AI Speeds Up Recruiting. It Has Not Made Hiring Any More Trustworthy.

The productivity story about AI in recruiting is largely true. The trust story is not. Recruiters are moving faster through pipelines, surfacing more candidates, and compressing time-to-screen by meaningful margins — but more than 2 in 3 recruiters told TestGorilla they spend an hour or more manually checking a candidate's credentials, even after AI tools have already flagged those candidates as qualified. Speed without confidence is not efficiency. It is just faster anxiety.

That finding, from HR Dive's reporting on TestGorilla's recruiter survey (2025), deserves to stand on its own: AI tools have accelerated candidate sourcing but have not reduced the manual verification burden recruiters carry for credentials and fit, leaving the average recruiter spending 60 or more minutes per candidate on manual checks. The implication is that AI has optimised the wrong half of the funnel. Finding people faster only matters if you can trust what you find.

Why hasn't AI confidence followed AI speed?

The answer is structural, not technological. Recruiting AI was built primarily for pattern matching against historical data — résumés, job descriptions, prior hiring decisions. It is genuinely good at that. What it was not built for is verification: Did this person actually do what they say they did? Did the outcomes they claim belong to them or to a team of twelve? These are judgment problems, not pattern problems. And organisations have not invested in the infrastructure — credential verification networks, reference-check automation, structured assessment pipelines — that would make AI-sourced candidates as trustworthy as they are abundant.

The parallel to software engineering is instructive. McKinsey's conversation with Cognition co-founder and President Russell Kaplan makes the point that autonomous agents in software development create measurable business value only when the surrounding systems — testing, review, deployment — are mature enough to absorb AI output without constant human firefighting. The same logic applies to recruiting: AI sourcing agents produce value only when the downstream verification and assessment infrastructure can handle volume without collapsing into manual spot-checks.

What does responsible AI-assisted hiring actually look like?

A useful frame comes from MIT Sloan Management Review and Boston Consulting Group's (BCG) fifth annual responsible AI study, which found that the defining challenge of responsible AI is designing systems that preserve human judgment when and where it counts. Hiring is precisely one of those high-stakes, high-consequence decisions where human judgment cannot be outsourced to a confidence score. The failure mode is not that AI picks the wrong person — it is that organisations stop asking the questions that would catch the wrong person, because the pipeline feels managed.

There is also a compensation signal worth watching. OpenAI's London employees received an average compensation package of £2,318,181 GBP in 2024, a figure that reflects the extreme wage premium being paid for the engineers who build the AI systems reshaping every other profession. The organisations paying those numbers are not trusting résumés. They are running extended technical trials, structured interviews, and reference networks that function more like due diligence than recruitment. The rest of the market is trying to use AI to compete for talent at the top of the funnel while still operating 1990s verification practices at the bottom.

The verification gap is a governance problem, not a tooling problem

Chief human resources officers (CHROs) who frame this as a technology procurement question are asking the wrong question. The right question is: what is the institutional cost of a mis-hire that AI sourcing accelerated through an under-resourced verification process? That cost — in onboarding spend, team disruption, rehiring cycles, and occasionally legal exposure — does not appear in the time-to-fill metric that AI vendors optimise for. It appears six to twelve months later, in turnover data and manager feedback that gets labelled a "culture fit" problem when it was actually an information problem.

The structural fix is unglamorous: invest in verification infrastructure at the same pace as sourcing infrastructure. That means automated credential checks, structured work-sample assessments, and reference processes that are not optional add-ons for senior hires only. It means treating candidate trust as a data quality problem, not a gut-feel problem.

Speed in the funnel is a feature. Confidence at the offer stage is the product. Right now, most organisations are buying the feature and calling it the product — and their hiring outcomes will eventually reflect that confusion.