AI Hiring Runs on Trust. Risk Is What Keeps It Honest.

← Back to Blog

Walk into any room of hiring leaders this fall and you will hear the same debate dressed in different words. Is AI making hiring faster and fairer, or is it quietly making it colder and riskier? At the hiring industry's big conference this week, trust researcher Rachel Botsman framed the question in a way that cuts through the noise: stop arguing about what the technology can do, and start asking what it takes for people to actually trust it — and who carries the risk when it is wrong.

Those are two different lenses. The trust lens asks whether the people on the receiving end of a decision believe the process is fair, explainable, and worthy of their effort. The risk lens asks what can go wrong — legally, financially, for your brand — and who is on the hook. Employers mostly operate through the risk lens. Candidates, who have far less power in the exchange, experience hiring through the trust lens. When the two sides are looking through different lenses at the same algorithm, they reach opposite conclusions about the same tool. That gap, not the technology itself, is the real story of AI in talent acquisition.

Trust Is a Confident Relationship With the Unknown

Botsman has spent two decades studying how trust forms between people who do not know each other, and how it breaks when we hand decisions to systems we cannot see. She wrote the early book on collaborative consumption, What's Mine Is Yours, when platforms like Airbnb were proving that strangers would rent rooms to strangers if a reputation system made the leap feel safe. Her follow-up, Who Can You Trust?, took a harder look at what happens when the systems we delegate to — ratings, algorithms, recommendation engines — become the very things we cannot verify.

Her working definition of trust is deliberately compact, and it maps perfectly onto hiring. Trust is what lets you move forward when you cannot calculate the odds.

Note what that definition does. Trust is not the same as confidence, which you earn through repeated evidence, and it is not the same as risk, which is about probabilities you can price. Trust lives in the space between the two: you cannot prove the outcome, so you take a leap based on signals — reputation, transparency, consistency, a sense that the other side has your interests in mind. Botsman's research showed that platforms succeeded when they moved trust from institutions to distributed networks of reviews and reputations. The same mechanism now applies to hiring, except the network doing the trusting is being asked to trust an algorithm instead of a person.

Here is the uncomfortable part. Hiring was already full of unknowns: whether a candidate will perform, whether a company is honest about the role, whether either side is showing its real self. AI adds a new unknown on top of both. For the candidate, the unknown is the black box that screens their resume, scores their interview, or decides they were not a fit — with no visible reasoning and no appeal. For the employer, the unknown is a vendor model trained on data the company never sees, making calls that shape the workforce. Every AI hiring tool is therefore a trust question wearing a technology costume.

Trust is a confident relationship with the unknown. — Rachel Botsman

Two Sides of the Same Algorithm: What Each Lens Sees

Start with the employer's view, and you will see mostly risk math. Hiring is expensive, so anything that promises to process more applicants for less money gets a hearing. Industry surveys reflect that pressure: a widely cited Resume Builder survey of hiring managers found 44 percent of companies using AI somewhere in hiring in 2024, and HR associations put routine use at roughly one in four organizations, with adoption climbing every quarter since generative AI put the tools in everyone's hands. When a hiring leader evaluates an AI screener, the questions are practical: Does it cut time-to-hire? Does it flag the right people? What does the contract say about liability if it misfires? That is a risk lens, and it is a reasonable one.

Now look at the candidate's view, and the questions change completely. A 2023 Pew Research Center survey found roughly two-thirds of Americans would not want to apply for a job where AI made the hiring decision. Job seekers who meet automated screeners report feeling judged by criteria they cannot see, for a job they cannot get feedback on, by a system with no one to appeal to. One survey found 45 percent of job seekers had already used AI to write their application materials — in other words, they are adapting to a system they do not trust by gaming it with the same tools. When candidates say AI hiring feels unfair, they are not making a technical complaint. They are making a trust statement.

Here is the asymmetry that matters. The party with power in the exchange — the employer — evaluates the tool through risk: cost, speed, compliance. The party without power — the candidate — experiences it through trust: fairness, transparency, dignity, recourse. A tool can pass the employer's risk test and fail the candidate's trust test on the exact same output. A recruiter sees an efficient knockout question; the candidate sees an arbitrary wall with no explanation. Both are looking at the same screen.

That mismatch produces perverse behavior on both sides. Candidates who cannot see the rules automate their applications to maximize volume and keyword coverage, which floods employers with noise, which pushes employers toward stricter automated filtering, which makes candidates trust the process even less. Talent analysts at the Indeed Hiring Lab have documented applications per opening climbing sharply since 2023, and hiring managers report that a growing share of that volume is AI-generated. The arms race is not a technology problem. It is a trust failure compounding itself.

The candidate-side trust gap: in a 2023 Pew Research Center survey, roughly two-thirds of Americans said they would not want to apply for a job where AI made the hiring decision.

One Visible Mistake Costs More Than a Thousand Silent Wins

The psychology research on people and algorithms is remarkably consistent, and it explains the hiring mess better than any product review. In a famous line of experiments, behavioral researcher Berkeley Dietvorst found that people abandoned algorithms after watching them make a single visible mistake — even when the algorithm was objectively more accurate than the human alternative. Humans got the benefit of the doubt, mistake after mistake. The algorithm got one strike. The researchers called it algorithm aversion, and it shows up exactly where you would expect: on the side of the person being evaluated.

The flip side is just as important. Researchers led by Jennifer Logg documented what they called algorithm appreciation: when people are the ones using an algorithm to judge someone else, they tend to trust it more than their own judgment — sometimes even when they have no evidence it works. Put the two findings together and you have a grim symmetry. The evaluator over-relies on the tool; the evaluated over-punishes it. An employer's hiring team sees an AI screener as a disciplined colleague. A candidate sees the same tool as an unaccountable judge. Both biases run on the same fuel: neither side can see how the decision was made.

This is why a single horror story — the qualified candidate auto-rejected, the talented applicant filtered out by a keyword quirk — does more damage than a thousand quiet successes. It is not just that the story spreads on social media. It is that the story confirms what the trust lens already suspected: the process is opaque, the stakes are high, and there is no one to argue with. Writers like Cathy O'Neil in Weapons of Math Destruction and Ifeoma Ajunwa in The Quantified Worker have spent years documenting what happens when scoring models quietly decide who gets opportunities, and the recurring theme is not malice. It is opacity plus stakes plus no recourse. That combination is the definition of an untrustworthy system, regardless of how accurate the model is.

For employers this creates a trap. Every dollar saved by automating rejection is spent twice in lost candidate trust — through applications abandoned, offers declined, and glassdoor reviews that read like warnings. But the answer is not to abandon AI, because the volume problem is real and getting worse. The answer is to design the system so the trust lens and the risk lens point the same direction: humans make the final call, the reasoning is visible enough to explain, and mistakes have a recovery path instead of a dead end.

Illustrative pattern from algorithm-aversion research: one visible error collapses reliance on an AI screener, while a human screener keeps getting second chances. The shape matters more than the numbers.

The Risk Lens: What Employers Are Actually On the Hook For

Let the trust lens set the tone, but do not let it hide the risk lens, because the risk is real and it is growing. Regulators have spent the last few years making one point unmistakably clear: when an algorithm is used in hiring, the employer is the responsible party — not the software vendor, not the model, not the data scientist. The US Equal Employment Opportunity Commission issued guidance in 2022 and 2023 stating that AI screening tools are selection procedures under Title VII and the Americans with Disabilities Act, that adverse impact analysis applies to them just as it does to human screens, and that employers can be liable for bias in tools they buy from someone else. The EEOC has since filed suit over AI hiring practices, including a case against a major screening vendor, and a separate age-discrimination settlement over an AI that automatically rejected older applicants.

The state and local picture is filling in fast. New York City requires an independent bias audit before an automated employment decision tool can be used, plus disclosure to candidates. Illinois has required notice and consent when AI analyzes video interviews. Maryland now requires employers to audit the vendors behind their hiring tools. The European Union's AI Act classifies employment AI as high risk, with obligations phasing in over the next two years. The pattern is identical everywhere: audit the tool, tell people it is being used, keep a human accountable. None of these rules bans AI in hiring. All of them assume the employer will behave like a fiduciary of fairness rather than a passenger on a vendor's algorithm.

Compliance is only part of the risk picture. There is brand risk: candidates talk, and an automated rejection experienced as arbitrary is a story that outlives the role. There is quality risk: an over-tuned screener quietly eliminates your best future hires, and you never know they existed because you never interviewed them. There is operational risk: AI-written applications flood the funnel, and a screening model trained on yesterday's data chokes on today's volume. And there is data risk: assessments, video interviews, and personality scoring generate sensitive personal data that becomes a liability the moment it is stored or shared.

A mature risk lens does not ask whether AI is dangerous. It asks where the exposure concentrates, so you can put controls exactly there. The exposure concentrates in three places: automated rejections, where a wrong call has no human to catch it; adverse impact, where a pattern of wrong calls harms a protected group; and vendor opacity, where you are accountable for a system you cannot inspect. Name those three, and the design principles for the rest of the article basically write themselves.

• Audit your outcomes, not your vendor's promises. Track pass rates and rejection rates by demographic group quarterly, the way you would track any other funnel metric. Most bias is visible in the data long before it appears in a lawsuit.
• Demand audit logs and explainability from every hiring tool you buy. If the vendor cannot show you how a decision was made, you are holding the liability and they are holding the convenience.
• Automate the sorting, not the verdict. Let AI rank, summarize, and prioritize. Reserve rejection — the decision with the real consequence — for human review.
• Disclose plainly. Tell candidates what the tool does, what data it uses, and who to contact for a human review. Disclosure costs nothing and removes the single biggest driver of perceived unfairness.
• Keep an appeal path that actually works. A candidate who can ask a human to look again is a candidate who trusts the process, and the occasional reversal is cheap insurance against a systemic miss.
• Treat your own AI-assisted applicants as signals, not spam. Volume is noise; a skills conversation with a real human is still the best verification tool ever built.

Designing for Both Lenses: Trustworthy by Default

Leadership scholars call it adaptive leadership: the discipline of updating your operating assumptions when the environment changes, instead of applying the old playbook more aggressively. AI in hiring is an adaptive problem, not a technical one. The old playbook assumed a symmetrical relationship — an applicant who wants the job writes their own resume, a recruiter who reads it makes the call, and both sides roughly understand the rules. Every one of those assumptions is now optional. Candidates can generate applications in minutes. Recruiters can screen thousands with a model. Neither side can see what the other is doing, and both are behaving accordingly.

The companies that win the talent race in this environment will be the ones that treat trust as a design requirement on both lenses, not a soft value. Start with the division of labor: AI assists, humans decide. Let the model do the exhaustive work — parsing every application, ranking against the job profile, drafting summaries, flagging gaps — and put a human on every consequential verdict. Candidates do not mind being screened by software nearly as much as they mind being dismissed by it. The difference between those two experiences is whether a person with a name and an email address sits at the end of the line.

Second, make the process explainable at the level of the candidate's experience, not the model's internals. Nobody needs your feature weights. What candidates need is a plain answer to four questions: What did you look at, what did you conclude, can I see it, and who do I talk to if it is wrong? Write those answers down before you deploy the tool, in the candidate's language, and you will have satisfied most of what the trust lens demands and a surprising amount of what regulators now require.

Third, be conservative with the knockout. When a tool is allowed to reject without human review, tune it to only auto-reject the clearly unqualified — the missing license, the conflicting availability, the application that fails a basic requirement stated in the job ad. Everything in the gray zone goes to a human. This single rule costs you almost nothing in efficiency and prevents the exact failure that erodes trust fastest: the qualified candidate who never had a chance.

Fourth, measure the relationship, not just the pipeline. Alongside time-to-hire and cost-per-hire, track the trust indicators: what share of screened-out candidates used your appeal path, what your offer acceptance rate is doing, what rejected candidates say when you ask them why they would not reapply. If application abandonment is rising and referral volume is falling, your process is telling you something no model accuracy metric ever will.

Botsman's deeper point is that trust is not a feature you bolt on after the model is built. It is the operating system. A hiring system that is transparent, accountable, and human at the point of judgment does not just feel better — it performs better, because it keeps the candidates you actually want in the conversation and keeps the regulators and the jury of public opinion out of it. Look at your AI hiring stack through both lenses, fix what each one reveals, and you will have done the adaptive work that most of your competitors are still avoiding.