Both sides of the job interview are running models now, and neither one says so

A recruiter sends you a link. You record four answers into a webcam with nobody on the other end. A model scores the transcript. A human skims the top of the pile.

Meanwhile, on your side of the screen, there is a second window open with the job description pasted into it.

Neither party mentions any of this. Both parties suspect the other.

The part that gets called cheating depends entirely on who went first

The usual framing is that candidates started it. It is worth checking the order.

Automated screening was in the pipeline long before consumer language models were. Keyword filters, then scoring, then recorded-interview scoring. By the time a candidate could open a chat window and ask for help with a cover letter, the other end of the process had already been machine-read for years.

That does not make anything permissible. It does mean the moral panic is pointed in a strange direction.

And it is not a wash, because one side had to consent

Here is the asymmetry that actually matters, and it is structural rather than a matter of opinion.

If a candidate uses a model to prepare, that is their own time and their own business. If an employer uses one to score, it is deciding an outcome about a person, often without telling them, and the person cannot see the rubric, cannot appeal it, and frequently cannot even confirm it happened.

Same technology. Very different position in the power arrangement.

The measured cost is not accuracy, it is who stays in the room

This is where it stops being a vibes argument.

In CESifo Working Paper 12573, researchers randomised over 3,000 real job applicants into asynchronous audio interviews, asynchronous video interviews, live online interviews, or a control group.

Asynchronous interviews caused an over 50% decrease in application continuation. The paper reports the decline happened "including among the most qualified applicants", and that it was largest for women.

Read that as a systems person rather than as a candidate. A stage that removes more than half your input, non-uniformly, and correlated with something other than the quality you are selecting for, is not a filter. It is a sampling bug. Anywhere else in engineering you would treat that as a defect in the pipeline rather than a property of the population.

So what does a candidate actually do about a system they cannot see

Not much, honestly. But not nothing.

Ask what is recorded and what is analysed by software. In the EU the AI Act treats candidate-evaluation systems as high risk. Illinois requires notice and consent before AI analyses a video interview. New York City requires a published bias audit for automated hiring tools. Elsewhere there is usually no rule, and you should ask anyway, because the shape of the answer is information about the company.

Then write for the reader you actually have. A transcript scorer reads structure, not charm. Name your structure out loud. "There were three constraints. The first was cost." That sounds faintly robotic to a person and it is exactly what survives being turned into text and graded.

And answer the follow-up nobody is there to ask. On a recorder, the second attempt does not exist. Whatever you said is the artefact.

The uncomfortable ending

Everyone reaches for the same closing line here, which is that we should go back to talking to each other. Nobody is going to do that. Interviewing a thousand people by hand is genuinely expensive, and that was a real problem, and the recorder was a real answer to it.

But if the cheapest way to interview a thousand people also removes the best of them before anyone is assessed, the saving is being booked on the wrong side of the ledger.

I build AI Applyd, which fills in and sends job applications on the employer's own careers page across fifteen of the hiring systems companies actually use, so the hours before any of this go somewhere better.

One question I do not have a good answer to, and I would take yours: if you were designing the first screening stage from scratch, knowing it has to scale to a thousand applicants and knowing the recorder costs you your strongest candidates, what would you build instead?