Forty-one percent of the hiring leaders we surveyed say that one in four or more of their recent hires turned out weaker than the hiring evidence suggested. Eighty-eight percent were involved in at least one such hire in the past year.
These are leaders describing their own decisions, made with the hiring file in front of them: the résumé, references, interviews, and credentials. The file supported the decision. In their judgment, the person’s subsequent performance did not match what that evidence suggested.
That raises a question for hiring leaders: what does the evidence in the file actually establish, and what are we inferring from it?
These figures are self-reported. The study did not independently evaluate employee performance or establish that a particular assessment method caused a weaker hire.
What the Human Evidence Project is
These findings come from The State of Human Evidence 2027, the first study of the Human Evidence Project, a public research initiative founded and funded by SignalVerified (opens in a new tab). The project centers on one question: what evidence do organizations trust when evaluating human capability in AI-enabled work?
SignalVerified is the company I lead, and it has a commercial interest in this subject. That is why the funding relationship is explicit and the project commits to reporting all findings, including results that challenge its funder. The principle is research first, company second. This is the first wave of what we intend to be annual research.
Why now: AI makes polished work easier to produce
AI changes what a polished output can tell us about the person who delivered it. Seventy-two percent of leaders in our study say AI now does around half or more of their teams’ work. This is their assessment of AI’s contribution, not a measured share of tasks or working hours.
When AI may have helped produce a piece of work, only 12 percent of leaders accept the quality of the output alone as proof of what a person can do. The most common answer, chosen by 43 percent, was work the person can explain, defend, and correct when questioned.
The output still matters. It simply no longer testifies on its own. A clean analysis may demonstrate that a useful artifact exists; it does not, by itself, establish who framed the problem, checked the claims, recognized the tradeoffs, or corrected the mistakes.
This is what we mean by human evidence: observable proof of how a person thinks, decides, communicates, and performs when doing real work.
What leaders distrust in the hiring file
We asked leaders to compare twelve kinds of hiring evidence side by side, choosing the most and least convincing across repeated sets. The conversational interview without set scoring was selected as least convincing in 42 percent of its appearances, the highest rate of all twelve. Résumés followed at 34 percent, and degrees at 33 percent.
Those percentages describe choices when each evidence type appeared in a comparison set. They are not the share of all leaders who reject interviews, résumés, or degrees outright, and they do not measure those methods’ predictive validity.
The distinction matters. The study records what leaders say they find convincing. It does not prove that the least trusted evidence is always the least useful. It does, however, give organizations a reason to examine how much weight they place on familiar signals, particularly interviews without consistent scoring.
The gap between what leaders want and what they collect
Trust is only one part of the question. The other is whether hiring processes collect the evidence leaders want. In the early findings, 57 percent want a finalist work task for most roles, while only 6 percent report running one.
A work task is not automatically better evidence. Its value depends on what it reveals about the person’s contribution, how consistently it is assessed, and whether the person can explain their decisions. The gap is a prompt to investigate, not a prescription to add another assessment to every hiring process.
Think of the last hire on your team who fell short of what the file suggested. Which part of that file supported the decision, and what would you want to see before trusting it again?
What this study can and cannot tell us
This study records what leaders say counts as proof, not what predicts performance on the job. It does not establish that AI caused hiring evidence to fail, nor that any single alternative will fix the problem. Testing which signals predict capability is a separate research question.
For now, the findings help make the assumptions visible. Hiring and promotion still require judgments about people, even when machines can produce much of the work. Organizations need to distinguish the quality of an answer from the capability of the human behind it.
Read the early findings
The three-page early findings briefing is available now for hiring leaders. It includes five headline results, three moves for hiring leaders, and the study’s method and limits. Share your name, work email, and organization to download it.
The full report releases in early January 2027 for research partners, alongside a public findings summary. You can get release updates or explore a research partnership.
About the study and author
The State of Human Evidence 2027 surveyed 148 U.S. hiring managers and executives at organizations with 50 or more employees online, October 6 to 9, 2026. Each had made or influenced a hiring or promotion decision for a knowledge-work role in the past 12 months. Recruiter, talent acquisition, and HR roles were screened out.
Unless stated otherwise, percentages refer to the 148 respondents. The evidence-comparison percentages above refer to the occasions when each evidence type appeared. Findings describe this sample and should not be read as representative estimates for all U.S. employers. See the early findings briefing for the study’s method and limits, and the FAQ for funding and release details.
Michael A. Campbell, Ed.D., is the founder and CEO of SignalVerified (opens in a new tab), which founded and funds the Human Evidence Project. This article is adapted from his October 14, 2026 piece for The Capability Signal. It discusses the research, not a product recommendation.