Explaining AI Recommendations to Candidates

To explain an AI recommendation to a candidate, say in plain words why the tool was used, what it considered, what its output did and how to ask for human review. An AI recommendation in hiring is a signal from inputs to help a person review a candidate. It may suggest a profile, rank supplied CVs or highlight evidence against stated criteria.

An AI recommendation is not a fact about the person, proof of qualification, an inference about a protected trait, a prediction of job performance or a hiring decision. This is general educational information, not legal advice. A qualified privacy, employment and accessibility owner must decide what rules and notices apply to your organisation.

What decision does the AI recommendation support?

Record whether the system only organises information, recommends a human next step or makes a decision without meaningful human involvement. The ICO's jobseeker guidance, dated 31 March 2026, says people should know when automated decision-making is used, how it may affect an application and how to request a human review. Do not describe a ranking as "objective" or imply that a lower position proves a lack of ability.

The GDPR, adopted 27 April 2016, includes information duties in Articles 13 and 14 and safeguards for certain solely automated decisions in Article 22. The EDPB's automated decision-making and profiling guidance, endorsed 25 May 2018, and its WP251 rev.01 listing (guideline last revised 6 February 2018) are useful interpretive references. They do not determine your lawful basis or whether a particular hiring workflow is compliant.

What should a candidate notice about an AI recommendation say?

A candidate notice should answer five practical questions:

• Why was the tool used?
• What categories and sources were considered?
• What did the output do in this stage?
• What did it not do?
• How can the candidate correct, challenge or ask for human review?

Explain factors in plain language. Do not disclose passwords, security controls, another candidate's data, confidential prompts or proprietary weights. Explain purpose, inputs, output role, limits and consequences rather than dumping model internals.

The plain-language notice below is a fictional starting example, not a universal legal notice:

We used an automated tool to help organise information for this role. It considered
the role requirements and the professional information you supplied or that we
identified from the stated source. It produced a recommendation for a human review;
it did not decide your application or predict your job performance. The information
may be incomplete or out of date. Ask us what information was used, correct an error,
request a human review or raise a concern at [contact and accessible route].

Adapt the wording to the actual data, stage, consequence, lawful-basis analysis and jurisdiction. Do not say "no human was involved" if a recruiter reviewed the output, or call review meaningful when the reviewer cannot override, investigate or record disagreement.

How to test data quality, accessibility and fairness

Keep a small provenance record for each material input: source, capture date, purpose, owner, access group and correction route. Mark missing provenance, stale profiles and ambiguous skills as limitations. Public visibility is not proof of accuracy, permission to reuse or permission to infer sensitive information. Minimise data to the stated hiring purpose. Do not infer health, disability, ethnicity, religion, age, family circumstances or personality from names, photos, language, schools, locations or model output.

Review whether the notice and challenge route work for screen readers, keyboard users, zoom, contrast, captions and assistive technology. Provide an equivalent human channel and translated or plain-language options where needed. WCAG 2.2 is a W3C Recommendation dated 12 December 2024; it offers testable guidance, not a certification of a fair hiring process. The EEOC's 12 May 2022 technical assistance warns that algorithmic tools can screen out people with disabilities and highlights accommodation and medical-information risks. Fairness requires job-related criteria, non-discrimination checks, an adjustment route and accountable human ownership; an explanation alone does not prove fairness.

Apply privacy and security controls to the explanation record. Restrict access, authenticate reviewers, log exports, protect candidate communications and set a purpose-based retention review. Record who handles correction, access, objection, appeal and incidents. The NIST AI Risk Management Framework 1.0, published 26 January 2023, is voluntary guidance for governing, mapping, measuring and managing AI risk, not legal approval or an accuracy guarantee.

Which release gates apply before an explanation goes out?

Use these states in the review record:

EVIDENCE: purpose, inputs, source dates, output role, notice, owner and version are recorded.
UNKNOWN: a material source, limitation, consequence, accessibility check or contact route is unresolved.
HOLD: a named owner must resolve an UNKNOWN before the recommendation is shown or relied on.
STOP: contain the workflow when data is exposed, a discriminatory proxy is found, the route is inaccessible,
      or no human can investigate, override, correct and communicate the result.

Record a version or revision ID, model or rules identifier, role-brief version, input changes, reviewer, approval date, next review date and incident history. A material change or misrouted candidate record should trigger containment, correction, candidate communication where appropriate, rollback and a dated post-incident review. Preserve the old explanation and decision context so a later reviewer can tell what the candidate saw.

Fictional example: a platform-engineer recommendation

In this fictional example, a recruiter uses a tool to organise a platform-engineer CV and dated public profile. The recommendation cites cloud experience, but the profile has changed and its source date is missing. The privacy reviewer marks UNKNOWN, then HOLD until provenance, notice wording and a human contact route are confirmed. The recruiter corrects the record, explains that the output only supported review and records a manual decision. Later, unrelated candidate text appears in an export; the team marks STOP, limits access, investigates and records the incident. This is not a benchmark or hiring outcome.

Where Talent Summoner fits

Talent Summoner is our product for AI candidate sourcing, CV ranking and outreach you confirm before it sends (checked 28 September 2026). Candidate sourcing starts from a role brief and searches public professional sources for human review. Candidate ranking organises supplied CVs and provides plain-English reasoning. The product does not choose a lawful basis, manage rights or retention, provide appeals, verify every claim, automatically reject candidates or make the hiring decision. Keep privacy and employment owners responsible for the candidate-facing explanation.

Next, choose one real recommendation stage, draft the notice from its actual inputs and consequence, run the accessibility and non-discrimination checks, and obtain privacy-owner sign-off before the next hiring cycle. Then link the approved process to candidate sourcing and candidate ranking without promising a hiring result.