Greenhouse's 411% Scissors: Why Inbound Screening Just Structurally Broke

Greenhouse dropped two reports in spring 2026 that talent leaders are reading as separate problems. They aren't. The April "Hire Standard" benchmark and the companion AI in Hiring Report describe one structural break, from two angles, and the honest read is that inbound screening has stopped being a reliable signal at companies hiring engineers.

The headline most outlets ran was "40% of candidates admit to prompt injection." That's the friendly version. The number under it is 41%, and the number behind that, the one nobody is putting on a slide, is 411%.

The scissors nobody wants to graph

Greenhouse analyzed over 6,000 companies and over 640 million applications from 2022 through 2025. Two lines move in opposite directions.

411%

Applications per recruiter are up 411%. Recruiting teams are down 55%. A recruiter who used to see roughly 600 applications a year is now staring at 2,500-plus, with fewer coordinators, fewer sourcers, and the same forty-hour week. That is not a funnel you can "tighten." That is a funnel that has stopped functioning as a funnel.

Sharawn Tipton, Greenhouse's Chief People Officer, walked through these numbers on the RecTech podcast without dressing them up. Daniel Chait, the CEO, has been blunter in print, calling it an "AI doom loop" and telling Fortune, "the hiring process today isn't working, and the more of today's AI people use, the worse it's getting." When the vendor selling you the ATS says the inbound model is broken, that's worth pausing on.

Why prompt injection is the wrong villain

The AI in Hiring Report surveyed more than 4,100 job seekers, recruiters, and hiring managers across the U.S., U.K., Ireland, and Germany. Of the 1,200 U.S. job seekers, 41% admit to using prompt injections or hidden text designed to bypass AI filters. Of the ones who don't, 52% say they're considering it. 91% of recruiters and hiring managers have spotted or suspected candidate deception. 65% of hiring managers have actually caught it, including 22% who pulled a prompt injection out of a resume and 18% who got deepfaked on a video call.

Those numbers are real and they're disturbing. They are also a distraction.

ManpowerGroup detects hidden text in about 10% of scanned resumes. Greenhouse's own data puts white-on-white prompt injection at roughly 1% of all resumes. Engineer Fielding Johnston published a working demo in March 2026: bottom margin of every page, 1-point text, #FAFAFA on #FAFAFA, invisible in the rendered PDF, fully extractable by any text parser. The technique is real and reproducible. It also mostly doesn't work, because most ATS pipelines strip formatting before the resume hits an LLM, exposing the hidden text in plain sight.

Prompt injection is what despair looks like when it learns to use a text editor. </pull> So the gap between "41% admit it" and "1% of resumes carry it" is the actual story. Almost half of U.S. candidates have lost enough faith in inbound that they're willing to sabotage their own resume to get past a robot they assume is between them and a human. They are mostly wrong about the mechanics. They are not wrong about the despair. ## The arms race nobody wins 61% of U.S. hiring managers, and 59% across the U.K., Ireland, and Germany, now run software to detect AI-generated content. 74% say they're more concerned about fake credentials, deepfakes, or misrepresented experience than they were a year ago. Only 21% of U.S. candidates believe most employers are using AI responsibly and transparently. Stack those numbers. Every filter you add to inbound degrades the employer brand of the candidates you most want to attract. Every prompt injection a candidate tries burns recruiter time you no longer have. The detection vendors are selling you the second half of a war the first half already lost. The "AI interview" backlash is the next shoe. 63% of candidates have now been interviewed by an AI, up 13 points in six months. 38% have walked away from a hiring process specifically because it included an AI interview, and another 12% say they would. Paddy Lambros, who runs the AI career-agent startup Dex, called first-round AI interviews "insulting and inhumane" in Fortune. Teams leaning on AI interviewers to absorb the 411% volume are filtering out the candidates they were trying to reach.

refolk prompt: Backend engineers in NYC who've shipped production Rust at a Series B or later, no current FAANG note: A ranked shortlist pulled from GitHub, LinkedIn, and the open web, with the signal you'd normally spend a day verifying already attached. slug: q44sdqgmvh

## The cost-per-hire math has already flipped

Gem's 2026 data, which Refolk wrote about separately, puts outbound candidates at 8x more likely to be hired than inbound. The nonexecutive cost-per-hire average is now $5,475. Just 0.5% of applicants receive offers. If you do the arithmetic on a recruiter handling 2,500 applications per year at a 0.5% offer rate, you get roughly 12 offers from the entire inbound stack, before you adjust for the deception cleanup, the AI-interview dropouts, and the fact that the strongest candidates rarely apply cold to begin with.

That is the case for outbound sourcing in technical recruiting, stated as an income statement rather than a philosophy. You can keep paying recruiters to triage a pile that produces 12 offers, or you can point the same hours at people you've already decided you want.

This is the friction [Refolk](/) was built around. You describe the person in plain English ("staff-level data platform engineer who shipped Kafka at scale, currently at a public company, open to startups"), and you get a ranked shortlist across GitHub, LinkedIn, and the open web. No boolean strings, no scraping, no LinkedIn Recruiter seat math. The point isn't to replace your sourcers. It's to stop pretending inbound is a substitute for them.

### Who's already reallocating

Look at where dedicated technical sourcers actually sit. Refolk's index shows roughly 1,255 technical sourcers in the U.S., heavily concentrated in the Bay Area, with Apple, Anthropic, MongoDB, Verkada, Zoox, EvolutionIQ, and Zipline among the most active employers. The aggregate recruiting headcount is down 55%, yes. The elite-talent companies are not cutting sourcers. They are reallocating into them. They figured out the 411%/55% scissors before Greenhouse published the chart.

```stat
number: 0.5%
label: Share of applicants who now receive offers
note: Per pin.com's 2026 cost-per-hire benchmark, against a $5,475 average nonexecutive cost-per-hire.
</stat>

## What to do on Monday

A few moves that follow from the data, not from vibes.

### Stop measuring inbound as a primary channel

If inbound is producing 12 offers a year per recruiter and outbound candidates are 8x more likely to convert, the dashboard you're showing your CFO is measuring the wrong river. Track outbound-sourced pipeline as the primary number, inbound as a secondary, and applications-per-offer as a sanity check on whether your inbound is even worth running unfiltered.

### Kill the AI first-round interview for engineers you actually want

The data is unambiguous: 38% of candidates have already walked because of an AI interview, and the candidates most likely to walk are the ones with options. Use AI interviewers for genuinely high-volume, low-context roles where the alternative is no review at all. For anyone you'd be upset to lose, put a human in the first round.

### Treat prompt injection as a sentiment indicator

When 41% of your applicant pool is willing to embed white text in their resume, you have a trust problem, not a fraud problem. Audit what your job postings, your screeners, and your AI interviewers communicate to a sharp candidate. 69% of U.S. candidates have encountered fake job postings. They are pattern-matching you against that, and most of them are losing.

### Build the sourcing muscle while the market is distracted

The companies poaching well in 2026 are the ones who decided in 2024 that "post and pray" was over. If you don't have a sourcing function, this is the cycle to build one. If you do, this is the cycle to give it leverage. Tools like Refolk exist precisely because the "describe the person, get the person" workflow is now cheaper than the "boolean, scrape, scrub, message" workflow that defined LinkedIn Recruiter's last decade. The 411% number is also an opportunity: your competition is drowning in the same pile you are, and most of them will respond by buying another filter.

## The honest version of the headline

Greenhouse's marketing rounded 41% down to 40% because 40% reads cleaner on a chart. Fine. The number that actually matters is 411%, and it isn't about candidates at all. It's about a model of hiring that depended on a recruiter being able to give a fair human read to an application, and that model broke sometime between 2022 and 2025 while nobody was looking. Prompt injection, deepfakes, AI scripts: those are what the breakage looks like from the candidate side.

Outbound is what working hiring looks like on the other side of it. You pick the person. You reach out. You start the conversation with trust on the table, not a detector running underneath it. The vendors who sold you AI screeners will sell you AI detectors next, then AI detector detectors. The companies that are quietly winning have stopped buying tickets to that arms race.

## FAQ

### Is the 40% prompt injection number actually accurate?

The underlying figure from Greenhouse's AI in Hiring Report is 41% of 1,200 surveyed U.S. job seekers admitting to prompt injection or hidden text. Greenhouse's press materials round that to 40%. Both numbers are self-reported survey data, not artifact counts. The artifact-level prevalence is much lower: roughly 1% of resumes according to Greenhouse, around 10% containing some form of hidden text according to ManpowerGroup. The gap between intent and execution is itself the most important finding.

### Doesn't outbound sourcing just push the same volume problem onto sourcers?

No, because the unit economics are different. Inbound produces a 0.5% offer rate against 2,500-plus applications per recruiter per year. Outbound, per Gem's 2026 data, produces 8x the hire rate per touched candidate. A sourcer running 50 well-targeted outreaches a week is competitive with a recruiter triaging 50 applications a day, and the candidates are better. The constraint is targeting quality, which is exactly where modern sourcing tools earn their keep.

### What about smaller companies that can't afford a dedicated sourcing function?

That's the case where plain-English search tools matter most. A founder or hiring manager who can describe the person they want, and get a ranked list of real candidates across GitHub, LinkedIn, and the open web, is doing the job a sourcer used to do without the headcount. The 411%/55% scissors hits small teams harder than big ones, but the tooling gap has narrowed enough that one careful hire-owner can run a credible outbound process.

### Will AI screeners get good enough to fix inbound eventually?

Possibly, but that bet ignores what the candidate data is telling you. The trust loss is already priced in: 79% of U.S. candidates don't believe employers are using AI responsibly, and 50% have walked or would walk from an AI interview. Even a screener that worked perfectly would be filtering an applicant pool that has learned to game it, distrust it, or avoid it. The cheaper move is to stop relying on inbound for the hires that matter and use AI where it actually has leverage, which is in finding the people you should be talking to in the first place.