Send the Session Where You Corrected the Model

Send the Real Session, Corrections Included

Send the session as it actually happened: the wrong turns, the moment you told the model it was wrong, and the parts you rewrote yourself. The guidance employers are given on reading these logs is built around finding exactly that material. A transcript with no visible correction in it is not a stronger submission; it usually reads as the weaker one, because it gives the reviewer nothing to check your judgment against.

The evidence on this is specific and worth taking seriously. In one peer-reviewed paper, an applicant whose letter disclosed AI help was trusted less than one who said nothing about how it was written ¹. A second experiment in the same paper tested the case that matters here: when a third party exposed a professional's unmentioned AI use, trust fell to the lowest point in the study, well below the level recorded for someone who had said so upfront ². The version that costs you the least is the honest one, sent once, not touched up afterward.

Think about what a reviewer actually does with the file. Nobody is counting your prompts or timing how long you spent on each one. A reviewer scanning a chat log is looking for the handful of moments where you stopped, disagreed with the model, and did something about it: rephrased the question, asked for a source, rewrote a paragraph in your own words, or threw out an answer entirely. Those moments are short, usually a few lines each, and they are exactly what disappears first when someone tries to make a log look tidier than the work actually was.

If the exercise allows a short note alongside the transcript, use it to point at one or two of those moments rather than summarizing the whole session. A reviewer working through a stack of submissions reads a pointer faster than a wall of chat history, and naming your own best moment costs you nothing while making the file easier to credit fairly.

What to Ask Before Anything Records

Before a screen recording starts, ask what's captured, who sees it, why, and how long it's kept. These are ordinary questions a reasonable employer can answer in writing, not objections that put you on the back foot. If the answer is vague or nobody can give you one, that vagueness is itself useful information about how the process is run.

Being asked to record isn't unusual, and it isn't aimed at you specifically. In a survey by Greenhouse, an applicant-tracking vendor, of 373 hiring managers in the UK, Ireland and Germany, 59% said they already use some form of software to monitor or detect AI use during interviews at least some of the time, and most of the rest said they were considering it ⁴. A request to see your session is the same instinct applied to a written exercise instead of a live one.

An accommodation request is a normal ask, and the costs employers actually report for granting one are small. In a Job Accommodation Network survey, most employers who reported what a workplace accommodation cost said it cost nothing to implement, and most of the rest reported a one-time cost with a median of $300 ³. If the recording format itself is a barrier because of a disability or a medical condition, say which of those it is when you ask for an adjustment, rather than describing the format on its own. You never have to name a diagnosis, though an employer can ask for documentation where the need is not obvious. How to ask for an accommodation on an AI interview covers what to write and who to send it to, and disability law differs by country and by state, so check what applies where you are applying.

What employers have to tell you before recording your screen covers the specifics employers are told to disclose: what's captured, who watches it, how long it's kept, and why the microphone often sits on its own separate switch from the video.

A reasonable employer treats these as routine questions, because they are the same questions any employer handling personal data should expect to answer. If a recruiter reacts to the question itself as suspicious, that reaction tells you something about the process worth weighing alongside the offer, separate from whatever the recording ultimately shows.

Don't Manufacture a Cleaner Log

Do not edit a transcript into something you did not do. A rewritten log that removes every wrong turn and every correction fails the same way a fabricated claim on a resume fails: it collapses the moment a follow-up question asks you to walk through a decision that, in the cleaned version, no longer has a visible reason behind it.

What a reviewer is actually looking for in the log is acts, not volume: where your first message aimed, which single claim you stopped and demanded a source for, and what the assistant produced that you threw out. Turn count and fluent phrasing are noise next to that. None of it survives being smoothed away. The same guidance also tells reviewers that a pasted log is a curated log, so the load-bearing parts get confirmed in a conversation about the work, and that conversation is where an edited transcript comes apart.

If you are genuinely unsure whether a section of the session is relevant, leave it in rather than guessing which parts to remove. A reviewer who has the full session can decide what matters; one handed a curated excerpt has to take your editorial judgment on faith, which is a strange thing to ask of someone deciding whether to trust your judgment in the first place.

The same holds for a screen recording, not just a text log. If a segment shows you pausing, backtracking, or restating the task to yourself, leave it in rather than trimming it out before you send the file: it is often the clearest evidence in the whole recording that the thinking was happening in real time rather than being reconstructed afterward to look tidier than it was. A recording that never pauses, never backtracks, and never shows a second thought is the one that earns a reviewer's closer look.

See what gets scored >