Skills-Based Hiring Changed Under 1 in 700 Hires: The Tests That Still Predict

Dropping degree requirements was supposed to open millions of jobs to people without a bachelor's degree. When the Burning Glass Institute and Harvard Business School tracked what actually happened, the change reached fewer than 1 in 700 hires in 2023. Skills-based hiring mostly rewrote job postings, not decisions. And the skills assessment meant to replace the degree is now something a model can sit for the candidate.

The short answer. Skills-based hiring has mostly failed at the decision stage, not the posting stage. Removing a degree line changed fewer than 1 in 700 hires. The assessments that still predict performance when candidates use AI are structured interviews, proctored job knowledge tests and work samples done live, with forced-choice formats for any personality measure.

Skills-based hiring announcements outran the hiring data

Between 2014 and 2023, the annual number of roles where employers dropped degree requirements grew nearly fourfold, according to the Burning Glass Institute. Large employers announced it publicly. Governors removed degree lines from state jobs. The story became that the paper ceiling was coming down.

The February 2024 report from the Burning Glass Institute and the Harvard Business School Project on Managing the Future of Work asked a narrower question. After a company removed the requirement, did it actually hire more people without degrees into that role?

What the Burning Glass and Harvard study measured

The researchers looked at 11,300 roles at large firms. Each role was one occupation at one employer, with enough hiring volume to compare at least a year before and after the requirement came off. Inside those roles, the share of hires without a bachelor's degree rose by about 3.5 percentage points on average.

That sounds modest but real. The problem is the denominator. Only 3.6% of roles dropped a requirement during the period, so the net effect across all hiring was about 0.14 percentage points. HR Dive's reading of the report puts that at roughly 97,000 workers out of 77 million yearly hires.

Three kinds of employer

The averages hide a split. The report sorted firms that removed degree requirements into three groups, and the groups behave so differently that treating them as one trend is misleading.

About 37% were what the authors call Skills-Based Hiring Leaders. They increased the share of non-degree hires in the analysed roles by nearly 20%, and nearly all of the real change came from them. About 45% made the change in name only, with no meaningful difference in actual hiring behaviour. The remaining 18% were backsliders, who made short-term gains and then drifted back.

The workers who did get through did well. Non-degree hires into roles that previously required a degree saw a 25% salary increase on average. At Leader firms, their retention rate ran 10 percentage points higher than that of degree-holding colleagues.

That last finding is the one I would put in front of any finance team. It means skills-based hiring is not charity. Where it is done properly, it produces employees who stay longer, which is a cost line, not a values statement.

Why skills-based hiring stalled at the screening stage

The report's own explanation is blunt. Companies reverted to proxies rather than doing the harder work of assessing an individual candidate's skills. Removing a requirement from a posting is a copy edit. Changing what a recruiter does with a stack of 400 applications is an operating change.

Think about where the degree actually did its work. It was never mainly in the posting. It was in the 30 seconds a screener spends on each resume, in the hiring manager's instinct about who will ramp quickly, and in the applicant tracking system's knockout questions.

The degree filter moved downstream

TestGorilla's own survey shows why. 67% of employers still use resumes in their process, and 71% spend less than 15 minutes reviewing one. TestGorilla itself notes that recruiters can easily see who does and does not have a degree from the resume.

So the degree did not leave the process. It moved from an explicit requirement to an implicit one, applied by a person in a hurry. A filter applied silently is harder to audit than a filter written down, which is why I would argue the in-name-only group is worse off than if it had never announced anything.

The fix is sequencing, not sentiment. If the skills assessment comes after the resume screen, the resume screen still decides who gets assessed, and the resume screen still reads degrees. TestGorilla's data suggests that only about a third of employers using tests run them before looking at resumes. That ordering is the whole game.

What employers say about skills-based hiring versus what they do

Survey data on skills-based hiring is abundant and almost all of it comes from companies selling assessments. That does not make it wrong. It does mean the numbers measure stated intent, and the Burning Glass work shows how far stated intent can sit from behaviour.

Put the two datasets side by side and the gap is visible without any interpretation.

The two sources measure different populations and periods, so the rows are not a like-for-like contradiction. The direction of the gap is the finding, not its exact size.

The third row is the one that should worry anyone running a skills-based programme. In the same survey where a majority claim the degree is gone, more employers say it matters more than say it matters less. Both can be true only if the degree has stopped being written down and kept being used.

My view is that most organisations would get further by dropping the label and keeping the measurement. Nobody needs to announce skills-based hiring. They need to show, quarter by quarter, the share of hires without degrees in roles where the requirement was removed. If that number does not move, the programme is an announcement.

Which pre-employment tests actually predict job performance

Replacing a degree requires something that predicts performance better than the degree did. The best available evidence on that question is the meta-analytic literature on personnel selection, and it was substantially revised in 2022.

For 24 years, practitioners quoted Schmidt and Hunter's 1998 estimates. Paul Sackett, Charlene Zhang, Christopher Berry and Filip Lievens showed that many of those estimates had been inflated by overcorrecting for range restriction. Their revised figures, published in the Journal of Applied Psychology, reorder the table.

Structured interviews came out on top at 0.42. Job knowledge tests sat at 0.40, empirically keyed biodata at 0.38, and work sample tests at 0.33. General cognitive ability fell from 0.51 to 0.31. Unstructured interviews fell to 0.19, the same as conscientiousness. Years of job experience fell to 0.07.

Education was never the strong signal

Years of education was one of the predictors the 2022 team could not re-estimate, because the older studies lacked the information needed to redo the correction. So there is no modern figure to quote, and I will not invent one.

What the revision does make clear is that the methods closest to the job win. Structured interviews ask every candidate the same job-related questions and score them against anchors. Job knowledge tests check what someone actually knows about the work. Both measure the role directly instead of inferring it from a credential. That is the actual case for skills-based hiring, and it is stronger than the equity case most programmes lead with.

Note also that the spread is wide. The structured interview estimate carries an 80% credibility interval of 0.18 to 0.66. A badly built structured interview can predict worse than a good cognitive test. The ranking tells you where to invest effort, not that the label guarantees the result.

How AI cheating attacks each skills assessment format

Here is the uncomfortable timing. Employers adopted skills assessments as the replacement for degrees at the same moment candidates gained a tool that can complete many of those assessments for them. The question for a recruiter is not whether AI helps candidates. It is which formats lose their signal when it does.

Personality questionnaires are the softest target

Jane Phillips and Chet Robie at Wilfrid Laurier University asked four language models to fake ideal responses on personality measures for a target job. They compared the results with students instructed to do the same. GPT-4 scored at the 85th percentile on extraversion and the 98th on conscientiousness relative to the faking students.

The same research found forced-choice formats, where candidates choose between equally desirable statements, harder to fake for both people and models. Rating-scale questionnaires were described as fairly easy to fake. If your process uses a transparent Likert-style personality screen as a gate, a model now produces a near-ideal profile in seconds.

Unproctored skills tests show the score drift

Coding and technical tests are the most measured case, because the vendors running them can see the behaviour. CodeSignal reports that 35% of proctored assessments were flagged in 2025, more than double the year before, with a 40% fraud attempt rate on entry-level roles.

The more useful CodeSignal number is the comparison. Organisations without proctoring saw average score increases of 7.76%, against 1.78% for those using it. That gap is the signal leaking out of unproctored tests. For a deeper look at how engineering teams have rebuilt the interview itself, see how technical interviews changed once everyone ships with AI.

The archival data says the panic is ahead of the evidence

Now the counter-case, which most coverage skips. Alise Dabdoub and Rebecca Pool analysed large archival samples of real job applicants on two cognitive assessments and one personality assessment, before and after ChatGPT's release in November 2022. They found statistically significant but practically negligible changes, with effect sizes near zero and inconsistent in direction.

Assessio, a psychometric test publisher, reports a similar pattern from its own studies. Fewer than 1 in 10 of 500-plus candidates used a tool that was not permitted, only 2.1% would consider using AI in a real assessment, and AI scored 44% on figural reasoning items. TestGorilla found 17% of job seekers admit cheating on skills tests, with 7 in 10 of those using AI.

These numbers range from 2% to 35% depending on who measured what. My reading is that both are true at once. Cheating is concentrated in specific formats, mostly unproctored coding and transparent questionnaires, and it has not yet moved aggregate scores on well-designed cognitive tests. That is an argument for targeted fixes, not for abandoning assessment.

A skills assessment stack that survives AI

The practical answer is not to find an AI-proof test. None exists that is also realistic. The answer is to take the formats with high validity and change the conditions under which they are taken, so that the candidate's own skill is what gets scored.

This table is the stack I would build for a mid-volume role such as an operations analyst, account manager or support lead. It assumes candidates will use AI wherever they are not watched.

Three design rules sit underneath that table, and they matter more than any single format.

Assess before you screen resumes. If a degree is not required, do not let the resume decide who gets assessed. Run a short, proctored job knowledge test or a structured phone screen first, then read resumes for the people who pass. This is the one change that separates the Leaders from the in-name-only group in practice.

Make the work live and short. A 25-minute live work sample with a mid-task change measures more than a four-hour take-home, and it cannot be outsourced to a model without the interviewer seeing it. The format matters more than the length.

Decide whether AI is part of the job. For many roles, using AI well is the skill. In that case, allow it and score the judgement. Our guide to interview questions that test AI fluency covers that version. What you cannot do is ban AI in an unproctored test and treat the result as unassisted.

Finally, measure the programme the way Burning Glass measured it. Track the share of hires without degrees in each role where the requirement came off, and their 12-month retention. If neither moves within two quarters, your stack has a leak, and it is almost always the resume screen.

Where this argument is weakest

Three objections deserve a straight answer, and the first one partly lands.

The hiring data is old. The Burning Glass analysis runs to 2023 and covers large firms. TestGorilla's 2025 survey shows the share of employers claiming to have removed degree requirements jumping from 30% to 53% in a year. If even a fraction of that is real, 2024 and 2025 hiring could look better than the 1 in 700 figure. I have not found newer measured hiring data to settle it, and I would rather say that than assume.

The validity figures are averages across decades of studies. Most were run before AI existed, on criteria such as supervisor ratings that carry their own biases. A 0.42 for structured interviews says nothing about your interview. It says the format can predict well when built properly. The wide credibility interval is the honest reading.

The cheating data conflicts, and most of it comes from vendors. CodeSignal, TestGorilla and Assessio all sell assessments and benefit from conclusions about how to run them. The flag rate of 35% measures suspicion, not confirmed cheating. The one peer-reviewed archival study points the other way. Treat every cheating percentage in this post as directional.

The objection that does not land is that skills-based hiring has failed and degrees should come back. Nothing here supports that. Non-degree hires at Leader firms stayed longer and earned more. The method works where it is actually used. What failed was the substitution of a press release for an operating change.

For a wider view of how much of the work itself is shifting toward things that are hard to test at all, see the piece on the premium on tacit knowledge and human skills.

Frequently asked questions

Where to start

Pull one number this week: for every role where you removed a degree requirement, the share of hires in the last 12 months who do not hold one. Compare it with the year before the change. If it has not moved, you are in the 45%, and you now know it from your own data rather than from a survey.

Then move one assessment in front of the resume screen for a single high-volume role, proctored or live, and run it for one hiring cycle. The question is not whether the candidates are better. It is whether different candidates reach the interview at all.

Related reading