Verified, inferred, or proven: why the résumé is breaking in the AI era

The signal everyone still hires on is now noise

For decades, the résumé was a convenient proxy. It was never proof that someone could do the job — it was a claim, written by the candidate, that a recruiter chose to believe. It worked well enough because writing a convincing, tailored résumé took effort, and effort correlated loosely with intent and ability.

Generative AI has severed that correlation. A candidate can now produce a flawless, keyword-perfect, role-tailored résumé in seconds — and rehearse polished answers to every predictable interview question just as fast. The artefacts hiring has always relied on to infer ability are now trivially manufactured. The proxy still looks the same; it just no longer points at anything real.

The industry's response has split into two camps, and it is worth being precise about the difference, because they are not the same thing and neither one is enough on its own.

Inference: predicting skill from data exhaust

One camp infers. Talent-intelligence platforms build enormous graphs from hundreds of millions of career profiles and use them to predict what a person can probably do: you held this title, at this kind of company, adjacent to these skills, so you likely have that capability too. It is genuinely useful for discovery — surfacing candidates a keyword search would miss, spotting adjacent skills, forecasting trajectories.

But inference is a probability, not a fact. It is built from the same résumé-shaped signals that AI has just learned to fabricate, plus association: people like you tend to have this skill. That is a reasonable prior for sourcing. It is a dangerous basis for a decision. Inference cannot tell you whether this person, today, can actually do the thing — only that the population they resemble usually can.

Testing: sampling skill at a moment in time

The other camp tests. Adaptive assessments measure a skill directly by asking questions and scoring answers, often with computerized adaptive testing so the difficulty tracks the candidate. This is a real measurement, and a large step up from inference: it observes the person rather than the population.

But a test is a sample. It captures how someone performed on a set of items, in one sitting, under whatever conditions applied. Two questions decide whether that sample is worth anything: was it the right difficulty to be informative, and was it actually that person, working unaided? A high score from a test that was too easy, or taken with a second browser open, measures nothing. Testing without calibration and without integrity is just a more expensive résumé.

Proof: measurement plus evidence plus confidence plus integrity

Proof is the third thing, and it is what a defensible hiring decision actually requires. A skill proof is not a predicted likelihood and not a raw score. It is a measured level, tied to the specific evidence that earned it, carrying a stated confidence, produced under verified conditions.

The distinction matters most at the moment a decision is questioned — by a rejected candidate, a hiring manager, a regulator, or your own future self reviewing why a hire did not work out. Inference gives you "the model thought they were a good fit." A raw test gives you "they scored 78%." Proof gives you "here is the answer, here is the skill it demonstrates at this level, here is the confidence, and here is the record that it was really them." Only one of those survives scrutiny.

Why this is not a subtle distinction

It is tempting to treat inference, testing and proof as points on a spectrum of the same thing. They are not. They fail differently.

Inference fails silently. It quietly encodes the biases of who held which title at which company, and it presents a guess with the confidence of a fact. Nobody sees the failure until a pattern of decisions turns out to have been built on association rather than capability.

Testing fails loudly but late. The score looks authoritative right up until you learn the assessment was gameable, or miscalibrated, or that the person who scored it was not the person who showed up. By then the offer is out.

Proof is designed to fail safely. Because every score is anchored to evidence and carries a confidence bound, the weak spots are visible before the decision, not after it. Low confidence means "measure more," not "guess anyway."

What proof asks of a hiring process

Hiring on proof is not a single tool you bolt on. It is a set of properties every capability signal has to carry:

• Direct measurement, not association. Observe the person doing the thing — through an adaptive interview or a calibrated assessment — rather than predicting from their history.
• Calibration. Use adaptive difficulty and item-response modelling so the measurement is informative and comparable, with a confidence bound, not a raw percentage from an arbitrary item set.
• Evidence anchoring. Every score points back to the specific answer that earned it, so the claim is inspectable rather than opaque.
• Integrity. Verify it was really that person, working unaided, so the evidence means what it says.
• Human oversight. Make scores reviewable and appealable, so an edge case gets a person, not an unaccountable verdict.

Miss any one of these and you fall back to inference or an uncalibrated test — a claim wearing the costume of proof.