Telling an AI Scorer to Rate You Highly Lands in the Transcript

Why the Resume Version Doesn't Transfer

The trick being copied is a written one, and an interview leaves it nowhere to sit. Hidden instructions on a resume live inside the file: white text on white, a two-point font, a line placed where no eye stops. Speech carries no formatting. Whatever you say to an AI interviewer arrives as words in a transcript, in the same plain type as the rest of your answer, on a file with your name on it.

The written version is common enough to have been surveyed. Greenhouse's November 2025 AI in Hiring Report, a hiring-software vendor's survey of 4,136 job seekers, recruiters and hiring managers across the US, UK, Ireland and Germany, found 41 percent of job seekers admitting to prompt injections, hidden text designed to bypass AI filters ¹. Read that as an admission rate on a survey panel rather than a count of real applications. The page gives no fielding dates and no margin of error, does not say whether that figure is US-only or pooled across the four countries, and nothing in it shows the tactic changing an outcome.

What the survey measures is how many people say they have tried it.

Keywords typed in white text on a resume fail in a way that matters here. Extraction ignores color, so the hidden block lands in the parsed record, and that record is the version a recruiter opens. The trick gets read by the person it was built to fool.

The mirror image of this lands on the employer's desk. The advice they get on hiding an instruction in a job posting is to skip it, because a planted line sorts applicants by how closely they proofread rather than by anything the job needs.

What Does the AI Actually Score?

One vendor's written answer is the words you say, transcribed, and nothing else. HireVue's 2025 Explainability Statement says its AI relies only on what the candidate says, using no video analysis and no other audio characteristics, and specifically not facial expressions, body language, background or tone of voice ². The pipeline it describes runs speech to text, then a natural-language model, then a score benchmarked against expert human raters' evaluations of answers to similar competency questions ².

That is a company describing its own product, and it is not an audit. It covers one vendor's AI-scored interviews and says nothing about what another system reads. What it does describe is the input on that platform: the transcribed words are the material the score is built from. The statement does not say what its model does when the words it transcribes are an instruction rather than an answer. Nobody outside the company can see that, and one vendor's answer would say nothing about another's.

The instruction does not sit beside your answer; for that stretch of the transcript it is your answer.

There is one measured place where a model's preferences bend, and no spoken sentence reaches it. In a controlled resume experiment, published so far only as a preprint, large language models preferred resumes generated by themselves over human-written ones, with content quality held constant. Self-preference against human-written resumes ran from 67 to 82 percent across major commercial and open-source models ³.

In simulated pipelines across 24 occupations, candidates using the same model as the evaluator were 23 to 60 percent more likely to be shortlisted than equally qualified applicants who had submitted human-written resumes ³.

Those pipelines were built by researchers, and no real candidate was shortlisted or rejected. The largest disadvantages showed up in business fields such as sales and accounting ³. The effect depends on the employer screening with a model, and on you happening to have used the same one.

Neither condition is something a candidate can see.

Who Reads What You Said?

A person on the employer's hiring team does, later, on at least one platform that says so in writing. Alex, the AI interviewer formerly named Apriora, says in its candidate FAQ that it makes no final hiring decisions and does not automatically reject candidates ⁴. The same page says its recordings, transcripts and notes are shared with the employer's hiring team, where humans make all hiring decisions ⁴.

What the employer does with that file is outside the vendor's control, and the page does not say how long any of it is kept. What it states is the destination. The transcript reaches the hiring team attached to your name.

Whether anyone opens the recording is a different question, and whether a human watches your AI-scored interview has its own answer. The short version is a split: a score narrows the pool, and a person makes the shortlist from what is left, usually reading the file.

The words are in the file either way.

At least one employer publishes what can follow AI misuse, though not this case by name. The UK Civil Service tells candidates that AI must not be used during any live assessment at any stage of the selection process, and that applications may be rejected where AI is used inappropriately at any stage ⁵. Before submitting, candidates confirm that the information they provide is true and accurate ⁵. That is a stated power rather than a measured outcome, with no rate, no examples and no appeal route on the page.

One employer's written rule is not evidence about anyone else's.

Put a Specific Thing in the Transcript Instead

Say the thing only you could say, because the content of the transcript is what gets read. Name the situation, the decision you made inside it, what it cost, and how you found out whether it worked. An answer about the migration you rolled back at the wrong hour carries information an instruction cannot, and it holds up when the next question digs into it.

The employer-side guidance on scoring an answer a candidate produced with AI points the same way, rating four acts separately: what you framed before generating anything, what evidence you demanded, what you refused from the model's output, and what you checked outside the conversation.

Polish counts for nothing in that rubric.

Being found out has been measured too. In a 2025 vignette study with 195 US panel members, a professional was trusted least when a third party exposed his undisclosed AI use ⁶. He fared better when he had disclosed it himself, and best of all when nothing was said ⁶.

The scenario was a tax adviser and his client rather than an applicant and an employer, so the ordering is what carries over and not the size of the gaps.

A spoken instruction never needs a third party to expose it: it is in the transcript the hiring team reads.

If the worry underneath the question is that the format will read you wrong, the useful move is not aimed at the model. Ask who reviews recorded answers and at which stage. Ask whether a written or live alternative exists, especially if the recorded format is a barrier for you because of a disability or a medical condition. Both are ordinary questions, and a person can answer them.

See how it works >