The Candidate Authenticity Score: Real, AI Persona, or Fabricated

Before you book screening time, you have to answer one question about a promising sourced or inbound candidate: is this a genuine person, or an AI-generated or fabricated identity? This guide gives in-house recruiters, sourcers, and founders doing their own hiring a fixed rubric to score a candidate's public footprint across LinkedIn, GitHub, and the open web at once, then route them to book, verify-further, or reject with a written reason two reviewers would agree on. It is built to be used at the gate, before you spend interview time, and it carries explicit guards so real career-changers and privacy-conscious candidates are not wrongly cut.

This is a different job from two adjacent standards. It is not the operations dedup teardown that asks whether one record is a single human, and it is not the pre-offer authenticity standard for remote engineers at offer stage. This scores fraud risk earlier, at sourcing and screening, across all three surfaces together.

Why this gate exists now

Fake candidates are no longer an edge case, and a passed background check does not close the question. Gartner predicts that by 2028 one in four candidate profiles worldwide will be fake, which turns authenticity from a rare escalation into a standing step in the funnel.

The prevalence numbers disagree on magnitude, and the disagreement is the point. Show both ends rather than averaging them, because the true rate depends on role, channel, and geography.

The Pindrop row is the most concrete: of 827 applications for one back-end developer role, roughly 100 came in on fake identities. Palo Alto Networks says a fake candidate can be assembled in as little as 70 minutes, which is why volume roles see clusters rather than one-offs. And the reason this gate is needed at all: SIA and Sterling research found 63% wrongly believed all background checks include identity verification. They do not. A passed-check note in your applicant tracking system is not evidence the person is who they claim.

Why no single check can carry the decision

Both the human eye and the automated detector sit near a coin flip on hard cases, so a defensible verdict only emerges from stacking independent signals. This is the load-bearing principle of the whole rubric: it is additive, never one gate.

Read the table as a warning against confidence. A hiring manager who feels sure a video call is fake is right about 56% of the time. A detector that flags a headshot as AI-generated is right maybe 50 to 60% of the time on a generator it has not seen. Gartner expects that by 2026, 30% of enterprises will stop treating identity verification and authentication solutions as reliable on their own. The rubric answers this by requiring several weak signals to agree before any reject.

The five scoring dimensions and what each proves

Score every candidate on the same five dimensions. Each one states what a clean result proves and what it looks like when it lies, so the score means the same thing across reviewers.

• Footprint age and depth. A multi-year, consistent presence is hard to fabricate at depth. It proves the identity predates the application. It lies when a farmed account fills a contribution graph to a goal, or when a real privacy-conscious candidate simply has a six-month-old LinkedIn.
• Cross-platform match. The same person, same face, same work history across LinkedIn, GitHub, and the open web is a strong positive. It proves independent surfaces corroborate. It lies when a fabricator builds all three from one template, so check that the surfaces reference the same specific projects, not just the same name.
• Contribution depth. For engineers, per-author commit counts and pre-creation-date work prove authored labor. Being listed as a contributor could mean 2,000 commits or one. The contributor list alone will not tell you which, and green squares are fabricable to an arbitrary goal.
• Photo authenticity. Specific artifacts around ears, teeth, jewelry, hair boundary, and background raise suspicion. This proves little on its own given weak detectors; it lies in both directions, flagging real people who used AI headshot tools and clearing polished fakes.
• Logistics and contact consistency. Contact details that recur across other identities, or a location that mismatches network and payment signals, are the most durable tell. This proves operational reuse, which the fraud economics require. It rarely produces a false positive because innocent candidates do not share a VoIP number with three other applicants.

Notice the inversion. The signals people reach for first, the headshot and the video call, are the weakest and the most bias-prone. The signals that actually decide cases sit at the bottom of the stack and take longer to gather. Weight your rubric accordingly.

The eight-step scoring procedure

Run these in order. The first five gather and grade signals; the last three route, test, and escalate. Times assume a single candidate at the gate, not a batch.

How the depth check actually works

Step three is where engineering candidates separate. GitHub join dates cannot be altered, but fixing commit history makes the contribution graph reflect any activity period you like, and public tools exist to fill a graph to an arbitrary goal. So the graph is theater. What resists fabrication is per-author commit counts, which GitHub does not summarize for you, and the gap between account creation and first commit. Farmed accounts tell on themselves: they host no original code, lots of tutorial-level code, an overcompensating README, and no commits before the creation date. The only reliable proof of ownership is GitHub OAuth, which gives cryptographic proof rather than a claim.

Market context changes what an empty GitHub means. Weigh it before you flag.

Both counts come from Refolk's index. The India pool is about 1.72x the US pool, which means a "no GitHub footprint" flag is a weaker signal in some markets than others. A flat global rule over-flags by geography, so let the depth check inform the score, not dominate it.

Routing: turning a score into an action

Map the rubric score to one of three actions, and log the reason with every routing call. The point of the bands is that two reviewers reading the same signal sheet reach the same action.

• Book. Consistent multi-year footprint, cross-platform match on specific projects, and clean contact consistency. Nothing suspicious in the depth check. Book screening time and move on.
• Verify-further. A thin footprint with one innocent explanation, or a single artifact with no corroborating signal, or a market where absent GitHub is expected. Do not reject. Add friction at interview and confirm one hard artifact.
• Reject. Multiple independent red flags stacking: reused contact details across identities, a fabricated-looking depth profile, and a location or logistics mismatch, with a written reason. Never a single-signal reject.

The bottom-right and top-left quadrants are the traps. A rich footprint with stacking fraud signals is a fabricator who invested effort, so it escalates rather than passes. A strong-looking signal on a thin footprint is usually a data gap, not fraud, so it verifies rather than rejects.

Gathering the corroboration in step two is the slow part, because it means finding the same person across LinkedIn, GitHub, and the open web and checking that the surfaces reference the same specific work. This is exactly the retrieval Refolk is built for: you can ask for the corroborated footprint directly instead of stitching it together tab by tab.

How this goes wrong: failure modes and false positives

This is the most valuable part of the standard, because a rubric that over-flags is worse than no rubric. Every failure mode below has a false positive that cuts a real person, and a check that prevents it.

• Photo detector over-trust. A confident AI-generated score is often wrong on new generators, which score 50-60% accuracy. The false positive is a real person who used an AI headshot tool, now common on LinkedIn. Check: require a second, differently-lit photo or a live camera moment, not the detector verdict.
• Thin footprint read as fraud. A privacy-conscious candidate or recent immigrant looks identical to a fabricated one. The false positive is a genuine career-changer with a six-month-old LinkedIn. Check: corroborate one hard artifact such as verified employment or an OAuth-confirmed GitHub, rather than counting connections.
• Contribution-graph theater. Green squares are fabricable to a goal, so a farmed account looks productive while a real developer on private or GitLab repos looks empty. Check: per-author commit counts and pre-creation-date commits, not the graph.
• Name-based suspicion. Non-Western names and unfamiliar schools trigger a "doesn't line up" read that is bias, not signal. Resumes with Western names get 50% more callbacks at identical qualifications. Check: never let name or school contribute to the score, and audit rejections for name clustering.
• The hand or turn test as proof. Modern face-swaps pass occlusion, so passing the test gives false confidence, and failing it can just be a bad webcam. Check: use unscripted knowledge probes plus logistics consistency, not one gesture.
• Confusing AI-assisted with fraudulent. A polished, AI-written resume is not a fabricated identity. Nine in ten HR workers reported a surge in low-effort AI-generated applications, but a strong candidate who used a chatbot to write bullets is not a fraud. Check: score identity authenticity separately from application polish.
• Reused-contact miss. A VoIP number or email recurring across different applicants is a strong tell that is easy to overlook at volume. Check: deduplicate contact fields across the whole pipeline, not per candidate.
• Single-reviewer drift. Human-only judgment is inconsistent and bias-prone. Check: require a written reason and a second-reviewer agreement threshold before any reject.

The signals people reach for first, the headshot and the video call, are the weakest and the most bias-prone.

The live authenticity test, and where sources disagree

Only run the live test at interview, and only for verify-further candidates. It adds friction that a fabricated identity struggles with in real time: unscripted questions with no clean answer to prepare, a request to change camera angle, and ID shown on camera. Document anomalies without confronting the candidate mid-call, because a false accusation on a real person is a legal and reputational cost.

Here the evidence splits, and you should know it. Some vendors still promote the wave-your-hand and turn-your-head tests as reliable deepfake detection. Gem and others report the opposite: the simple tests are already obsolete because modern face-swap technology handles the gesture with minimal artifacts. Treat any gesture result as one weak input, never as proof. The durable substitutes are unscripted knowledge probes, watching for lip movement out of sync with speech, absent or unnatural blinking, artifacts at the edges of the face, and delays in expression reactions, always weighed against the possibility of a slow connection or an old webcam.

Candidate: [candidate ID, not full name if pipeline is anonymized]
Rubric score: [number] / band: verify-further
Footprint age/depth: [real / thin / suspicious] - [one specific observation]
Cross-platform match: [corroborated on projects X, Y / inconsistent because ...]
Contribution depth: [per-author commit finding, or N/A for non-engineering role]
Photo: [artifact noted, or clean; detector score advisory only]
Contact/logistics: [phone and email unique to pipeline: yes/no]
Reason to verify rather than book: [one line]
Live test to run at interview: [specific unscripted probe]
Second reviewer: [name] agrees: [yes/no]

What to verify before you call the score done

Run this before logging any routing decision. It is the difference between a score two reviewers agree on and one reviewer's gut.

Keeping the rubric current

The rubric is stable but its inputs decay, so re-check the two moving parts on a schedule rather than trusting yesterday's calibration. First, detector reliability drifts as generators improve; the 50-60% figure on out-of-distribution images tells you to keep detectors advisory permanently, not to chase a better one. Second, prevalence shifts by role and channel, so if a single posting starts returning a Pindrop-style cluster of fake identities, tighten the verify-further band for that channel and log the change with its business reason.

Two controls keep the gate honest over time. Re-review a random sample of low-ranked and rejected candidates to see whether strong profiles are being missed, and document every criteria change and the reason for it. That audit is what protects real career-changers and privacy-conscious candidates from a rubric that quietly drifts toward over-flagging. It is also what lets you defend the gate: in Refolk's index there are 91,813 US-based recruiting and talent-acquisition professionals who apply judgment like this, and the ones whose decisions hold up are the ones who wrote down the reason and had a second reader agree.

The last thing to keep current is your own confidence. Humans read deepfake video at 55.54% and the best open detectors barely clear a coin flip, so the moment the rubric feels like it is giving you certainty from one glance, it has stopped working. Certainty comes from signals agreeing, from a reused number surfacing in a dedup, from a depth check that contradicts a claimed prominence. Stack the evidence, write the reason, and let a second reviewer break the tie.