Can an employer fire someone based on an AI detector score?

You can, and that is the problem. A detector returns a statistical likelihood that text resembles machine-generated writing; it does not observe the act of generation and cannot identify a tool, an account or a moment. Acting on the number alone leaves you with no evidence of the conduct you are alleging, which is a weak position in an unemployment hearing, a wrongful-termination claim, or a defamation suit over what you told colleagues about why the person left.

Why do AI detectors flag human writing?

They score how predictable text is. Writing that is plain, formulaic, or produced by someone composing in a second language tends to be more predictable than idiosyncratic native prose, so it scores as machine-like. The same is true of technical documentation, templated business writing and anything edited toward clarity. Grammar assistants and translation tools push scores up further. None of that is evidence of anything except that the writing is conventional.

Is there a discrimination risk in using detectors?

Yes, and it is the sharpest of the exposures. If a screening tool's false positives concentrate on writers whose first language is not English, a policy of acting on those scores produces adverse outcomes correlated with national origin. That is the classic shape of a disparate-impact claim: a facially neutral practice with an unequal effect that the employer must then justify as job-related and consistent with business necessity. A vendor accuracy claim is not that justification.

Does it matter if a third party runs the detection?

It can change the legal machinery entirely. When an outside vendor assembles information about a person and you use it for an employment decision, the arrangement can fall under consumer-reporting rules that require disclosure, authorization, a pre-adverse-action notice with a copy of the report, and an opportunity to dispute it. Many teams paste candidate writing into a third-party tool and skip all of that because it does not feel like a background check.

What should we do if we genuinely suspect undisclosed AI use?

Treat the score as a prompt to investigate, never as the finding. Ask the person, in writing, to explain their process and show working material — drafts, version history, sources, notes. Compare against work you know is theirs. Document what you found rather than what the tool reported. If the outcome is adverse, the file should be able to state the conduct without mentioning a percentage at all.

Should we have a written AI use policy instead?

Almost always. Most detector disputes are really policy vacuums: nobody said whether AI assistance was allowed, so there is no rule to have broken. A short policy that states where AI is permitted, where it must be disclosed, and where it is prohibited converts an unprovable accusation about how text was produced into a checkable question about whether a disclosure was made.

AI Detector False Positives 2026: The Legal Risk of Acting on a Score

What the Number Is Measuring

These tools do not recognise text. They estimate how surprising a passage is to a language model — roughly, how often the next word is the word a model would have picked, and how much that varies across the passage. Machine-generated prose tends to be smoother and less surprising than human prose, so smoothness becomes the proxy for machine authorship.

Everything that makes writing conventional therefore raises the score. Writing in a second language. Writing to a template. Technical documentation. Legal and medical prose. Anyone who ran their draft through a grammar assistant, which is most people. The score is a measurement of style, and style is not conduct.

Four Different Claims, One Bad Decision

Employment exposure

•Disparate impact if false positives track national origin or disability
•Wrongful termination where the stated cause cannot be evidenced
•Unemployment benefit disputes the employer loses on proof
•Retaliation claims when detection is applied selectively after a complaint
•Union and contract grievances over just-cause standards

Reputational and procedural exposure

•Defamation risk in repeating the accusation to colleagues or references
•Consumer-reporting duties when an outside vendor scores a candidate
•State AI and automated-decision notice rules triggered by screening use
•Confidentiality breach from pasting client work into a third-party tool
•Contract termination disputes with freelancers and agencies

The Disparate-Impact Problem, Precisely

A policy does not have to be aimed at anyone to be unlawful. If a neutral screening practice produces meaningfully worse outcomes for a protected group, the employer carries the burden of showing it is job-related and consistent with business necessity — and that a less discriminatory alternative was not available. A detector whose errors cluster on second-language writers fails every part of that test at once: the connection to job performance is speculative, the vendor cannot explain any individual score, and an obvious less-discriminatory alternative exists, which is asking the person for their drafts.

The Accused Cannot Win, Which Is the Tell

Ask what evidence would clear someone. Version history can be manufactured. Drafts can be fabricated. A confident denial reads as a confident denial whether or not it is true. When a process offers the accused no way to establish innocence, it is not an investigation — it is an accusation with paperwork, and that is the characteristic every adjudicator is trained to notice. The same structural defect is why academic institutions that adopted detectors early have been walking the practice back.

A Process That Survives Review

Before Any Detector Is Used

☐Publish an AI use policy: permitted, must-disclose, prohibited
☐Decide in advance that a score can start an inquiry and cannot end one
☐Check whether pasting the work into a vendor breaches a client confidentiality term
☐Confirm whether third-party scoring pulls you into consumer-reporting duties
☐Apply any detection uniformly, not just to people already under scrutiny

If You Suspect Undisclosed Use

☐Ask in writing for process, drafts, sources and version history
☐Compare against known samples of the person's work
☐Document the conduct you found, not the percentage you were shown
☐Give a real opportunity to respond before any adverse step
☐State the cause in terms you could prove without the tool

Frequently Asked Questions

Are AI detectors accurate enough to rely on?

Not for an adverse action against an individual. Even a tool with a genuinely low error rate across a large corpus produces a meaningful number of wrong answers at the level of one person's document, and it cannot tell you which answer is the wrong one. Accuracy claims are aggregate statements; a termination is a single case.

Why do second-language writers get flagged more?

Detectors score predictability, and writing produced in a second language is often more conventional in vocabulary and sentence construction than native prose. Translation and grammar tools smooth it further. The result is a higher machine-likeness score for reasons that have nothing to do with how the text was produced.

Can we be sued for accusing someone wrongly?

Repeating a false accusation of dishonesty to people who do not need to know it is the core of a defamation claim, and internal circulation is not automatically privileged. Separately, an adverse action you cannot evidence supports wrongful-termination and unemployment claims, and a screening practice with skewed error rates supports a discrimination claim.

Does the FCRA apply to running a candidate's writing sample through a tool?

It can. If an outside vendor assembles or evaluates information about a person and you use the result in an employment decision, consumer-reporting obligations may attach — disclosure, authorization, pre-adverse-action notice with a copy of the report, and a chance to dispute. Many teams skip these because a detector does not feel like a background check.

What about freelancers and agencies rather than employees?

The employment statutes may not reach the relationship, but the contract does. Terminating for cause on an unprovable allegation invites a breach claim and a fight over unpaid work, and if you told anyone else why the engagement ended, the defamation exposure is identical.

What is the actual fix?

A written policy plus a process requirement. State where AI is allowed and where it must be disclosed, and require that any adverse action rest on documented conduct — an undisclosed use the person acknowledges, a fabricated citation, a verifiably copied passage — rather than on a score. That converts an unfalsifiable accusation into a checkable fact.

The Same Question, Asked of Your Website

The failure here is acting on a number nobody can trace back to a specific, nameable defect. That is the difference between a score and a finding, and it is worth applying to every automated report your team acts on.

A scan of your own site returns the rule, the element and the line rather than a percentage. Run a free scan and see what a finding looks like when it names the thing at fault.

AI Detector False Positives: The Legal Risk of Acting on a Score