Every Signal You Trusted in Hiring, AI Can Now Fake

Every signal your hiring process trusts is now cheap to fake. The résumé, the portfolio, the take-home, the writing sample, the tidy answer in the screening call. A candidate with a browser tab open can produce any of them in an afternoon, at a quality that clears your bar. This is not a warning about the future. It is a description of the shortlist sitting in your inbox right now.

A signal was never the thing itself. It was a stand-in for the thing, and it worked for one reason only: faking it used to cost about as much as having the real thing. To write a sharp strategy memo you had to be able to think strategically. To ship a clean portfolio you had to be able to do the work. The artifact was expensive to counterfeit, so its presence was evidence. That was the whole trick, and it held for a hundred years.

AI broke the trick in eighteen months. It did not make people more capable. It made the artifacts almost free to produce, while leaving the underlying capability exactly as rare as it always was. The gap between those two things, cheap artifact against rare capability, is the crack your entire process now falls through.

The short version:

• A hiring signal works only when faking it costs roughly what having the real thing costs. AI collapsed the cost of the fake to near zero and left the cost of the real capability untouched. Every artifact-based signal broke at once.
• The CV, the portfolio, the take-home, the writing sample, the cover letter, the coding screen. Each was a proxy for capability. Each is now produceable by someone who does not have the capability. You are no longer measuring the candidate. You are measuring their tool.
• This is not cheating at the edges. It is the median candidate, using AI the way everyone now uses it, submitting work that looks like yours and isn't theirs.
• The one signal AI cannot fake for a candidate is what they do when the model is confidently wrong in front of you. That is judgment, and it is the only thing left that still separates the shortlist. We built the AI Operator Profile to measure exactly that.

---

A signal is a bet on the cost of faking it

Think about why a portfolio ever meant anything. When a designer showed you ten screens, you were not admiring the pixels. You were reading the pixels as proof of a mind that could solve the problem behind them. The proof held because the only way to produce those screens was to be able to solve the problem. Effort was the collateral. The artifact was the receipt.

Take the collateral away and the receipt is worthless. That is the position every hiring signal is now in. Not weakened. Not noisier. The mechanism that made it a signal, that producing it required the capability it stood for, is gone. We wrote about this collapse at the level of the whole system in hiring signals going obsolete. This piece is about walking down the list, signal by signal, and watching each one fail.

Walk the list. Every one of them falls.

The résumé. It was always the weakest signal, a self-reported summary you took on faith. Now the faith is misplaced in a new way. Candidates feed the job description to a model and get back a résumé tuned to your exact keywords, your exact competencies, phrased to clear the filter you screen with. The applicant-tracking system you bought to handle volume is now reading text written by a machine to satisfy a machine. Nobody in that loop is a person.

The cover letter. Dead on arrival. It once told you something about how a person writes and thinks under pressure. It now tells you which model they prefer. There is no version of "read the cover letter for voice" that survives, because the voice belongs to the tool.

The writing sample. This one hurts, because writing was supposed to be the honest signal — thinking made visible. It is now the easiest thing in the world to counterfeit. A candidate who cannot structure an argument can submit a structured argument. The sample proves the model can write. It says nothing about whether the person can.

The take-home assignment. The one HR trusted most, because it looked like real work. It is now the one that lies to you most confidently. You send a realistic task, you get back a polished response, and you have learned precisely nothing about the candidate. You learned that they know how to prompt. The take-home has moved from your best filter to your most expensive theatre, and it deserves its own autopsy, which is why I gave it one.

The coding screen. Same story, faster. The models write competent code to competent specs. A live-share screen with the AI muted tells you a fraction more, right up until you realise the candidate has a second monitor.

Read the pattern. It is not that one or two signals slipped. It is that the entire artifact layer, everything you collect before the person is in a room being unpredictable, reports the same corrupted number. That is what the collapse of talent signals means on the ground.

The signals, and what's left of them

Signal	What it used to prove	How AI fakes it now	What still can't be faked
Résumé	A track record worth summarising	Model tunes it to your exact keywords	Whether the track record happened
Cover letter	How they write under light pressure	Model writes it; voice is the tool's	Nothing — retire it
Writing sample	Thinking made visible	Structured argument from an unstructured mind	Can they defend it live, line by line
Portfolio	A mind that can solve the problem	Output produced without the underlying skill	What they'd change and why
Take-home	Capability on realistic work	Polished answer, zero effort collateral	What they do when the AI's answer is wrong
Coding screen	Can they build it	Competent code to a competent spec	Can they catch the bug the model shipped

This isn't the cheater at the edge. It's the median.

The comfortable way to hear all this is: "a few dishonest candidates will game the process, and we'll catch them." That framing is a decade out of date. The person using AI to build their take-home is not a cheat hiding in the corner of your funnel. It is your median applicant, using the tool the way their whole generation uses it, with no sense that they are doing something wrong. Because they are not. They are using the instrument that sits on every desk.

So you cannot screen your way out of this by hunting for fakes. The fake is the baseline. When everyone submits AI-shaped artifacts, the artifact stops carrying information about the person. You are grading the models against each other. Here is the part that should worry you more than the cheating: the candidate who leaned hardest on the tool, and checked it least, is the one whose artifact looks most finished. Wharton researchers Steven Shaw and Gideon Nave, in their study Thinking, Fast, Slow, and Artificial, ran three experiments with 1,372 people. Accuracy collapsed to 31.5% when the AI was wrong, below the 46% people managed with no AI at all, and they followed the wrong answer around 80% of the time, growing more confident as they got it wrong. The most finished-looking submission in your pile is the one produced with the least thought.

Then what is left to measure?

Not an artifact. A behaviour.

Everything AI can fake has one thing in common: it is a thing the candidate hands you, produced somewhere you cannot see, at a time you do not control. Every signal that survives has the opposite shape. It happens in front of you, in real time, on a problem the candidate cannot pre-cook.

The behaviour that still separates people is what they do the moment the machine is confidently wrong. Hand a candidate a task with an AI answer already attached: clean, plausible, and quietly incorrect. Say nothing about the error. The AI user accepts the gift and builds on it. The AI Operator gets an itch, checks the load-bearing claim, finds the crack, and rebuilds the part that was rotten. Same task, same twenty minutes, opposite hire. That is the case I made in full in stop hiring AI users, start hiring AI Operators.

That behaviour cannot be faked for one simple reason. Faking it would require the exact judgment you are testing for. The candidate would have to know the AI was wrong, know where, and know how to fix it — which is precisely the capability that makes them worth hiring. The signal is unfakeable because passing it is the thing.

The honest correction

Stop treating the artifacts as evidence. The résumé is a routing document, not a verdict. The portfolio is a conversation starter, not a conclusion. None of the pre-collected work tells you what you need to know anymore, and pretending it does is how a confident, unchecked hire ends up on your team looking exactly like your best one — a person AI hasn't just helped, but replaced in the parts that mattered, described more fully in why the whole process broke.

The work moves to where it can't be faked: a live moment, a wrong answer, and a close read of what the person does next. It is more effort than scoring a take-home. It is also the only measurement left that measures the candidate instead of their tool.

Ivanooo built the AI Operator Profile to make that measurement repeatable — to put a candidate in front of a confidently wrong machine and read the one signal AI can't produce on their behalf.

---

Frequently asked questions

Which hiring signals can AI fake? All the artifact-based ones: the résumé, cover letter, writing sample, portfolio, take-home assignment, and standard coding screen. Each was a proxy that worked only because producing it required the underlying capability. AI made producing them cheap while leaving the capability rare, so each now reports the quality of the candidate's tool, not the candidate.

Can't I just detect AI-generated applications and filter them out? No, and chasing detectors is the wrong fight. AI-assisted work is now the median, not the exception — your typical honest candidate uses the tool the way everyone does. When the baseline is AI-shaped, the artifact stops carrying information about the person regardless of whether you flag it. The fix is to change what you measure, not to hunt fakes.

Is the take-home assignment still worth using? Not as evidence of capability. A polished take-home now proves the candidate can prompt, not that they can do the work unaided. If you keep it, treat it only as a conversation starter for a live session where you probe what they'd change and why.

What hiring signal can't AI fake? Judgment under a live wrong answer. Hand the candidate a real task with a plausible-but-incorrect AI output attached and watch whether they catch it. Faking that would require the exact judgment you're testing: knowing the AI is wrong, where, and how to fix it. Passing the test is the capability itself.

Why do the most polished applications sometimes come from the weakest candidates? Because the candidate who leaned hardest on AI and checked it least produces the most finished-looking artifact. Wharton's Shaw and Nave found people followed wrong AI answers around 80% of the time while growing more confident. Polish now correlates with surrender to the tool, not with capability.

How should I redesign hiring around fakeable signals? Demote every pre-collected artifact to routing information, and move the real assessment into live, unpredictable moments the candidate can't pre-cook — especially a wrong-answer scenario. Measure the behaviour, not the deliverable. A structured way to do this is the AI Operator Profile.