What Is AI Hallucination? Risks in Recruiting Contexts
AI hallucination describes a language model producing confident, fluent text that is factually wrong or entirely made up. In low-stakes writing tasks this is an inconvenience. In recruiting — where outputs drive decisions about real people — it can mean fabricated qualifications, unsupported scoring explanations, or false compliance guidance.
What Hallucination Is
Language models generate text by predicting probable next tokens given their input and training. They do not look things up in a live database or verify facts before writing. When a model's training data doesn't cover a topic, or when a prompt pushes the model toward a confident-sounding answer, it can generate plausible-sounding text that is simply wrong.
Hallucinations are distinct from vagueness or uncertainty. A hallucinating model often states false information with the same confident, fluent style it uses for accurate information — there is no built-in signal that something is wrong.
Hallucination Risks Specific to Recruiting
- Fabricated candidate qualifications: An AI summary may state a candidate has a certification, degree, or experience they did not mention — especially if the model infers from similar-sounding profile language.
- Invented outreach facts: AI-generated personalised messages may reference a candidate's employer, project, or experience inaccurately.
- Unsupported scoring explanations: A scoring summary may explain a decision with reasoning that doesn't trace back to the candidate's actual answers.
- False legal citations: When asked about compliance requirements, a model may cite regulations, case law, or agency guidance that does not exist or was misquoted.
- Incorrect benefit or offer details: AI-drafted offer summaries may misstate compensation figures, benefits, or start dates.
Hallucination vs. Retrieval Error vs. Bias
These are related but distinct failure modes:
- Hallucination: The model generates content not grounded in any source — it is fabricated.
- Retrieval error: The model retrieved the wrong document or the wrong passage, and accurately described it — the content is real but misattributed or out of context.
- Bias: The model produces outputs that systematically favour or disfavour certain groups based on patterns in training data — the outputs may be internally consistent but unfair.
Fixing hallucination (better grounding, source citation) does not fix bias. Both require separate attention.
How to Reduce Hallucination Risk in Recruiting Workflows
- Ground outputs in source documents: Systems that answer from retrieved documents (RAG) are less prone to fabrication than models generating from general training alone — but retrieval can still fail.
- Require citations: Ask the system to link every claim to a source — a transcript excerpt, a profile field, a document section. Unsupported claims become visible.
- Human review before action: No AI output should trigger a consequential action — a rejection, an offer, a compliance decision — without a human checking the supporting evidence.
- Test with adversarial inputs: Ask the system questions whose correct answer is "I don't know" or "this information isn't in the source." A well-calibrated system should express uncertainty; a hallucinating one will invent an answer.
- Separate generation from retrieval: Keep the data layer (what the candidate actually said) distinct from the generated summary layer. If a recruiter can inspect the original transcript, they can catch summaries that misrepresent it.
When Hallucination Is Especially Dangerous
The risk is highest when AI output directly influences a consequential decision with limited human review:
- Automated scoring or ranking where the score drives advancement without recruiter review of the evidence
- Legal or compliance guidance generated on the fly (recording consent rules, AEDT applicability, accommodation requirements)
- Offer letters or agreements where numbers or terms may be stated incorrectly
- Candidate-facing communications that assert facts about their own application or qualifications
In each case, the mitigation is the same: a human who can verify the claim against primary sources before the output is acted on.