Problems with AI hiring: the real risks, and the ones you can avoid

Here is the list, each item with the design decision that causes it. Note what is not on it: the AI being wrong about somebody.

Human first rounds are wrong constantly too. What differs is scale and silence.

• A rejection nobody read. The biggest risk is not a bad score, it is a score wired to an action. Once software can mark somebody rejected, nobody can explain the decision, because nobody made it.
• A number you cannot check. A ranking with no quote, transcript or reasoning behind it cannot be audited, defended to a candidate, or overruled with confidence.
• Bias learned from past hiring. A model built on text people wrote, or tuned on decisions people made, learns what those decisions rewarded, including the parts nobody would defend.
• The wrong thing measured. Fluency is not competence, a keyword is not a skill, and fixed questions with a published rubric measure preparation.
• Applicants who walk away. Some people will not talk to a machine, and no vendor can tell you what share of yours will.
• Records you now hold. Transcripts and recordings of people you did not hire are personal data with a retention question attached.

Yes, and so does every other kind of screening, so the useful question is which step loses the most.

Work through your own funnel. A resume keyword filter rejects on wording: the applicant never said anything, so nothing they could have said would have changed it. An unread pile is worse, because then the order applications arrived in decides.

A fifteen-second skim judges a document written to be skimmed. In all three the candidate is never asked a question. That arithmetic run by hand is in how to screen 200+ applicants without a recruiter.

Now compare an eight-minute voice interview that every applicant gets. It is not better because it is a machine. It is a wider gate because it is a conversation everybody is admitted to: somebody who writes a weak resume and interviews well becomes visible for the first time, and a follow-up about one specific week is a test no polished document passes.

Which is also the warning. A wider gate is only wider if it is genuinely open. An interview in one language, needing a quiet room and a microphone, happening once, is narrower than a resume for some people, and those are the applicants you never hear from.

It can be, and the honest answer to the second half is: not unless somebody measures it.

Kyra Wilson and Aylin Caliskan tested the embedding models used to rank resumes and reported that they "significantly favor White-associated names", with the worst outcomes for Black male names (their study, read 30 September 2026). The same learned association can reach a score written from a transcript, and a transcript is itself a machine's guess at what was said, not equally good for every accent. We have not measured how Career1's varies, so we claim no figure for it.

So here is the whole of what Career1 can honestly say about its own bias: no audit has been published, nothing has been measured, and there are no customers whose selection rates could be examined. The interview is not blind; the name and the resume are in front of the model, and culture fit is the score to discount hardest. What is true is that the reasoning behind every score sits in the report in the candidate's own words, so a bias you suspect is one you can go and look for.

The controls that help are in how to reduce bias in first-round screening. The first is not a vendor feature: check your own selection rates by group, whichever tool you used.

The list a vendor should publish and mostly does not. For Career1's interview:

• It does not meet anyone. No handshake, no office, no sense of a room.
• It does not test work. No portfolio review, no coding exercise, no take-home. If the job is making something, it cannot tell you whether they can make it.
• It hears one conversation on one day. About eight minutes, instructed never to run past ten. Somebody ill, rushed or in a noisy room gets the version of themselves that day produced.
• It is in English only. A candidate who would do the job in another language is assessed on the wrong thing.
• It cannot know your team. It has your job description, not your standards, your last two bad hires, or the person the new hire sits beside.
• It cannot confirm identity or take up a reference. Separate problems, covered in fake job applicants.
• It has no outcome data. No score has been linked to whether a hire worked out, because there are no hires yet.

No. It should gather evidence, and a person should decide. That is not a philosophy but a question of which actions the software may take.

Here is Career1's, as a mechanism rather than a promise. Every applicant to a role is invited to interview. The AI recruiter speaks with each of them in the browser for about eight minutes, asks about the claims on their resume, and follows up when an answer is thin.

Then it writes a report: scores for match and overall performance, a recommendation from strong hire to no hire, five scored sections with their reasoning, and a ledger of which claimed skills were actually demonstrated, with the quote that shows it.

What it cannot do is reject anyone. There is no automatic rejection in the product, no score threshold that moves an applicant, and no setting that turns one on.

The AI produces an ordering and a recommendation. A person reads it and decides.

The overrule path matters more than the ranking, so be concrete:

1. You disagree. You think the eleventh person belongs near the top.
2. Read the report. Each score carries its reasoning, and the skill ledger quotes the sentence it was judged on.
3. Read the transcript, or watch the recording. Both are attached.
4. Decide the other way. Nothing is locked, because nothing was decided.

That is the difference between decision support and an automated decision: disagreeing is checkable rather than a matter of trust. A complete sample report and the pricing are published, so you can check both before paying.

That depends on where you hire. This is not legal advice, and it is the part of this page where you should take no vendor's word, including ours.

From an official page: New York City publishes that an employer may not use an automated employment decision tool "unless the tool has been subject to a bias audit within one year" and unless "certain notices have been provided to employees or job candidates" (nyc.gov, read 30 September 2026).

Obligations differ by country, state and city and they are changing, so this page names none of them. Two things hold wherever you are.

• A bias audit is something Career1 does not have. If a rule where you hire requires one before a tool of this kind may be used, Career1 does not meet it today. That is a reason not to buy, and better found here than later.
• Disclose, and keep the record. Say in the advert that an AI conducts the first interview, how long it takes, that it is recorded, and who reads it. A stored conversation explains a decision better than a memory.

The non-fit list, the same one as AI screening for small teams:

• You hire one person a year. Five applicants are five phone calls. Buy nothing.
• The role needs a work sample. For a designer, a writer or an engineer, ask for the work and pay for it.
• You need an applicant tracking system. Several roles and interviewers, agencies and referrals, is a coordination problem. Career1 is not an ATS and does not sync with one: see ATS or AI interviewer.
• Most of your pile is plainly ineligible. Form questions do that free, and resume screening is the cheaper filter.
• You interview in a language other than English.
• You need an independent bias audit or customer references first. Career1 is new and has neither.

Reading that list and walking away is a correct outcome.

Both pages were opened on 30 September 2026. Nothing here is legal advice.

• Wilson and Caliskan, "Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval": arxiv.org/abs/2407.20371.
• New York City, automated employment decision tools: nyc.gov.