submit is narrow. Take any sentence in the box and picture being asked about it next week: where it happened, who else was involved, what you would do differently now. If a sentence has nothing behind it, it is not yours to submit yet. That test works whether or not the employer checks anything, and whether or not a rule exists.

It is also the test a good interviewer is already running.

The measured downside is largest when someone else does the exposing. In an online experiment with 195 US panel members, a professional was trusted least when a third party exposed his use of generative AI ⁶. He was trusted more when he disclosed it himself, and most when nothing was said ⁶. The scenario was a freelance tax adviser and his client, not an applicant and an employer, so the ordering is what transfers.

None of that makes silence a plan: whether someone else ever raises it is the branch you control least.

The same question with an explicit rule attached is covered in what happens if the take-home says no AI, and it starts in the same place: read what the employer actually asked for.

See how it works >