A skill you cannot verify is a missing figure.

A plant manager asks which of her electricians hold a current high-voltage certification. The answer is in the learning records, with dates. The same manager asks which of them could lead a commissioning project. The HR system has an answer for that too, inferred from job titles, courses and a self-assessment two years old. Both answers arrive in the same font.

What people say about themselves

A 2024 analysis of more than 22,000 domain assessments compared self-ratings with tested performance. Fifty-six per cent of people underestimated their level, 32% overestimated it, and 11% were accurate. That is one vendor's population and the figures should be read as an order of magnitude, but the direction matches decades of research on self-assessment: people are poor judges of their own skill, in both directions, and the ones who underestimate are often the ones you want.

What a model infers

Inferred skills are better than self-ratings and still far from facts. The TalentCLEF shared tasks are the closest thing the field has to a public exam. In 2025, with 76 teams and more than 280 submissions, the best job-title matching system scored 0.534 mean average precision across languages and the best job-to-skill prediction scored 0.360. In 2026, with 113 teams, the best job-to-person matching reached 0.71 and the best graded job-to-skill matching reached 0.81. The 2026 corpus is open, and so is a 13-task benchmark published in March 2026 under an open licence, which makes the claims testable.

Two cautions. The benchmarks test matching on curated text; a real tenant has twenty years of inconsistent job titles. And a vendor's claim of “over 90% inference accuracy” is a claim until the test set is public. The open benchmarks are the standard to ask for.

Four origins, applied to skills

We label every figure in a brief by where it came from: read from a record, computed by code, quoted from a document, or the Coworker's own reasoning. Skills get the same treatment, and it changes what the plant manager sees.

1. Read: a certification with an issue date and an expiry date in the learning record. A fact, with a source.
2. Computed: “has led two projects over 500,000 euros in the last three years”, from project and position history. A fact derived in code, with the inputs shown.
3. Quoted: the job profile's required skills, from the document your HR team maintains, down to the line.
4. Reasoning: “could lead a commissioning project”, inferred from the three above. Labelled as the Coworker's judgement, never shown as a fact.

A skill that is none of these, because the only evidence is a self-rating or a profile nobody updated, is reported the way a missing figure is: “not verified in what I could read”. The manager can still ask about it. She just cannot mistake it for a record.

What to ask a skills vendor

1. Which public benchmark did you score on, and what was the score?
2. For each skill on a profile, can you show me the evidence it was inferred from?
3. What happens to a skill the person rejects? Does it lower related ones, or just disappear?
4. How old is the average skill on a profile in a tenant like ours?