AI Resume Screeners Agree 14% of the Time. That's Worse Than Random.
AI resume screeners produced 14% shortlist overlap on identical data, worse than a coin flip. Here's why sourcing on verifiable signal beats the paper layer.
Recruiterflow's mid-July 2026 analysis, citing recruitment veteran Greg Savage, reported something the screening-tool vendors would rather you not internalize: when the same AI resume screener was run twice on the same candidates, the two shortlists agreed only 14% of the time. If you paid for an AI screening layer this year, this is the number that should be on your dashboard.
The rest of the pipeline is not helping. Roughly 40% of tech candidates are believed to have materially inflated their resumes, 38.5% of live interviews now trip AI-cheating flags, and Google, McKinsey, Deloitte, and Cisco have all reintroduced mandatory in-person rounds. The paper layer is finished. The question is what replaces it.
What the 14% number actually means
The 14% figure is the shortlist overlap when the same AI resume screener was rerun on identical candidate data. It is not a benchmark against humans. It is the tool disagreeing with itself.
Greg Savage's test, cited in Recruiterflow's July 2026 recruitment-trends piece, worked like this: multiple AI screening tools were each asked to produce a shortlist of five from the same candidate pool. Between tools, overlap was 14%. Running the same tool again the next day produced a different five. In several runs, the strongest candidates were missing entirely.
Here is why that is worse than it sounds. If you draw two independent shortlists of five from any reasonably sized pool, chance overlap is around 20%. A screener that produces 14% overlap with itself is not noisy. It is anti-correlated with itself, which means it is pattern-matching on unstable features (resume formatting, keyword density, LLM perplexity artifacts) rather than anything durable about the candidate.
The mechanism is not mysterious. Modern screeners score embeddings of resume text. AI-written resumes now dominate the input distribution. So the model is largely ranking prompt-engineering quality, and prompt-engineering quality is not correlated with ability to ship.
The numbers behind the collapse
The screening layer is breaking at the exact moment adoption is peaking, which is the worst possible combination. Here are the figures worth memorizing before your next vendor call.
| Metric | Value | Source |
|---|---|---|
| Shortlist overlap, same AI screener rerun on identical data | 14% | Greg Savage via Recruiterflow, Jul 2026 |
| Random 5-of-5 chance-overlap baseline | ~20% | Combinatorial baseline |
| Tech candidates with materially inflated resumes | ~40% | Recruiterflow / The Savage Truth |
| Live interviews flagged for AI-cheating (Jul 2025 to Jan 2026) | 38.5% of 19,368 | Fabric AI |
| Hiring managers who suspected AI-generated interview answers | 91% | Greenhouse, 2026 |
| Untrained human deepfake detection accuracy | 55.54% | Wiley meta-analysis |
| Companies planning to use AI resume review in 2026 | 83% | Resume Builder |
| Adoption-to-consistency ratio | ~5.9x | Derived (83% / 14%) |
That last row is the punchline. Nearly six times as many companies are adopting AI resume review as the tool agrees with itself. The market is not pricing the failure mode.
Resume inflation and AI screeners are on the same curve
Candidates use LLMs to write resumes. Employers use LLMs to read them. The humans in the loop are spectators, and the paper artifact in the middle has lost most of its information content.
Look at the pincer:
- 83% of companies plan to use AI to review resumes in 2026 (Resume Builder).
- Roughly 40% of tech candidates have inflated resumes (Recruiterflow).
- 38.5% of live interviews now trip AI-cheating detection (Fabric AI, across 19,368 sessions between July 2025 and January 2026, a rate that tripled in three months).
- 91% of US hiring managers report encountering or suspecting AI-generated interview answers (Greenhouse, 2026).
- Palo Alto Networks researchers showed someone with no image-manipulation background can build a fake candidate that passes a video interview in roughly 70 minutes.
- Gartner projects one in four job candidates will be fake by 2028.
Candidates use AI to write, employers use AI to read, and the humans in the loop are watching two language models negotiate.