Almost every AI hiring tool on the market today is trying to do the same thing faster: read resumes, scan cover letters, score video interviews, or make candidates solve abstract puzzles. We went a different direction. Our screener ignores all of that. It just asks candidates questions about the actual job — in plain text, asynchronously — and works from their answers.

That's a deliberate response to what's broken in everything else.

The problem with reading resumes

The most famous cautionary tale in AI recruitment is Amazon's scrapped hiring tool. It was trained on a decade of resumes, and because those resumes came mostly from men, the system taught itself that male-associated language was a positive signal and penalised resumes that signalled women. Amazon eventually abandoned it.

The lesson wasn't "Amazon built it badly." The lesson was deeper: resumes are dense with bias. They carry your name, your school, your employment gaps, your age, your background — a hundred proxies for who you are rather than what you can do. Any model trained to read them inherits all of that. You can try to scrub the bias out afterward, but you're cleaning a contaminated input.

So we removed the input. No resume, no cover letter. The screener never sees where you went to school or what your last job title was. It can't reward prestige or pattern-match on a name, because it never receives that information in the first place.

What hiring research actually tells us

Decades of selection science have studied what actually predicts whether someone will perform well in a job. The findings are consistent and they're not flattering to the resume.

Years of experience and educational credentials — the things resumes are almost entirely made of — turn out to be among the weakest predictors researchers have found. They tell you where someone has been, not what they can do. Structured interviews, where every candidate is asked the same job-relevant questions and evaluated against the same criteria, consistently rank near the top of the hierarchy. The gap between the two is large and has held up across decades of replication.

The takeaway is straightforward: resumes are a weak signal dressed up as a strong one. Structured questions about the actual job are a stronger signal that most hiring processes underuse. We built around the stronger signal and dropped the weaker one.

Why not video or puzzles either?

Video interviews bring their own problems — they reintroduce everything a resume leaks (appearance, accent, perceived age, race) plus a new layer of judging people on how telegenic they are. Facial-analysis hiring tools have been so heavily criticised that several have been pulled or regulated.

Puzzle and game-based assessments have the opposite problem: they're often too abstract. Ranking someone on a logic game tells you they're good at the logic game. Whether that transfers to the actual role is a shakier bet, and these formats favour people who've had the time and resources to practise that specific style of test.

Questions about the job avoid both traps. You're not judging a face or a hobby skill — you're asking someone to engage with the work itself.

There's also an accessibility case that gets less attention than it deserves. Text-based, asynchronous screening genuinely levels the field for candidates who freeze on video, have social anxiety, or simply don't perform well under synchronous pressure. The format gives people time to actually think through their answers rather than react in the moment. That's not a concession to lower standards — it's a recognition that how someone performs under artificial interview pressure often has little to do with how they'll perform in the role.

Two regulatory regimes now govern AI hiring tools, and they matter enormously for how you build.

New York City's Local Law 144 requires that automated employment decision tools undergo an independent bias audit, with audit summaries posted publicly, advance notice in job listings, and an opt-out option for jobseekers. Enforcement was slow to start, but the city has committed to a more rigorous phase. Employers should expect more scrutiny, not less.

The EU AI Act is the larger exposure for global employers. It classifies AI used in recruitment and selection as "high-risk," subjecting it to strict oversight, mandatory disclosure to candidates, and rights to explanation for individual decisions. The penalties are substantial enough that non-compliance is a material business risk, not just a reputational one.

A text-based structured screener that never ingests demographic data and produces rubric-based scores is far easier to bias-audit and defend than a video tool analysing faces and voices or a black-box resume parser. Compliance-friendly by design isn't a selling point we bolted on — it's a consequence of the approach.

The hardest problem: candidates using AI to answer

This is the most honest section we can write, and skipping it would date this article instantly.

Candidates using generative AI to answer screening questions is now a mainstream behaviour, not an edge case. Text-based screeners are the most exposed format — the loop is trivial: copy the question, paste it into ChatGPT, paste the response back. There's also a subtler problem called rubric hacking: AI models are trained on the same documentation hiring managers use to build scoring rubrics, so they reliably generate the textbook-perfect answer that interviewers are conditioned to reward.

There are real answers, but surveillance isn't one of them. Eye-tracking, browser lockdowns, and other monitoring tools destroy candidate trust, are easily bypassed with a second device, and send the message that you approach candidates as suspects. The credible response is deterrence and question design.

Our answer is to ask questions where the reasoning matters more than the polish. A question like "tell me about a specific time you had to make a decision without all the information you needed — what did you decide and what happened?" is far harder to credibly fake than "how would you handle a conflict with a colleague?" The first requires a real story with real details that hold up under follow-up. The second produces the same polished non-answer whether a human or a language model wrote it. We also tell candidates directly that we're looking for their genuine thinking, not a textbook answer — deterrence paired with question quality is more effective than any monitoring tool.

Being honest: what we're still working on

Research on conversational AI assessments has found a consistent tension: formats that reduce bias sometimes show lower predictive validity than the tests they replace. In plain terms, the fairer method isn't always the best predictor of who will succeed. That's a real tension, not a footnote, and it means the questions have to work harder to carry the predictive weight that the resume used to carry — imperfectly, but still.

Text also isn't perfectly neutral. Vocabulary and phrasing still correlate with education and first language, so a system that rewards polished writing can quietly disadvantage non-native speakers. We score for substance and reasoning, not prose style — but it's something any text-based tool has to guard against actively, not assume away.

These aren't reasons to abandon the approach. They're reasons to take question design seriously, which is where we spend most of our product effort.

Why this bet is worth making

An AI screener that asks about the job — instead of reading resumes, watching videos, or running puzzles — isn't a softer version of the others. It's a bet that the most useful, least biased signal you can get from a candidate is how they think about the work itself. The science on structured interviews backs that up. The compliance case backs it up. The accessibility argument backs it up.

The honest challenge is making the questions good enough to carry the predictive weight, and designing them to resist the AI-assisted-answering problem that every text-based tool now faces. Those are the problems worth solving. They're the ones we've chosen.

Frequently asked questions

Does this work for technical roles?

Yes, with the right questions. We don't replace a coding test for roles where writing code is the primary output — but for the majority of technical roles, the bottleneck isn't identifying who can write a binary search tree. It's identifying who can communicate technical decisions clearly, who has good instincts about tradeoffs, and who will work well with the rest of the team. Structured questions about past technical decisions and situational judgment capture those things better than abstract puzzles do.

What about volume hiring — doesn't quality suffer at scale?

The opposite, actually. The consistency argument for structured screening gets stronger at volume, not weaker. When a recruiter reads hundreds of resumes manually, the last one gets a fraction of the attention the first did. When every candidate answers the same questions evaluated against the same rubric, everyone gets the same scrutiny regardless of where they fall in the queue. The problem with volume hiring has never been that structured screening doesn't scale — it's that manual structured screening is too slow. That's exactly what we've automated.

How do you stop candidates just using ChatGPT?

You can't guarantee it, and anyone who claims otherwise is selling something. Our answer is question design: ask for specific personal examples with enough detail that a generic AI answer is obviously hollow, use follow-ups that probe the reasoning behind an answer, and be transparent with candidates that we're evaluating thinking, not polish. Deterrence and question quality together are more effective than any monitoring tool.

Is this compliant with hiring laws like the EU AI Act?

We're designed to make compliance easier, not to make legal promises we can't keep. Because our screener never ingests demographic data — no name, no school, no location — there's no demographic signal to audit out after the fact. It was never in the input. Our scoring is rubric-based and explainable. We provide a Data Processing Agreement and full subprocessor list. For specific legal questions, your employment counsel will know your jurisdiction's requirements better than we do — but we can tell you exactly what data we process and how.

What makes a good screening question?

Three things: it should be specific to the role, it should ask for a concrete example or a situational judgment rather than a generic opinion, and it should have a natural follow-up that tests depth. "Tell me about a time you disagreed with a decision your manager made and how you handled it" is far stronger than "how do you handle conflict?" The first requires a real story. The second produces an answer anyone could write in thirty seconds. The questions are the product — which is why we spend more time on them than on anything else.