What Is Hallucination?

Hallucination happens when generative AI generates false information confidently. The AI doesn't "know" it's making something up. It's predicting the next token of text based on patterns it learned. Sometimes those patterns lead to invented details.

Example: You feed a resume to an AI with the prompt: "Summarize this candidate's key accomplishments." The resume says "contributed to Django project, worked with PostgreSQL database." The AI outputs: "Led refactoring of Django application to improve database query performance by 40%."

The AI hallucinated details that weren't in the resume. It pattern-matched "Django" and "database" to skills it knows matter, then invented accomplishment magnitude (40% improvement) that sounds plausible.

Why does this happen? Generative AI is a language prediction model. It learned patterns like "software engineers lead refactoring, database improvements, performance gains." It pattern-matched your candidate to these templates and filled in details. This is its core failure mode.

Hallucinations in Resume Summaries

Common Patterns

• Invented seniority: Resume says "worked on infrastructure projects." AI summarizes: "Led infrastructure migration to cloud." (Hallucinated leadership)
• Fabricated metrics: Resume says "improved efficiency." AI summarizes: "Improved operational efficiency by 25%." (Hallucinated number)
• Overstated scope: Resume says "contributed to React project." AI summarizes: "Architected React component library used across 5 products." (Hallucinated scope)
• Invented technologies: Resume mentions "APIs." AI summarizes: "Designed REST API architecture." (Resume might not specify REST vs. GraphQL vs. other)

Real Recruiting Consequence

You read the AI summary. You see "led infrastructure migration to cloud." You form an impression: senior engineer with cloud expertise. During the interview, you ask detailed questions about the migration. The candidate looks confused: "I didn't lead that. I contributed to it."

You now have false information baked into your assessment. You might overestimate seniority or expertise. The candidate doesn't get the job partly because of AI-invented accomplishments.

Hallucinations in Interview Synthesis

Some AI tools transcribe and summarize interviews. They can hallucinate accomplishments, misquote candidates, or invent details.

Examples

• Misquotation: Candidate said: "I learned Python at my last job." AI summary: "Candidate has advanced Python expertise." (Changed qualified claim to absolute expertise claim)
• Invented expertise: Candidate mentioned "worked with Kubernetes" once offhandedly. AI summary: "Kubernetes expert with strong understanding of container orchestration." (Overstated based on single mention)
• Tone misinterpretation: Candidate said hesitantly: "I think I could learn that technology." AI summary: "Confident in ability to learn new technologies." (Misread hesitation as confidence)

The Confidence Problem

The most dangerous part of hallucination is that AI presents false information confidently. A human who's unsure might say: "I think the candidate mentioned Docker... maybe?" An AI says: "Candidate has Docker experience" as a definitive statement.

You're more likely to trust confident assertions. So hallucinations backed by specific details ("40% improvement," "led migration") are believed more than tentative claims would be.

Information Loss in Summarization

Beyond hallucination, there's information loss: the summary omits important details.

Examples

• Context loss: Resume mentions "managed team of 5." AI summary: "Experience with team management." (Lost the scale—managing 5 is different from managing 50)
• Nuance loss: Resume says "led project from conception through launch, involving 18 months of planning and design before development started." AI summary: "Project leadership experience." (Lost the timeline and emphasis on planning)
• Skill level loss: Resume says "familiar with Python, expert in Java." AI summary lists both as skills without distinction in expertise level

Summaries prioritize brevity over completeness. Important nuances get lost. You make decisions on incomplete information.

How to Mitigate Hallucination Risk

1. Never Trust AI Summaries Alone

If you use AI to summarize resumes or interviews, always verify the summary against source material before using it for hiring decisions. Read the actual resume. Listen to key interview clips. Don't rely on the summary alone.

2. Treat Summaries as Starting Points

Use AI summaries to quickly understand content structure. "What are the main projects this candidate mentions?" Then read the source material yourself for details, nuance, and accuracy.

3. Look for Specificity as Red Flag

Specific numbers, percentages, and detailed claims should trigger verification. "Improved efficiency by 25%" should lead you back to the resume to check: did the candidate claim 25%? A specific number is either accurate (from the source) or hallucinated. Verify.

4. Compare Multiple Sources

For interviews: compare the AI summary to the interviewer's notes and (if available) the recording. Do they match? If the interviewer said "candidate wasn't sure about this topic" but the AI summary says "confident understanding," something's wrong.

5. Use AI for Information Extraction, Not Assessment

AI can extract facts: "candidate lists Python, Django, PostgreSQL." That's pattern-matching and works reasonably well. AI shouldn't assess: "candidate is confident in Python." That requires context and judgment AI doesn't have.

6. Build Verification Into Your Process

If your ATS uses AI summaries, require that hiring managers flag any claims they want to use in interviews or decisions. Have someone verify those claims against source material before the interview. This catches hallucinations before they influence decisions.

Practical Checklist

AI Output	Verification Step	Used For
Resume summary with accomplishments	Read full resume; verify accomplishments stated	Screening only; not hiring decisions
Interview transcript	Listen to recording for key quotes	Reference only; assessment done by interviewer
Skill extraction ("Candidate has Python")	Verify skill explicitly mentioned in source	Filtering and screening
Competency assessment ("Strong communication")	Do not use without human verification	Only after human validates

This lesson addresses fundamental concepts about AI in recruiting. Understanding these principles is essential for responsible use of AI-assisted recruiting tools and processes.

Practical Applications

In practice, these concepts apply across recruiting workflows. From sourcing through offer, understanding how AI impacts each stage helps you make informed decisions about tool adoption and usage.

Risks and Considerations

Every use of AI in recruiting carries considerations. Bias risk, candidate experience, legal exposure, privacy concerns. This lesson helps you identify and mitigate these risks.

Key Takeaway

Frequently Asked Questions

How can we tell if AI has hallucinated in a resume summary?

Look for claims that are more specific or impressive than what the source likely said. Resume says "worked on projects," AI says "led 5 major projects." Specific numbers, percentages, and definitive claims should be verified. Also compare the summary tone to the resume tone. If the resume is humble and the summary is boastful, hallucination likely occurred. And if specific claims matter for your decision, verify them before deciding.

Should we stop using AI summaries entirely?

Not if you verify. AI summaries are useful for quickly getting the gist of a resume or interview. The key is treating them as starting points, not gospel. Use them to identify candidates worth deeper review. Then do your own reading/listening before making decisions. The verification step is essential but manageable for a reasonable flow of candidates.

What's the difference between hallucination and information loss?

Hallucination is inventing false information. "Candidate led project" when the resume doesn't say "led." Information loss is omitting true information. Candidate has 5 years Python experience, but summary just says "Python experience" without the years. Both are problems, but hallucination is more dangerous because it's false. Information loss you can mitigate by checking source material yourself.

Can we use AI to detect hallucinations in other AI outputs?

Not reliably. AI tends to validate other AI outputs (they pattern-match each other). The only reliable way to detect hallucination is comparison to source material. Have a human read the source and compare to the AI summary. This is the verification step. You can't AI your way out of AI hallucination.

What should we do if we discover we hired someone based on hallucinated information?

If you discover during onboarding that someone doesn't have skills the AI hallucinated, address it immediately. You might: ramp them slower, provide more training, or reassign them if the hallucinated skills were critical. Don't blame the candidate—they didn't hallucinate. Blame the process. Use this as a learning event to strengthen your verification process. And consider whether you can provide accommodations or training to help the hire succeed anyway.