Dana coordinates recruiting at a 90-person fintech startup, and she is the only person on the team who screens with AI every day. On a Tuesday she pulled up an AI-generated assessment of a backend engineering candidate and read a sentence that stopped her cold: "Strong candidate, though she may want to prioritize family over the on-call rotation." Nothing in the resume said anything about family. The model had invented a concern out of a first name and a two-year gap, and it had attached it to a hiring recommendation Dana was about to forward to the engineering manager. That moment is the whole subject of this lesson. AI output is not pass or fail. Some of it is good, some is 80 percent good and needs a quick fix, and some carries a defect serious enough that the right move is to throw it out or hand it to a human. Knowing which is which is a skill, and Dana learned to make the call in seconds instead of forwarding a liability.

Three Verdicts, Not Two

The instinct most recruiters bring to AI output is binary: the assessment is either usable or it is not. That framing causes two opposite mistakes. It tempts you to keep flawed output because redoing it feels slow, and it tempts you to escalate everything because a flawed output once burned you. The fix is to recognize three verdicts, not two. An output can be fixed, in which case you correct a small error and move on. It can be rejected, in which case the defect is severe enough that nothing in the output can be trusted and you start over. Or it can be escalated, in which case the output is reasonable on its face but the decision it informs needs a second human in the room. The art is matching the verdict to the defect, because using a reject-grade output to make a decision and escalating a routine output both cost you, just in different currencies.

The Five Reject Flags

Reject means the output is unreliable and no amount of light editing rescues it. Five patterns earn that verdict. The first is multiple hallucinations. A single invented fact you catch and correct is an editing task; two or more fabricated facts in one assessment tell you the model was confabulating rather than reading, and you can no longer trust the claims you did not independently check. The second is a major omission of a dimension that matters to the role. If a screen for a senior engineer never touches technical depth, or a leadership assessment never addresses how the person manages people, the output is not wrong so much as hollow, and a hollow assessment quietly steers you toward whatever it did cover.

The third reject flag is obvious, unfixable bias. Dana's "prioritize family" line is the textbook case: the model encoded an assumption about caregiving from a name and a gap, and you cannot edit your way out of a recommendation built on that foundation. Discriminatory inferences about parenthood, age, national origin, or disability are not stylistic problems to soften; they are reasons to discard the output and tighten the prompt. The fourth is a fundamental misunderstanding of the role. If the model penalizes a backend engineer for thin customer-facing experience, every downstream judgment rests on the wrong standard. The fifth is major claims with zero evidence. "Strong leader, great culture fit, high potential" with nothing cited is not an assessment, it is a horoscope, and acting on it means acting on the model's confidence rather than on anything real.

Bias arrives in more disguises than the caregiving inference, and it helps to have the family of them in mind. "Does not have a top-tier degree, so there are likely gaps" is educational-pedigree bias dressed up as an inference about skill. "Seems like someone who might be looking to coast at this stage" is age and career-stage bias. Each of these reads as a professional judgment and each is a conclusion the model reached from a signal that has no business in a screen. Once you can hear that pattern, you catch it in outputs that look far more polished than Dana's Tuesday assessment.

Each reject flag also carries a specific corrective move, and the move is what turns rejection into improvement rather than frustration. Multiple hallucinations mean reject and start over. A major omission means reject or completely rewrite the assessment, because patching a hole leaves you with a document you cannot vouch for. Obvious bias means reject and then build a constraint into the prompt so the same inference cannot recur. A misunderstanding of the role means reject and clarify the role requirements in the prompt, since the model was working from the wrong standard. Unsupported claims mean reject and require evidence in the prompt, instructing the model to cite the specific source line for every judgment it makes.

The Five Escalation Flags

Escalation is different in kind from rejection. The output is plausible and probably accurate, but the decision it feeds touches judgment, fairness, or accountability in ways one person reviewing alone should not absorb. Five situations call for it. The first is contradictory signals: the model rates someone strong on technical skill and simultaneously flags them as a poor fit for the team, and resolving that tension requires human context the model does not have. The second is any cultural-fit assessment. "Seemed quiet, not sure they fit the culture" is the single most common doorway for bias in screening, and culture judgments should be made by people who can interrogate their own assumptions, never accepted from a model as a verdict.

The third escalation flag is the genuinely borderline candidate, the one the model itself cannot place cleanly between strong and weak. That uncertainty is a signal for discussion, not a coin flip. The fourth is the candidate who is clearly qualified but different, an unconventional background, a nonlinear career path, an unusual mix of skills. Different is exactly where pattern-matching systems and busy humans both screen people out for the wrong reasons, so a second set of eyes protects against quietly discarding a strong nontraditional hire. The fifth is any decision you can predict will be questioned later, a departure from your usual profile or a candidate whose rejection might be challenged. Escalating early means the decision is defensible because more than one person stood behind it, which matters both for hiring quality and for the disparate-impact scrutiny that the EEOC applies to any selection procedure, AI-assisted or not.

The Decision Tree

In practice Dana runs every AI assessment through a fixed sequence, and the order matters because reject checks come before escalate checks. First she asks the five reject questions: multiple hallucinations, a major omission, obvious unfixable bias, a misunderstanding of the role, or major claims without evidence. A yes to any one of those ends the review immediately, because there is no point escalating output you already know is unreliable. Only if the assessment clears all five reject gates does she move to the five escalation questions: contradictory signals, a cultural-fit judgment, a borderline candidate, qualified-but-different, or a decision likely to be questioned. A yes there routes the case to a colleague or the hiring manager rather than to the trash. If the output survives both passes, it is usable, and she treats it as a well-organized input to her own decision rather than a decision in itself. The sequence takes well under a minute once it becomes habit, and its real value is that it makes the call systematic instead of dependent on whether Dana happens to be paying close attention that afternoon.

A Worked Example: Catching the Defect

Return to the backend engineer. The AI assessment Dana received read, in part: "Maria brings six years of distributed-systems work and led the migration to a microservices architecture at her last company. Strong technical depth. That said, given the two-year gap on her resume she may want to prioritize family over the team's on-call rotation, which could affect reliability." The numbers here are illustrative of how the review plays out, not a benchmark from any study. Dana ran the decision tree. Reject gate one, hallucinations: the resume said nothing about family, caregiving, or on-call preferences, so the model had fabricated a motivation from a name and a gap. That alone is a fabricated fact. Reject gate three, unfixable bias: the fabrication was specifically a caregiving assumption, the exact territory Title VII and the ADA treat as protected, because gaps correlate with parental leave, illness, and disability. Two reject flags on one assessment. Dana did not escalate it and she did not edit the offending sentence out and keep the rest. She rejected the whole output, because a model that invents a discriminatory concern in one paragraph has already shown she cannot trust its other judgments without re-verifying each one, at which point she has done the work herself anyway.

What she did next is the part that separates rejecting from sulking. She rewrote the prompt to forbid exactly this failure: evaluate only against the stated technical requirements, treat any resume gap as neutral unless the candidate explains it, never infer motivation, family status, or availability from a name or a date, and cite a specific resume line for every claim. She reran it. The second assessment dropped the family sentence entirely, noted the two-year gap as "unexplained, recommend asking about it in the screen," and rated Maria strong on the distributed-systems requirement with a cited example. That version was usable. The gap had become a question to ask rather than a verdict to act on, which is what a sound screen produces. One more step closed the loop: because the first run had produced a biased inference, Dana flagged the prompt template for her own future reference and ran a quick check that her strong-fit rate across the candidates she could estimate had not skewed in a way the four-fifths rule would catch, the same disparate-impact screen she would apply to any selection step.

Three Anti-Patterns

Three habits undo the whole framework. The first is accepting unreliable output because redoing it feels slower than living with it. It is not slower in any sense that matters, because a decision built on a hallucinated fact costs far more to unwind than a second prompt costs to run. If an output trips a reject flag, reject it. The second is failing to escalate decisions that genuinely need discussion, deciding alone on cultural fit or a borderline call because you are the one at the keyboard. That habit concentrates judgment that should be shared and removes the second perspective that catches bias. The third is the mirror image: escalating everything. If every routine assessment goes to a committee, you have reintroduced all the delay AI was supposed to remove and trained your colleagues to tune out your escalations. Escalate the decisions that actually need human judgment and let the clean ones through. The discipline is knowing the difference, which is exactly what the decision tree is for.

Building It Into Your Process

A decision framework that lives only in your head degrades the moment you are busy, which is precisely when you most need it. Dana wrote hers down: a short reject checklist and a short escalation checklist taped, in effect, to every assessment review, plus a one-line answer to what happens after each verdict. When output is rejected, the prompt gets fixed and the run repeated, and a recurring failure mode gets logged so the team's prompt templates improve. When output is escalated, it goes to a named person, the hiring manager for fit questions or a peer recruiter for borderline calls, with the contradiction or concern spelled out so the conversation starts in the right place. Making the criteria explicit also makes them teachable. A new recruiter inherits the same reject-or-escalate logic instead of relearning it through their own near-misses, and the team applies a consistent standard rather than ten private ones, which is itself a fairness safeguard.

Practice

Use assessments from your own live requisitions for these, because the calls only get faster when you have made them on candidates you actually care about.

• Identify the reject flags. Pull three recent AI outputs and run each one against the five reject questions. Note every flag you find and, for each, the specific line that triggered it.
• Identify the escalation flags. Take the same three outputs through the five escalation questions. For each flag, decide whether you would genuinely bring it to a colleague, and say why or why not.
• Practise the decision tree. Take five AI-generated candidate assessments and run each through the full sequence. Record the verdict for each: reject, escalate, or use.
• Build your red flag response process. For your team, define what happens after each verdict. When someone rejects an output, who do they talk to and who fixes the prompt? When someone escalates, what happens next and how quickly? Write the process down.
• Write your own criteria. Based on your roles and your team, draft the reject criteria and the escalation criteria specific to your context. The generic five of each are a starting point, not a finished policy.

Terms Worth Knowing

• Reject. The decision not to use an AI output at all, because it is unreliable or contains major errors.
• Escalate. The decision to bring an AI output to human judgment because the call needs discussion or carries ethical and fairness implications.
• Reject flags. The signs that output is unreliable: hallucinations, bias, omissions, a misread of the role, and unsupported claims.
• Escalation flags. The signs that output needs human judgment: contradictions, cultural-fit assessments, borderline candidates, qualified-but-different profiles, and decisions likely to be challenged.

Putting It to Work This Week

Knowing when to reject and when to escalate is what protects your hiring quality. It is the mechanism that keeps AI in the role of a tool supporting your decision-making rather than one quietly making decisions for you, and it is the difference between a recruiter who uses AI confidently and one who either over-trusts it or abandons it after one bad experience.

So implement the decision tree this week. Run it on five to ten candidate assessments, and keep a note of what you rejected, what you escalated, and what you used as is. The pattern in those notes is the most useful thing you will produce: it tells you which failure modes your prompts keep generating and which decisions your team should be making together. Refine your criteria from those actual decisions rather than from theory.

Reflection

• In your recent hiring, when did you override an AI assessment with your own judgment? Were those good calls, and how would the decision tree have changed the way you reached them?
• What is your biggest concern about relying on AI output? Is that concern a reject flag or an escalation flag, and does naming it that way change what you would do about it?
• How would you explain to your hiring team, in two minutes, when to reject an output and when to escalate it? If you cannot do it in two minutes, the criteria are not yet written clearly enough.

Related Lessons

The reject flags assume you can recognize a defect when you see one, which is what Common AI Errors in Recruiting: Hallucinations, Misinterpretations, Omissions teaches, and Verifying Candidate Information: Spotting Hallucinations and Inaccuracies gives you the verification habits that turn a suspicion into a confirmed fabrication. Avoiding Automation Bias: Staying Active and Skeptical addresses the underlying reason recruiters accept reject-grade output in the first place, which is the pull of a confident-sounding machine.

On the escalation side, Decision Rules: Explicit Criteria for Escalation turns the five flags here into written team rules, and Escalation Paths: When to Involve Legal, Compliance, DEI, or Leadership answers the question this lesson deliberately leaves open: who exactly receives the escalation once you have made the call. Because a framework only protects an organization when everyone applies it the same way, Team Agreements: Building a Culture of Responsible AI Use is where this decision tree scales from your desk to the whole team.

Treat the tree as your safety mechanism and use it consistently. It is what stands between an unreliable output and a hiring decision you would struggle to defend.

Key Takeaways

• Use three verdicts, not two. AI output can be fixed, rejected, or escalated. Treating review as pass-or-fail leads to keeping unreliable output and to escalating routine work, and both have real costs.
• Reject on any of five flags. Multiple hallucinations, a major omission of a role-relevant dimension, obvious unfixable bias, a fundamental misunderstanding of the role, or major claims with no evidence. Any one means the output is unreliable; do not edit around it.
• Every reject has a repair. Hallucinations mean start over, omissions mean rewrite, bias means add a constraint, a misread role means clarify the requirements, and unsupported claims mean require cited evidence. Rejection without a prompt fix produces the same output again.
• Escalate on any of five flags. Contradictory signals, a cultural-fit judgment, a borderline candidate, a qualified-but-different candidate, or a decision likely to be questioned later. These are plausible outputs whose decisions need a second human.
• Run reject checks before escalate checks. There is no point escalating output you already know is unreliable. Clear all five reject gates first, then test the escalation gates, then treat what survives as an input you decide from.
• Bias is a reject, not an edit. An inferred concern about family, age, or disability rests on protected characteristics under Title VII and the ADA. Discard the output, tighten the prompt, and rerun rather than softening the sentence.
• Do not over-escalate. Sending every assessment to a committee reintroduces the delay AI removed and trains colleagues to ignore you. Escalate only decisions that genuinely need human judgment.
• Write the criteria down. A framework in your head fails when you are busy. A documented reject-and-escalate checklist with a named next step makes the call systematic, teachable, and consistent across the team.