TL;DR: Keyword-based resume screening worked for two decades because writing a resume that matched a job description perfectly took effort, and effort correlated with interest. Generative tools removed the effort and the correlation with it. Across 34 high-scoring threads, recruiters describe match scores that no longer discriminate, applications that read identically, and — at the far end — candidates who are not the person doing the interview.
The evidence
PainHunt holds 34 posts scoring 10 or higher out of 15 on this pattern, averaging 11.4/15 with a pain intensity of 7.7/10 and a willingness-to-pay signal of 7.3. Sources are Medium 10, Mastodon 9, Reddit 6, Hacker News 6, with single posts from Dev.to and Substack. It is a smaller cluster than most we publish on, and the honest reading is that this is an early signal with unusually consistent content rather than a large market already shouting.
What recurs, in order of how directly it hits the employer.
Match scores stopped discriminating. Threads describe AI-optimized resumes flooding applicant tracking systems until "100% match" scores are meaningless and candidates are indistinguishable from one another. The score still computes. It just no longer sorts.
The filter fails in both directions at once. The same discussions report qualified candidates being filtered out over minor technology-stack mismatches while AI-generated applications pass through. A system that rejects real people for surface differences and admits synthetic ones for surface similarities is not badly calibrated — it is reading a signal that has been destroyed.
Applications converge on one voice. Recruiters describe modern CVs increasingly resembling one another: different labels, same content. When every applicant writes with the same assistant, differentiation moves out of the document entirely.
Screening cost moved downstream. Because the filter no longer removes weak fits, the work lands in interviews. Threads describe recruiters spending significant time interviewing people who passed initial screens and cannot perform the actual work.
At the edge, identity itself is in question. A smaller set of threads describes tools that assist candidates live during technical interviews, proxy candidates sitting interviews on someone else's behalf, and impersonation using synthetic video. These are few in number and should be read as an emerging pattern, not a measured rate.
Why now
The cost asymmetry flipped this year, not gradually. A candidate can now produce fifty tailored applications in the time one used to take. Employer-side screening capacity did not change at all. Any filter whose implicit assumption was "applying is costly" is now calibrated against a world that ended.
Volume is arriving at teams without recruiting infrastructure. Small companies posting one role receive application counts they have no process for, and the tools sold to them are the same keyword filters that stopped working.
The verification layer was always implicit. Nobody built candidate identity verification because the interview itself performed that function. Once an interview can be attended by someone else, or coached in real time, the function has to become explicit — and there is no established product for it.
The wedge
Building a better ATS is not the opportunity. The threads come from people who already have one and cannot switch. The opening is a layer that supplies the signal the ATS no longer has.
- Rank on evidence that is expensive to fabricate. Verified employment history, public artifacts with commit or publication history, references that respond. Text is free to generate; provenance is not. This is the whole thesis in one line.
- Score fit against work performed, not words matched. The recurring example in the threads is a candidate who "led migration of legacy reporting infrastructure" being invisible to a search for the job description's phrasing. Matching descriptions of work to requirements is a genuinely different computation than matching keywords, and it fails in the opposite direction — toward including near-matches rather than excluding them.
- Give the hiring manager a summary with its sources attached. Threads specifically describe managers receiving raw resume data when they need a short, reviewable assessment. Make every claim in the summary clickable back to what it came from, because an unsourced AI summary is the same trust problem one layer up.
- Sell interview integrity as a separate, opt-in product. Identity checks and live-assistance detection are a different purchase with different sensitivities. Keep them modular; a screening tool that ships surveillance by default will lose deals over the surveillance.
- Start with roles where artifacts exist. Engineering, design, and writing all leave verifiable public traces. Land there, prove the ranking beats keyword search, and expand into roles where evidence is harder to gather.
Risks and honest caveats
- This dataset is small. Thirty-four threads is a signal, not a market. Every sample is on-topic and the pattern is coherent, which is why it is worth writing about — but anyone building here should confirm demand with recruiters directly before committing a year.
- Bias exposure is severe and asymmetric. Automated candidate ranking is regulated in a growing number of jurisdictions, with audit requirements attached. "We rank on verified evidence" is a defensible position, but only if you can actually produce the audit trail, and building that is most of the work.
- Verification is adversarial and gets harder. Any identity check you ship becomes a target the week after you ship it. This is a permanent operating cost, not a feature you complete.
- ATS vendors will absorb the obvious version. Semantic matching is already appearing as a feature. The durable part is the verification and provenance layer, which requires data partnerships and continuous adversarial work rather than a model swap.
- Candidates are not the enemy, and a product that treats them as one will be hated. People are using the tools available in a market that ignores them. A design that punishes candidates rather than surfacing evidence will generate the backlash the threads already show against AI interviews.
How to validate this further
The test that decides the business is whether employers will pay for verification or only for better ranking — the first is a compliance-adjacent sale, the second a productivity one. Use the PainHunt dashboard to read the employer-side threads separately from the candidate-side ones, since they describe the same broken filter and want opposite fixes. Then check whether verified work history alone changes who gets interviewed, using idea validation.
Related reading: ATS resume checker: where the existing tools fall short — the candidate's side of this same filter — and verifying ghost job postings.