When AI-optimized resumes converge into a uniform blur, the old signals stop sorting candidates. These nine signals still hold because they reward lived specificity and judgment that AI assistance can’t shortcut. Use them to rebuild what your filters lost. The full framework lives in the AI resume screening pillar guide.

Quick Comparison

1. Specific Decision Narratives

A candidate who describes a real decision in concrete terms — the constraint, the choice, the result — gives you signal AI text rarely matches. “We had 40% I-9 drop-off, moved the step, cut it to 8%” beats any bullet list. Watch what happens when you push on that sentence. You ask why the drop-off was happening, and a candidate who lived it says the form sat behind a login wall that new hires hadn’t set up yet. You ask why moving the step worked, and they tell you it landed before the credential requirement instead of after. Each answer produces more specificity, not less, because the detail is recalled rather than invented. The mechanism is the cost of texture: a real decision carries a web of surrounding facts the candidate can keep pulling on, while a generated narrative is a flat surface with nothing behind it. Push on the surface and it tears. This is why the narrative outranks every credential on a resume: a credential is a claim you have to verify elsewhere, while a decision narrative verifies itself the moment you ask the second question, and it does so in the candidate’s own voice rather than a recruiter’s reconstruction.

• Ask for one decision and its outcome.
• Probe with “why that over the alternative?”
• Watch for lived specificity vs generic fluency.
• Notice whether detail deepens or thins as you keep asking.

Verdict: The most trustworthy signal you have.

2. Reasoning Under Ambiguity

Judgment on a problem with no clean answer reveals how a candidate thinks. AI can’t shortcut it because there’s no key to steal. See the behavioral questions AI can’t coach. Hand a candidate a genuine dilemma — a key hire wants to leave the week before a launch, and you can either counteroffer hard or let them go and absorb the risk. There is no answer sheet to retrieve, so a chatbot produces a balanced-sounding essay that names every consideration and commits to none. A strong candidate does the opposite: they pick a direction and tell you what they are willing to lose to get it. The mechanism is the absence of a target. Gameable signals all share a fixed right answer that a tool can reverse-engineer, and ambiguity removes the target entirely, which is why it stays trustworthy while keyword matches and test scores collapse.

• Present competing priorities and missing facts.
• Score the tradeoff they name.
• Accept multiple defensible conclusions.
• Penalize answers that refuse to commit to a direction.

Verdict: High trust; the core of AI-resistant screening.

3. Follow-Up Resilience

Fabrication breaks down under a second and third question. A candidate who lived the work answers fluently, and one who borrowed it stalls. The tell is timing. A first answer can be rehearsed and delivered smoothly, so the first answer tells you almost nothing. The second and third questions go where no script anticipated, and that is where the gap shows: the candidate who did the work keeps the same conversational pace, while the one who memorized a story pauses, hedges, and starts speaking in generalities. The mechanism is that rehearsal covers a fixed surface area. You can prepare an answer, but you cannot prepare every branch of every follow-up, so depth of questioning eventually exits the rehearsed zone. A live conversation is your only tool that can do this, which is why it outranks every artifact on this list.

• Always ask at least two follow-ups.
• Listen for detail that can’t be guessed.
• Note where fluency suddenly drops.
• Treat a sudden shift to generalities as the strongest signal of borrowed experience.

Verdict: A live conversation is your best detector.

4. Work-Sample Output

A short, role-relevant sample shows real ability the resume can’t. Score the reasoning behind it, not the surface polish. Give an analyst candidate a messy dataset and a single question, and ask for the answer plus the reasoning. One candidate returns a clean number with no defense. Another returns the same number, flags that two rows looked like duplicates, and explains the assumption they made to handle them. The second candidate is the hire, and no resume on earth would have told you that. The mechanism is that a sample asks for the work rather than a description of the work, and judgment leaves fingerprints — the caveat, the assumption named out loud, the edge case spotted — that polished prose about skills never contains. There is a second-order benefit worth naming: a small sample reveals how a candidate behaves when the data fights back, which is the daily reality of the job and the exact thing a tidy resume hides. The applicant who pauses to question a suspicious row is showing you a habit you cannot interview into someone. Keep the sample small so a strong candidate spends twenty minutes, not an evening, route it to a human rather than a parser, and score the thinking the sample exposes rather than the formatting it arrives in.

• Keep it short and realistic.
• Judge the thinking, not the formatting.
• Reward a candidate who names their assumptions over one who hides them.

Verdict: Strong when the prompt demands judgment.

5. Structured Phone Screen Performance

Fifteen structured minutes reveal more than any filter. Run the same questions for everyone and score against a rubric. The structure is what converts a pleasant conversation into a comparable signal. If you ask one candidate about teamwork and the next about deadlines, you have two impressions you cannot line up against each other, and impressions are where bias hides. Ask all three candidates the identical decision question and score each on the same rubric, and now you have a column of numbers that means something. The mechanism is holding the input constant so the output becomes a measurement. The cost worry — fifteen minutes times every candidate — dissolves once you automate the scheduling and reminders, leaving the recruiter with only the part that requires a human.

• Use a fixed three-question script.
• Score comparably across candidates.
• Automate the scheduling so the only human cost is the conversation.

Verdict: High trust; cheap to run with automated logistics.

6. Consistency Across Stages

A candidate whose written answers, phone screen, and interview tell the same story is giving you a stable signal. Sudden capability jumps between a polished application and a flat conversation are a flag.

• Compare claimed vs demonstrated ability.
• Flag large application-to-interview gaps.

Verdict: Useful cross-check.

7. Resume Keyword Match (Low Trust)

Keyword presence no longer separates qualified from unqualified. It’s free to fabricate.

• Stop using it as a competency gate.
• Keep it only for hard, verifiable requirements.

Verdict: Low trust. Demote it.

8. Assessment Correctness Score (Low Trust)

Fixed-answer scores are solvable with a tool in another tab. A perfect score is now a yellow flag.

• Stop reading scores as a clean ranking.
• Redesign for judgment instead.

Verdict: Low trust without redesign.

9. Honest Uncertainty

A candidate who names what they don’t know, and reasons anyway, signals integrity and real understanding. Reverse-engineered answers rarely admit doubt. Listen for the candidate who says “I’d want to check the actual churn numbers before committing, but my instinct is to fix onboarding first, and here’s why.” That sentence does two things at once: it admits a gap and still commits to a direction. A borrowed answer almost never does this, because confident completeness is what generated text is tuned to produce, and admitting a limit reads to a model as a weaker answer. The mechanism is that real expertise knows its own edges. People who have actually done hard work know where their knowledge stops, and they say so. Flawless certainty across the board is the tell that the candidate is performing knowledge rather than holding it.

• Reward defensible uncertainty over false precision.
• Treat flawless confidence with mild suspicion.
• Score the candidate who names a limit and still reaches a decision.

Verdict: Underrated; high trust.

How We Evaluated

Each signal was rated on whether AI assistance can produce it without real experience. Signals that survive follow-up and demand lived specificity scored high. For the audit that proves which signals predict your hires, see how to audit your screening-to-hire correlation and the pillar.