What skills-based hiring actually assumes
Skills-based hiring is the practice of selecting candidates on demonstrated capability rather than on proxies such as degrees, job titles or years served. In practice it usually means one of four things: a structured work sample, a timed assessment, a portfolio review, or a structured interview built around real tasks.
Every one of those methods depends on an assumption that used to be safe. The assumption is that the artifact and the ability travel together. If someone hands you a clean analysis, a working script, a persuasive memo or a well-argued case, you infer the capability that produced it. The artifact is the evidence, and the inference from artifact to ability is the whole mechanism.
That inference is what has come apart. Generative tools can produce a competent artifact for a person who could not have produced it alone, and, just as importantly, they can produce a competent artifact for a person who could. From the outside, the two look the same.
Why "just detect it" does not resolve the problem
The first instinct is to detect assistance and discount it. Two things make that a dead end for hiring teams.
The first is practical. Detection of machine-written text is contested and improves and degrades on both sides continuously. Building a selection process on a signal that unstable is a poor bet, and getting it wrong means rejecting a real person for a machine judgement you cannot explain to them.
The second is conceptual, and it matters more. In most roles, using these tools well is now part of the job. A process that penalises assistance is selecting against the behaviour the role requires. The question is not whether the person used a tool. It is what the person contributed when they did.
What is left to judge
When production becomes cheap, the scarce human contributions are the ones that sit around production rather than inside it. In our reading of how organizations actually make these calls, the candidates for evidence look something like this:
- Framing. Deciding what problem is worth solving, and what the deliverable should even be.
- Selection. Choosing between several plausible outputs and being able to say why one was kept.
- Verification. Knowing which claims need checking, and checking them.
- Correction. Recognising that an output is confidently wrong, and knowing what to do next.
- Accountability. Standing behind the result in front of a client, a regulator or a colleague.
Each of these is observable in principle. Whether organizations trust them, weigh them, or can even see them in a hiring process is an empirical question, and it is the one we are studying.
Why we are doing this study
Most discussion of AI in work asks what the systems can do. Far less is known about the people and institutions on the other side of the output: the managers, reviewers and hiring teams who must still judge capability, and who are doing it without an evidence base for how to do it.
That gap matters because the decisions are high-stakes and the practices are changing fast. Hiring and promotion are the two places where organizations already collect structured evidence about people, and where the shift to AI-enabled work is most likely to make existing evidence unreliable. We do not yet know what organizations will trust instead, or whether the new signals will be fairer, noisier, or simply different.
The State of Human Evidence 2027 is designed to find out. It asks one core question: what evidence do organizations trust when evaluating human capability in AI-enabled work? The survey, hypotheses, analysis plan and methodology are published before a single response is collected. The anonymized data, the limitations and all findings are published after, including results that challenge SignalVerified, which founded and funds the initiative. The principle is research first, company second.
Practical adjustments while the evidence is thin
We are not in a position to tell you what works, because nobody has established that yet. What we can do is set out the adjustments that follow logically from the problem, so they can be argued with.
- Assume assistance and design for it. Tell candidates the tools are permitted and ask them to describe how they used them. A hidden variable becomes an observable one.
- Score the reasoning, not only the deliverable. Ask what was rejected and why. A person who cannot defend the choices did not make them.
- Introduce a live element. A short conversation about the submitted work tests ownership without requiring detection.
- Give the same task to every candidate. Structure was always the strongest part of skills-based hiring. It matters more now, not less, because it is what makes two answers comparable.
- Write down what you counted as evidence. Most teams cannot reconstruct this afterwards, which is precisely why the practice is hard to improve.
The open question
What evidence do organizations trust when evaluating human capability in AI-enabled work?
That is the single question behind the State of Human Evidence 2027, focused on the two decisions where the stakes are highest and the practice is most established: hiring and promotion. The survey, hypotheses, analysis plan and methodology are published before fielding; the anonymized data, the limitations and all findings are published after, including results that challenge SignalVerified, which founded and funds the initiative.
If you make or influence these decisions, you can read how the study is designed and funded, or ask to be notified when it opens.