Somewhere in the world right now, a candidate is using ChatGPT to write a cover letter for a job whose description was probably also written by ChatGPT. That cover letter will be screened by an applicant tracking system that has been doing statistical resume parsing since the late 1990s — and a recent flavour of LLM since about the time Sam Altman last gave a TED talk. The candidate, if they're savvy, has prompted GPT-5 to "optimize for ATS keywords from this job description." The recruiter, if they're savvy, has prompted GPT-5 to "summarize and rank these 1,400 applications."
We are all, increasingly, doing the same thing: pointing language models at each other and waiting for a job offer to fall out the other end.
If this sounds like a uniquely 2026 problem, it isn't. It's just that the loop got tighter.
"AI in hiring" is older than your AI co-founder
The first commercial applicant tracking systems shipped in the early 1990s — Resumix in 1988, Restrac in 1991. They used statistical text parsing to extract structured fields (name, employer, dates, skills) from unstructured resumes. By the mid-90s these databases held millions of resumes and could rank candidates against a job description in seconds. None of this was called AI at the time. It was called resume parsing. The math wasn't dramatically worse than what's running inside half the AI screening tools today.
What changed isn't the existence of automated screening. What changed is the scale of the input (the often-cited industry figure is that around 75% of applications still never reach a human reviewer), the breadth of the signal (resumes, video, online presence, code samples), and — crucially — the marketing budget.
The "AI revolution in recruiting" is, in part, a rebrand. A useful one, sometimes. But a rebrand.
The dirty secret of every ATS that's ever existed
In 2021, Harvard Business School published a report — Hidden Workers: Untapped Talent — that surveyed over 2,250 executives. Their finding became one of the most-cited statistics in recruiting: 88% of executives agreed that qualified, high-skill candidates were being vetted out by their own hiring processes.
Not unqualified candidates. Qualified ones. Filtered out by the systems the executives themselves had purchased and deployed.
The mechanism is depressingly mundane. Hard keyword matching ("must contain Kubernetes") doesn't capture "led migration of fleet to container orchestration." Hard rules on employment gaps reject anyone who took parental leave, served as a caregiver, or — between 2020 and 2022 — happened to be in the wrong industry at the wrong time. Geographic filters on roles that are supposedly remote-eligible. The list is long, but the punchline is the same: the AI didn't make these mistakes. The AI was trained on humans making them, and then made the same mistakes at machine speed.
In August 2023 the EEOC settled its first major lawsuit involving AI hiring discrimination. iTutorGroup paid $365,000 to settle charges that its hiring software automatically rejected female applicants over 55 and male applicants over 60. The model wasn't sentient. It was just doing exactly what the historical data told it to do.
The candidate countermove
When the gatekeeper is a model, the path of least resistance is to learn what the model is reading for. So candidates did.
Resume keyword stuffing is now a documented genre. There are entire subreddits dedicated to it. The most aggressive version involves typing relevant keywords in white text on a white background — invisible to a human reviewer, fully visible to a parser. (Most modern parsers now strip white text. The arms race continues.)
ChatGPT cover letters are pervasive. Recruiters claim they can spot them; the studies on whether they actually can are inconclusive. What is clearer: when 1,400 candidates apply to a posting and 1,200 of them are running the same prompt against the same job description, the cover letter has stopped being a useful signal for anyone.
We have, collectively, built a system in which the cheapest possible move from both sides of the table is to delegate the entire interaction to a language model. The math says this should converge to gibberish optimized for other gibberish. The empirical results are roughly that.
What actually got better
The cynicism is fun but it isn't the whole story. Some things genuinely improved.
- Time-to-hire compressed. Median time-to-hire for tech roles in 2014 was around 42 days; current industry surveys put it closer to 28 for teams that have actually re-tooled. The compression is real, and it's not all AI — but a non-trivial chunk of it is AI handling first-pass screening that previously waited in a queue for a human.
- Structured screening replaced vibes. A well-designed AI video interview, scored on explicit dimensions (communication, role-specific knowledge, problem decomposition), is more consistent than the 15-minute phone screen it replaced. Note the word consistent — not necessarily better, but consistent. For roles with a thousand applicants, "consistent" is a step up.
- Candidate experience modestly improved. When AI handles screening interviews, every applicant gets one. There's no recruiter triage layer deciding who's worth fifteen minutes. For a candidate who would previously have been ghosted, getting a structured chance — even with a robot — is a meaningful change.
- HireVue dropped facial analysis in 2021 after independent researchers (and a coalition led by AlgorithmWatch) showed the scientific basis for inferring traits from facial micro-expressions was effectively non-existent. This is what progress looks like: not the model getting smarter, but the industry quietly walking back the parts that were never defensible.
What didn't
The systems still encode the bias in their training data, and now they do it at scale. NYC Local Law 144, which took effect July 5, 2023, requires employers using automated employment decision tools to commission bias audits and publish the results. Compliance, in practice, has been spotty — but the framework is the first real regulatory acknowledgement that "the model made the decision" is not a defence.
The "ghost job" phenomenon — postings that exist but will never be filled — has gotten worse, not better. Some of these are pipeline-building exercises. Some are suspected to be training-data harvests, where an employer wants candidate data without ever offering a role. There is no AI fix for this, because it's not a screening problem. It's an incentive problem.
And the loop tightens. By 2026 you can buy a Chrome extension that auto-applies to a thousand jobs a week using GPT-generated cover letters. You can buy a different SaaS that uses GPT to filter those applications back out. Both companies have raised meaningful Series A funding. Both, in the limit, are using the same language model on opposite sides of a transaction. The economic surplus generated by this entire ecosystem is, approximately, the API bill of a mid-sized OpenAI customer.
Where this is heading
The interesting question isn't "will AI replace recruiters." (No. The same way calculators didn't replace accountants.) The interesting question is which part of the recruiting workflow becomes mostly automatic, which part becomes mostly human, and where the line lands.
The emerging consensus, insofar as there is one: the top of the funnel (resume parsing, initial screening, scheduling) is moving fast toward automation. The middle of the funnel (structured interviews, technical assessment) is moving toward AI-augmented — human-in-the-loop, where the AI proposes and the recruiter disposes. The bottom of the funnel (offer negotiation, closing, references) stays human, because the stakes are individual and the social signal matters.
The teams winning right now aren't the ones with the most AI. They're the ones whose AI is bundled into the workflow (so recruiters use it without thinking about per-interview cost), structured (so the output is auditable, not "the model said yes"), and integrated (so the candidate record is the same across every stage).
Which, since you read this far, is the part where we tell you what LeapOne does.
The LeapOne take
LeapOne was built on three opinions about AI in recruiting that fall out of everything above:
- AI screening should be bundled, not metered. If you charge per AI interview, recruiters ration them. If recruiters ration them, the bias problem gets worse, because the candidates who don't get an AI screen get screened by gut feeling instead. Every LeapOne plan includes AI video interviews — 40, 250, or 800 per month, depending on tier — with no per-use fee.
- Scores should be structured and visible. Our AI interview scoring runs on nine explicit dimensions (communication, role knowledge, problem-solving, and so on), and every score is exposed alongside the transcript. "The model liked them" is not a hiring decision. "The candidate scored 4 out of 5 on technical decomposition and 5 out of 5 on communication — here is the transcript" is.
- The AI is a stage, not a verdict. Every AI-screened candidate flows into a normal pipeline where a human recruiter reviews the transcript and scores before any reject decision. The model proposes. The recruiter disposes. It's the only design we've found that delivers the speed benefits without inheriting the EEOC-settlement risk.
We also publish our pricing on a public page, include unlimited recruiter seats in every plan, and don't have a "contact sales" tier for anyone below Enterprise. None of that is AI. It's just the way the rest of the industry already should work.
The robots are reading each other's resumes. They will continue to. Our small bet is that the recruiting teams that win the next decade are the ones who keep a human on at least one end of the loop — and who use software that makes the model's reasoning legible enough to argue with.