An AI Interview Notice Needs to Name What Gets Scored

Picture the notice screen immediately before a recorded interview. The applicant has already uploaded a résumé, entered contact details, and clicked a link from the employer. A checkbox says the company may use artificial intelligence to evaluate the interview. The only available buttons are “I agree” and “Continue.”

That screen confirms almost nothing. It does not reveal whether software merely transcribes the answers, scores their content against a rubric, ranks the applicant, or analyzes voice and video signals. It does not identify who receives the score, whether a low result can end the application, or when the recording will be deleted.

Some jurisdictions impose specific notice duties for certain hiring systems. Other details in a useful notice are design controls rather than universal legal requirements. Employers need to separate those categories, then map the notice to the system they have deployed rather than copying a broad AI disclosure from another workflow. This explainer is not legal advice.

Start with the decision the system supports

The first useful sentence on that notice screen should name the system’s purpose and its place in the hiring sequence. “Software will transcribe your recorded answers and score them against the communication and customer-service criteria for this role” conveys more than “AI may be used.” So does a statement that a recruiter will review the transcript and score before deciding whom to advance.

The distinction matters because interview platforms do different jobs. One product may generate a transcript and summarize it for a recruiter. Another may compare answers with employer-defined criteria, producing a recommendation or rank. A third may combine the interview with résumé data or an assessment completed elsewhere.

Calling all three “AI-assisted interviewing” conceals the operative mechanism.

The notice should also state whether the output can remove an applicant without human review. If a recruiter can override the score, the employer should describe when that review occurs rather than promising generic “human involvement.” A person who checks only the highest-ranked candidates is not reviewing a rejection produced below the cutoff.

Illinois supplies a concrete requirement for a narrower workflow. Its Artificial Intelligence Video Interview Act covers employers using AI to analyze applicant video interviews for Illinois-based positions. Among other duties, the law says an employer must:

“provide each applicant with information before the interview explaining how the artificial intelligence works and what general types of characteristics it uses to evaluate applicants.”

The law also requires advance notice that AI may analyze the video and consider the applicant’s fitness for the position, plus consent before that evaluation. Those requirements are enforced rules for covered interviews, not a general template for every hiring tool in every state. Still, the quoted language exposes the weakness in a bare consent box: consent follows an explanation of operation and evaluated characteristics.

Name the inputs, not just the recording

The notice screen should distinguish collected data from inferred data. Collected data includes the recording, audio track, transcript, typed responses, timestamps, and any assessment answers sent into the system. Inferred data means conclusions generated from those inputs, such as a score for a job qualification or a prediction that an answer matches an employer’s rubric.

A camera does not prove that facial expression is scored. A microphone does not prove that vocal tone is scored. Employers should not imply that every available signal is analyzed, but they also should not hide an active analysis behind the broad term “interview data.” If the platform scores only the words in a transcript, the notice can say that visual appearance, facial movement, background, and vocal characteristics are not evaluation inputs.

If another signal is used, it should be named and connected to the relevant job criterion.

This specificity gives applicants a practical check on likely failure modes. Automatic speech recognition, software that converts speech into text, can mishear names, specialized terminology, accented speech, or words recorded through a weak connection. A content score built on that transcript may then evaluate text the applicant never said. Video and audio analysis can also interact with disabilities, assistive technology, speech differences, lighting, or an inability to maintain conventional eye contact.

The employer therefore needs an operational fallback. It might preserve the original recording so a recruiter can compare disputed text with the audio, allow an applicant to correct a transcript before scoring, or offer a live structured interview that asks the same job-related questions. Each option costs reviewer time and can slow the hiring queue, but an automated score without a correction path transfers the system’s error to the applicant.

Return to the checkbox. If the employer cannot fill in the sentence “The system examines ___ to produce ___, which is used by ___,” the deployment is not ready for meaningful notice. The missing words usually indicate that procurement, recruiting, and the vendor have not agreed on the data flow.

Retention needs an owner and an endpoint

“Data is retained as required by law” is not a retention explanation. A useful notice identifies which records remain after the interview, who holds them, the retention period or the rule used to determine it, and how an applicant can request deletion where that option exists. It should distinguish the raw recording from derivative records such as a transcript, score, reviewer notes, or audit log.

Deletion is harder than removing a video from the recruiter dashboard. Copies may sit with the interview vendor, a transcription provider, the employer’s applicant tracking system, or a backup service. A deletion workflow needs an owner who can propagate the request and confirm which records were removed, which were retained, and the stated reason for any retention.

Illinois again imposes a specific rule within its scope. After an applicant asks for deletion, a covered employer must direct others holding copies to delete the videos, including electronically generated backup copies, and deletion must occur within 30 days after the request. That statutory duty concerns the videos. Employers still need to determine how other records, including transcripts and scores, are handled under applicable policies and laws.

New York City’s automated employment decision tool rules take a different approach. For covered tools, employers or employment agencies must provide advance notice that the tool will be used and identify the job qualifications and characteristics it will assess. Instructions must also be available for requesting information about the type and source of collected data and the retention policy. The city’s required bias audit, which compares selection or scoring rates across specified groups, does not explain an individual applicant’s score or replace a data notice.

Accommodation and review must be usable

An accommodation statement should appear before the applicant records an answer, not behind a general careers-page link discovered afterward. It needs a working contact method, examples of the stage that can be changed, and enough timing information for the recruiting team to pause the automated route while handling the request.

The Americans with Disabilities Act already constrains how employers use selection procedures, including software that may screen out a person because of a disability rather than an inability to perform the job. Federal guidance has warned that algorithmic tools can create this problem. That does not create one universal AI-notice script, but it makes a functional accommodation route more than a courtesy.

An alternative route should preserve the substance of the assessment. A live structured interview, accessible written response, or recruiter-reviewed recording may be appropriate depending on the role and the barrier. The alternative should not quietly impose a tougher test or move the applicant to an unmonitored inbox.

New York City requires notice that a candidate may request an alternative selection process or reasonable accommodation. City guidance also distinguishes the ability to request an alternative from an automatic duty under that AI law to provide one; separate disability law obligations may still apply. That is the difference between describing an available request and promising a particular result.

Appeal information is less consistently mandated, but it is a sound control whenever a score can materially affect progression. The notice should identify what can be challenged, where the request goes, whether a human will inspect the source material, and whether the application will be held open during review. Calling a recruiter does not count as an appeal process if the recruiter cannot see the transcript, model output, or scoring rubric.

Treat the notice as a system record

The final version of the notice screen should be retained with a version identifier, the date range in which it was shown, the job workflow it covered, and evidence that the applicant saw it before analysis began. Consent logs should point to that version rather than storing only a checked box.

This creates a receipt. If the vendor later adds transcript scoring, changes the characteristics evaluated, or begins sending data to another processor, the employer can compare the new workflow with the notice applicants received. The update trigger should be a material change in processing, not the next annual policy review.

The same discipline prevents overclaiming. A notice should not promise that a human “makes every decision” unless the workflow enforces that review and records it. Nor should it claim that recordings are never shared when a contracted vendor processes them. Policy text cannot compensate for a product configuration that does something else.

For the applicant waiting at that original screen, the useful version is longer than one sentence but does not need to become a privacy policy. It needs a clear account of the input, output, decision role, retention rule, accommodation channel, and review route, with links for the operational detail. The checkbox comes last.

Questions people ask

Can an employer use AI without telling every applicant?

Notice duties depend on the jurisdiction, tool, applicant, and workflow. Illinois has requirements for covered AI-analyzed video interviews, while New York City regulates covered automated employment decision tools. A company may also choose broader notice as a governance control, but that voluntary practice should not be described as a universal legal requirement.

Should a notice say whether facial expressions or voice are analyzed?

Yes, if those signals are evaluated, the notice should name them and explain the job-related purpose. If the system uses only a transcript or answer content, saying that appearance and vocal characteristics are not scored helps applicants understand the mechanism and avoids implying capabilities the platform does not use.

Does human review solve the problem?

Only if the reviewer can inspect the relevant source material, understands the score’s limits, and has authority to change the result before rejection. A person who sees only applicants above an automated cutoff cannot correct errors affecting everyone below it, even if the company describes the workflow as human supervised.

What should an applicant save before starting?

Save the notice, privacy link, consent language, accommodation instructions, and any stated retention or deletion terms. A screenshot of the pre-interview screen can establish what was disclosed at that time, especially if the platform later changes its language or the applicant needs to identify the data and score used.