Write What You Heard, Never Why You Think It Happened

Which words in a screen note create the problem?

The words that create the problem are verbs of explanation. Every one of them puts a cause behind a behavior you did not watch: used, prepped, memorized, panicked, understood, cared. A screen gives you sound and words in a twenty-minute window, and nothing about how the answer was assembled the night before.

A note that claims otherwise asserts something the call never contained.

One line shows the whole class, and it is easy to type without pausing. Clearly used ChatGPT to prep; answers sounded rehearsed.

The test that catches it runs mid-call. Read the line back and ask whether a recording would settle it. Named three sources and could not say which one the number came from is settled by the recording. Did not really understand the tool is settled by nothing.

Every unsupportable line has a supportable one underneath it, usually shorter:

• Instead of *clearly used ChatGPT to prep*: repeated the phrasing from the careers page, twice, in two different answers.
• Instead of *sounded rehearsed*: gave the same example for two unrelated questions.
• Instead of *seemed nervous*: nothing. Cut the line. No part of it bore on a requirement.
• Instead of *not really an AI person*: after two follow-ups, could not name a task they had stopped handing to a model.
• Instead of *vague*: the vague sentence, quoted.

Manner is the sharpest edge here. A line about how a person spoke rather than what they said records something other than the answer: nervous, halting, rehearsed, hard to follow.

Those words move with a stammer, with anxiety that meets the definition of a disability, and with speaking a second language.

Since July 26, 1992, when the ADA's employment provisions took effect, it has been discrimination to use qualification standards, employment tests or other selection criteria that screen out or tend to screen out an individual with a disability. The defense is that the criterion as used is shown to be job related for the position in question and consistent with business necessity ⁶.

A note starts working like a criterion the moment a reader treats it as one.

Whether an AI notetaker's summary calling a candidate nervous and rehearsed can go in the hiring file is the same question asked of a tool, and the free-text box is where it gets asked of you.

Who reads a screen note after the call?

Someone who has the file and not the call, and usually much later than you expect. Federal rules already treat a screen note as a record. Any personnel or employment record an employer makes, hiring records and application forms included, has to be kept for one year from the making of the record or from the personnel action, whichever comes later ¹.

One year is the floor, and a charge moves it.

Once a charge of discrimination is filed, or the Commission or the Attorney General brings an action, the employer has to preserve all personnel records relevant to the charge until the charge or the action reaches final disposition ¹.

Relevant reaches well past the person who complained. The rule defines it to cover records on all other employees holding positions similar to the one the aggrieved person held or sought. The definition also reaches the application forms and test papers of every other candidate for the position that person applied for and was rejected from ¹.

One candidate's charge can pull in what you kept on everyone else who applied for that position.

California pushed the horizon out further. In the final text of its automated-decision rules, dated March 2025, the state's Civil Rights Council extends employment-records retention from two years to four and names automated-decision system data in the list of what is covered ⁵. The rule reaches California employers under state fair-employment law, and the four years run from the later of the record's creation or the personnel action.

Four years can outlast the screener who wrote the note.

The person explaining your sentence may not be you, and they will have the file and none of the room. What has to be kept from an AI-assisted hiring decision, and for how long is worth settling before the first charge rather than after one.

Why can't a note say the candidate used AI?

Because nothing in a phone screen establishes it unless the candidate says so, and the instruments sold for the job fail in a direction that matters. A candidate who prepared with a model, one who prepared with a friend and one who rehearsed alone in a car can produce the same fluent answer over a phone line. The note records a guess and reads back later as a finding.

The written form of the same guess has been measured. Seven GPT detectors were run over 91 human-written TOEFL essays by non-native English speakers. The average false-positive rate was 61.3 percent, and at least one of the seven flagged 97.8 percent of those essays as AI-generated ².

The figure needs reading carefully: a false-positive rate averaged across seven 2023 tools on academic essays, not an accuracy score and not a constant.

Which way the error runs is the part that matters. The same detectors were near-perfect on essays written by US eighth-graders ², so the misfire lands on people writing in a second language.

A screener guessing authorship from a call has no better instrument than those detectors.

Whether AI detectors work on resumes and cover letters is settled enough that no note should rest on one. Hiring managers rejecting candidates for sounding like AI is the same error one rung up the loop.

The observation underneath the guess is usually worth keeping. Repeated the phrasing from the careers page twice is a fact about the call. Could not say where the framework in the answer came from is a fact about the call.

Neither names a tool, and both are more useful to a hiring manager than a verdict about ChatGPT.

Something real does sit behind the suspicion, and a follow-up reaches it. Ask which part of the answer would change if the number were different. Ask what they could not settle in the time they had. Then write down what came back, in the words it came back in.

Replace the guess with the requirement

Anchor every line to something the role requires, and the note stops being an opinion about a person. Four fields do it: the question you asked, the words that came back, the follow-up you ran, and the requirement the answer did or did not evidence. The shape asks for no objectivity, only for what happened in the order it happened.

A screener can fill the four fields at minute nineteen.

Notes that name what the job requires carry signal about how a hire later does.

Researchers text-mined the post-interview notes written about 7,650 candidates who passed the interview and joined a large Chinese internet technology company. The more job-related capabilities an interviewer had named, the better that hire's later performance and promotion record, and the lower the turnover ³.

The limits are large. Only candidates who passed could be observed, so what the notes would have predicted about rejected people is invisible by construction.

The design is correlational, at one firm, in technology roles, and the effect is small. One standard deviation on the match between the notes and the company's own job analysis went with about a 2 percent rise in performance ratings ³.

Naming the capability went with better outcomes, by a little, in one building. Small as the result is, it points the direction a note can usefully go.

Measured gaps also concentrate where a role leaves most to judgment, which the authors read as discretion at work. A resume audit of 36,880 applications to 9,220 US job advertisements for new college graduates found callback gaps widest in roles combining high analytical and interpersonal demands with little routine content ⁴.

In management occupations, callbacks ran 28 to 43 percent lower for Black men, Black women, White women and Hispanic men than for otherwise identical White men.

The audit is a preprint under review. It measures callbacks rather than offers, and the 28 to 43 percent range spans four groups.

The free-text box is the most discretionary square inch of a screen.

Two readers who wrote down behavior can settle a disagreement. Two who wrote down verdicts can only vote, which is how a split between one interviewer calling it strong delegation and another calling it outsourcing the thinking becomes a meeting.

Cutting the guesses does make the note thinner.

A screener who writes only quotations hands over a transcript nobody reads, and the impression that actually moved the call goes unrecorded, which is its own kind of hidden.

So take the sentence back and write the same twenty minutes the other way. Asked how they check what a model gives them. Named three sources, then could not say which one the figure came from. On follow-up, said the model had listed them. Requirement, verifies before shipping: not evidenced.

Nothing in that names a tool, a mood or a motive. In year four, it still says what happened.

Read the evidence >