Can employers use AI chatbots to review background checks?

Once a background check completes, employers routinely find themselves looking up unfamiliar language. In that sense, AI chatbots have been a really helpful evolution of traditional search engines. But that's also where the temptation begins and where an information search can quickly turn into over reliance on an unverified hiring recommendation.

A few things to keep in mind before you paste a report into a chat window:

The report is full of personal data. To start, background checks usually contain the candidate's full name, date of birth, addresses, aliases, and every record found. Reports with investigatory screenings like employment and education verifications contain even more sensitive details that would make an identity thief drool. When you paste it into an account your company doesn't control, your company may have no oversight of what is retained, reviewed by people, or used for training their models.^[1] Additionally, your screening agreement may restrict sharing reports with third parties, so an upload can create a contractual problem before the chatbot even answers.

Record interpretation is nuanced. Criminal-record terminology in particular often depends on the specific jurisdiction and statute the record comes from. For example, New Jersey uses classifications such as "indictable offenses" and "disorderly persons offenses," rather than the familiar felony/misdemeanor labels. And a word like "solicitation" can describe very different conduct (solicitation of prostitution or asking someone to commit another crime).^[2]

Compounding this issue, most reports are still designed as a "document" first with each provider delivering their own style and formatting. Extraction can lead to important text getting dropped, columns being merged, or dates being misread. The model then writes confidently about record data that may not match the original report. These mistakes can send the record into the wrong policy category and change how the candidate is ultimately assessed. Detailed prompting helps, but the interpretation still needs support from the statute and record details—not simply the meaning most commonly associated with the word.

Consistency is not their strength. Ask a chatbot the same question about the same report tomorrow and you may get a different paragraph, and maybe a different conclusion. This is very different from code, where classification rules are written deterministically. Without that, you can't show that two similar candidates were treated the same way, you can't see how your reviewers are actually deciding, and you can't hand anyone a meaningful audit log. Unexplained differences in how similarly situated candidates are treated can become evidence in a discrimination claim, leaving the employer to explain why one person's record cost them a job while another's did not.^[3]

It's missing critical context. It's very easy to misread the confidence in chatbot responses as accuracy. But a general chatbot simply doesn't have the proper context about you or your organizational needs to conduct a proper (or compliant) review. Federal, State and local law all come into play in the evaluation process, and determining whether an individual record is relevant is only part of the process. Even with your policy in the prompt, a model may count a lookback period from the offense date rather than the conviction date your criteria require, changing whether the record gets marked for review.

A better prompt doesn't establish a reliable process

Suppose your reviewer uses an approved business account with the right no-training settings, provides the job duties and adjudication matrix, and even asks the model to consider the offense, elapsed time, and candidate's circumstances. That definitely improves the starting point, but several gaps remain.

A convincing explanation can conceal blanket assumptions. A model might reason that a theft conviction matters for a warehouse job because employees have access to inventory. But that connection could justify rejecting someone with almost any theft history from almost any job involving property. It does not explain why this particular record (given its age, circumstances, and the candidate's subsequent history) warrants exclusion from this particular role.^[3]

The danger is that the answer sounds individualized because it mentions the candidate's offense and job title. While underneath, it simply restates a broad and potentially flawed assumption: someone who stole before presents a theft risk now. Polished language can make that assumption easier to accept without examining whether the evidence actually supports it.

The process should be consistent, even when outcomes differ. Individualized assessment means a candidate's circumstances can change the preliminary decision. Evidence of rehabilitation may justify a different outcome, and the reviewer should document why it mattered.^[4] But a reworded prompt or model update can change how the same facts are weighed without any deliberate change to your review process. That makes it harder to explain whether a different recommendation reflects the candidate's circumstances or the tool's variability. The review record needs to connect the evidence, applicable criteria, and reviewer's judgment to the decision.

Decision history helps establish a baseline

What if the model also had your organization's past approvals and rejections? That might help it produce recommendations that resemble previous decisions. But it raises another question: which parts of that history should guide the next review?

Decision history can carry old problems forward. Giving a model your adjudication history can certainly help reproduce past behavior. But that history might include unintentional blanket exclusion or restrictions with little connection to the job. If those decisions become examples of the "right" answer, the model can carry the same weaknesses into new recommendations, expanding the employer's exposure rather than improving the review process.^[5]

Historical analysis is valuable for identifying inconsistent treatment, undocumented exceptions, and policy drift. Those findings should guide scrutiny of the process, with relevant research and workplace outcomes helping test the assumptions behind it. But past decisions are evidence of what an employer did—not proof that it should keep doing it.

What to do instead

To start, evaluate any review tool (whether it's a chatbot or an AI feature inside your screening platform) by how it supports compliant assessments, handles uncertainty, and documents the employer's decision.

Protect candidate information. Prohibit uploads to unapproved tools and review how any provider handles report data, including access, retention, and reuse.^[6] An approved account is the starting point for evaluating whether a tool is appropriate for this work.

Establish criteria you can justify and test. Document screening criteria by role, review their basis in evidence and applicable requirements, and test how they handle different charges, dispositions, dates, and incomplete records. Use decision history to identify inconsistent application and policy drift, while periodically examining whether the criteria themselves remain defensible.

Keep the final decision with a trained reviewer. A human being should be making the final call. Reviewers should be trained to apply screening policies consistently, to conduct defensible individualized assessments, and to document the rationale behind every decision. Preserve that reasoning alongside the policy version, candidate response, and final decision so any change can be traced to its substantive basis.

How Kizuna supports a better review

Kizuna is built to make defensible background check review practical at scale. Our classification engine accounts for the complexity and jurisdictional nuance of record data. That data is then evaluated against the employer's own criteria, defined through hundreds of configurable settings so the rules can reflect actual job risks rather than broad category bans. Flagged records route to a reviewer equipped with research-backed risk signals, candidate context, and the structure for a compliant individualized assessment.

Together, these capabilities bring record interpretation, policy application, research, and documented human judgment into the same process, giving employers everything they need to make smart, fast, and defensible decisions.

So, can employers use AI chatbots to review background check reports? With the right controls, they can certainly help explain unfamiliar concepts and identify questions worth investigating. But using them to assess a candidate requires a process that verifies the record, applies justified criteria, considers individual circumstances, and documents the employer's judgment. A convincing answer in a chat window does not establish that those steps occurred.

Kizuna is built to give employers that structure.

Book a demo to see how record interpretation, explicit policy rules, and research-based context come together in a review your team can explain and support.

This post is general information, not legal advice. Talk to your counsel about how these issues apply to your organization.

---