Dana is a technical recruiter at Cartwright Logistics, a 600-person freight company hiring backend engineers for its dispatch platform. Last spring she nearly forwarded a candidate to a final loop on the strength of an AI-generated summary that called him a "senior backend engineer with 8 years building distributed systems." The hiring manager asked one question Dana could not answer: which 8 years? She pulled the resume. It showed 5 years as an engineer and 2 as a tech lead, and "distributed systems" appeared nowhere. The AI had not lied. It had smoothed plausible-sounding numbers over a gap, and Dana had almost let it. Verification is the discipline that stands between a confident AI sentence and a hiring decision built on a number nobody checked.

What Verification Is, and What It Is Not

Verification is the systematic work of checking whether what the AI told you is actually true, measured against source material. It is not the same as spot-checking for obvious errors. Spot-checking catches the sentence that reads wrong. Verification catches the sentence that reads perfectly right and happens to be false, which is the more dangerous category, because a fluent, confident, well-formatted claim is exactly the kind a busy recruiter waves through. The cost is real: verification takes time. The payoff is that it keeps a fabricated "8 years" or an invented certification out of a decision that affects a person's career and your company's compliance posture.

The word "systematic" is doing real work in that definition. Verification is deliberate and methodical, and it always runs against a source: the resume, the interview notes, the professional profile, the reference, the actual reality behind the claim. When the AI asserts something, you do not weigh whether it sounds plausible, you go and look. That habit is what this lesson builds into a repeatable toolkit.

Three Domains You Verify

Dana verifies in three domains. The first is facts about the candidate: years of experience, specific technologies, named projects, locations, education. For each specific fact in the AI output, she identifies it, checks it against the resume, interview notes, and LinkedIn profile, flags discrepancies, and resolves them with a follow-up question. Her dispatch-platform example is exactly this: the AI said 8 years backend, the resume implied something between 5 and 7, and the resolution was a direct question to the candidate rather than a guess.

The second domain is sources and research. When the AI describes a passively sourced candidate, Dana checks whether the source actually exists, whether it is attributed correctly, and whether the interpretation holds. When an AI sourcing tool reported one candidate was "active on GitHub with 50 repositories," Dana opened the profile. There were 50 repositories, but 44 were forks of tutorials and 6 were original, two of them relevant. The number was true and the impression it created was false, which is why the rule is to look, not to trust the count.

The third domain is claims and assessments. When the AI says "strong problem-solver," Dana isolates the claim, finds the cited evidence, and asks whether the evidence actually supports it. "Debugged a complex production incident" is one data point, not a general trait, and treating a single anecdote as a conclusion is how an evaluation drifts from evidence into narrative.

There is a fourth move in this domain that recruiters skip most often, and it is the one that catches the subtlest errors. Once you have matched the claim to its evidence, ask whether the same evidence could reasonably mean something else. Debugging a gnarly production issue might indicate strong general problem-solving. It might equally indicate deep familiarity with one particular system, or persistence, or simply that the candidate was the person on call. Each of those reads differently against a role. If a single piece of evidence supports several competing interpretations, the honest output is not a verdict but a question you take into the next conversation, along with a note about what further evidence would settle it.

A Tiered Framework: You Cannot Verify Everything

Dana screens roughly 15 candidates a week. Verifying every clause of every AI output would consume her week, so she tiers her effort. Tier 1 is always verified: specific quantitative facts (years of experience, team size, scope), major qualifications (languages, frameworks, key skills and certifications), and anything that would change the decision if it were wrong. These are also, not coincidentally, the facts most likely to be fabricated, because specific numbers are exactly what a language model will confabulate to sound precise. Tier 2 is often verified: assessments of soft skills, claims about motivation or fit, and sourcing accuracy for passive candidates. Tier 3 is spot-checked only: general descriptions, background narrative, and non-critical descriptive detail. The framework is not about doing less work; it is about spending the verification budget where a wrong fact does the most damage.

Five Concrete Techniques

Resume verification is the workhorse: pull the resume, line up the AI's claimed facts against it, and document every discrepancy rather than mentally noting and forgetting them. Interview-notes verification applies the same move to behavioral claims: for every assertion about what the candidate said or did, find the supporting line in the actual notes. Independent source checking means opening the GitHub, the blog post, or the conference page yourself rather than trusting the AI's description of it, which is how Dana caught the forks-versus-original-repos problem. That check has three parts worth naming: is the claimed profile real, does it match the description the AI gave of it, and is any contribution attributed to the candidate genuinely theirs?

Reference verification matters most for senior roles, where a brief pre-loop reference conversation can confirm scope, tenure, and a couple of key skills before the team invests a full interview day. And the follow-up conversation is the catch-all: when a claim matters and cannot be verified from documents, ask the candidate directly. "Your background mentions distributed systems; walk me through the largest one you worked on" turns an unverifiable AI sentence into firsthand evidence, and it does so in a way that is fair to the candidate, who gets to speak for themselves.

Anti-Patterns to Avoid

The first anti-pattern is not verifying critical facts, especially specific numbers, on the assumption that a confident output is a correct one. Specific facts are the most hallucinated, so the Tier 1 rule exists precisely because skipping it is so tempting. The second is assuming sourcing is accurate, using the AI's description of a sourced candidate without ever looking at the underlying profile; the fix is to independently verify the source every time it feeds a real decision. The third is failing to account for changed information: reusing research from three weeks ago as if a candidate's profile, availability, or public work is frozen in time. For high-priority candidates, Dana re-verifies the load-bearing facts close to the decision, because stale truth and fresh falsehood produce the same bad outcome.

A Short Glossary

Verification is the systematic process of checking whether AI-generated information is accurate by comparing it against source material. Spot-checking appears in two senses in the material behind this lesson, and it is worth holding both. In the first sense it means sampling-based verification: you do not check everything, but you check the key facts, which is exactly what the tiered framework formalises. In the second sense it is contrasted with verification altogether, as a quick scan for errors that look obviously wrong rather than a deliberate comparison against a source. Read carefully, the difference is one of rigour rather than of kind, but be aware that the term is used loosely, and when you write your own protocol, say which sense you mean.

Tier 1 facts are the critical facts that would change your decision if they were wrong, and they are always verified. Tier 2 claims are important but not decisive, and they are often verified. Tier 3 details are non-critical information, and they are spot-checked only.

Practice: Build Your Own Verification Toolkit

Start with a role you are actively hiring for and write out your own three tiers. Which facts about a candidate for this specific role would change your decision if they were wrong, and therefore belong in Tier 1? Which claims are important but not decisive, and therefore sit in Tier 2? What is genuinely Tier 3 background narrative? The exercise is role-specific on purpose, because a certification that is decisive for a compliance hire is background colour for a sales one.

Next, take a real AI summary or assessment from your current pipeline and verify it properly. Work through your Tier 1 facts one at a time, check each against the resume, the notes, or the profile, and write the discrepancies down rather than carrying them in your head. Time yourself while you do it, because you will need that number for the protocol you write later.

Then practise source verification on its own. Find an AI output that makes research claims about a candidate, of the "active in open source" or "shipped five products" variety, and independently check each one. Which claims turn out accurate, which are inaccurate, and which are technically true while creating a false impression? That third category is the interesting one, and it is where Dana's fork-heavy GitHub profile would have landed.

With that experience in hand, design a verification protocol for your team's hiring process rather than just for yourself. What gets verified, at what stage, by whom, and roughly how long does it take? A protocol nobody has costed in minutes is a protocol that quietly gets skipped in a busy week.

Finally, turn the protocol into a template you can reuse: a simple structure that captures the claim, the source you checked it against, the outcome, and any follow-up needed. Be deliberate about what decisions the template is meant to support, because a verification record that nobody reads at decision time is paperwork rather than a safeguard.

Reflection Questions

What AI-generated information do you currently trust without verifying? Name it specifically rather than in general, then ask whether it belongs in a tier that requires checking.

Think back over your recent hiring. If you had verified the AI outputs you relied on, what might you have caught, and would any decision have changed?

What is the real time cost of verification in your process, measured rather than estimated? Weigh it honestly against the cost of a hire made on a fact that was never true.

And how would you build verification into your team's process rather than keeping it as a personal habit? A discipline that lives in one recruiter's head protects only that recruiter's candidates.

Putting It Into Practice

Verification is your safety mechanism. When you check AI outputs systematically, you catch errors before they influence hiring, and that is precisely the difference between AI-assisted hiring and AI-misled hiring. It costs time, and the time is the point: it is the price of knowing that the facts under a decision are real.

This week, implement the process for a single role. Pick one candidate, verify every Tier 1 fact in the AI output about them, and note two things as you go: how long it took, and what you found. Those two numbers are what will convince you, and later your team, whether the protocol is worth running at scale.

Related Lessons

Verifying Candidate Information: Spotting Hallucinations and Inaccuracies is the closest companion to this lesson. It focuses on recognising the fabricated claim in the first place, where this one focuses on the systematic process of confirming or refuting it.

Common AI Errors in Recruiting: Hallucinations, Misinterpretations, Omissions explains why specific numbers are the most fabricated category, which is the reasoning behind putting quantitative facts in Tier 1 without exception.

Scenario Practice: Review and Critique puts the verification habit to work on full candidate summaries, including a side-by-side comparison of an AI draft against the resume it was supposedly built from.

Avoiding Automation Bias: Staying Active and Skeptical addresses the underlying reason verification gets skipped: a fluent, confident output invites agreement, and staying skeptical is a deliberate act rather than a default state.

Documentation and Evidence: Building a Trail for Compliance takes the discrepancy log from your verification template and shows how records of what was checked, and when, become the evidence that a decision was made on verified information.

Where This Goes Next

Good hiring is built on verified information. AI is a tool, and you remain responsible for the accuracy of everything that feeds your decisions, regardless of which system drafted the sentence. Having built a personal verification toolkit, the work ahead is to move from reviewing individual outputs to building the systems and the team culture that make responsible AI use the default rather than the exception.

Key Takeaways

• Verification checks AI claims against source material; spot-checking only catches obvious errors. The dangerous claim is the fluent, confident, perfectly formatted one that happens to be false, and only deliberate verification catches it.
• Verify in three domains: candidate facts, sources, and assessment claims. Each has its own move: compare facts to the resume, open the source yourself, and trace every "strong X" back to the evidence cited for it.
• Tier your effort so the budget lands where a wrong fact hurts most. Always verify Tier 1 quantitative facts and major qualifications, often verify Tier 2 soft-skill and fit claims, spot-check Tier 3 narrative detail.
• Specific numbers are the most hallucinated category. Years of experience, team size, and scope are exactly what a model confabulates to sound precise, which is why they sit in Tier 1 without exception.
• Open the source instead of trusting the description. A true count like "50 repositories" can create a false impression; looking yourself is the only way to separate the number from the meaning.
• Re-verify load-bearing facts near the decision. Stale research and fresh fabrication produce the same bad hire, so refresh the critical facts for high-priority candidates before the final call.