AI generated applications are flooding your inbox

By SkavanderHR · 13 September 2026

Greenhouse’s own survey found 74% of US job seekers use AI somewhere in their application, and 28% admitted using it to fabricate work samples. Gartner measured real interview fraud at 6% of candidates surveyed. The useful response is verification you can defend, not a detector, and one compliance date worth knowing.

Key takeaways

• Greenhouse’s 2025 report found 74% of US job seekers using AI in their application process, and 49% applying to more postings specifically to beat automated filters.
• 28% admitted using AI to generate fake work samples and 32% to claiming AI skills they do not have, in the same survey.
• Gartner’s own survey of about 3,000 candidates found 6% admitting to interview fraud, and its analysts forecast that one in four candidate profiles worldwide will be fake by two thousand twenty eight. That last figure is a forecast, not a measurement.
• GoodTime’s hiring report named fraudulent and AI assisted candidates the top anticipated hiring challenge of 2026, ahead of the talent shortage itself.
• The high risk obligations that attach to hiring tools under the EU AI Act moved from August two thousand twenty six to December two thousand twenty seven under the Digital Omnibus.

How much of the inbox is machine written now?

Most of it, on the only primary numbers available. Greenhouse surveyed US job seekers for its 2025 report and found 74% using AI somewhere in the application process, and 49% saying they apply to more postings than before specifically to get past automated filters. The same platform data puts the average posting at 254 applicants.

Those two findings compound in one direction. Writing an application used to cost the applicant something, and that cost was the only filter working in your favour before a human read anything. It is now close to zero, so the volume rises and the average relevance falls. Nothing about the reader’s side got faster.

Read the market wide percentages with care. The figures moving fastest here come from vendor surveys and platform data, sometimes re-cited between blogs with no shared definition of what counts as AI generated. Greenhouse’s own report is the most defensible source in the set, and it is a survey of US job seekers, not a measurement of your inbox.

Is any of it actually fraud?

Some of it, measurably. In the same Greenhouse survey, 28% of job seekers admitted using AI to generate fake work samples and 32% to claiming AI skills they did not have. Gartner surveyed about 3,000 candidates in 2025 and found 6% admitting to interview fraud, meaning posing as someone else or having someone pose as them, and only 26% saying they trust AI to evaluate them fairly.

Gartner’s analysts also forecast that one in four candidate profiles worldwide will be fake by two thousand twenty eight. It is worth repeating that this is a prediction from a research firm, not a count of anything that has happened.

What the published surveys found

What was measured	Figure	Published by
US job seekers using AI in their application process	74%	Greenhouse, 2025
Applying to more postings to beat automated filters	49%	Greenhouse, 2025
Admit using AI to generate fake work samples	28%	Greenhouse, 2025
Admit claiming AI skills they do not have	32%	Greenhouse, 2025
Candidates admitting interview fraud, survey of about 3,000	6%	Gartner, 2025
Applicants who trust AI to evaluate them fairly	26%	Gartner, 2025

Sources: Greenhouse, 2025 AI in Hiring Report. Gartner press release of the thirty first of July two thousand twenty five, survey of about 3,000 job candidates, with the fake candidate forecast reported by HR Dive in August of the same year.

What verifies a candidate without punishing the honest ones?

Four things that hold up, and one that does not.

Writing style is not evidence. A fluent cover letter proves nothing now, in either direction, and a candidate who used a model to fix their second language grammar is not committing fraud. Screening on polish selects for tool use, not ability, and detection tools that claim to spot generated text are not something a hiring decision should rest on.

What does work is asking for work you can examine live. A short conversation about something the candidate actually built, where they chose a trade off and can explain why, cannot be prepared in advance by anything that was not there. So can a live walk through of their own history: the projects, the team sizes, the parts they owned. Keep the questions the same for everyone at that stage, write the answers down, and keep a human making the decision. That last point is both the fairer process and the one that is easiest to defend if a rejected candidate ever asks how the decision was made.

For identity specifically, a live video conversation early, with a camera on, does most of the work at this company size. It is a low ceremony version of what large employers are buying dedicated tooling for.

What do the EU AI Act deployer duties mean for you?

This is not legal advice, and the detail of what applies depends on what a tool does. The plain shape is this: a company that buys and uses a hiring tool is a deployer, not a provider, and the heavier duties under the EU AI Act attach to uses classed as high risk, which is where employment and worker management sit. Guidance published through 2026 records that those high risk obligations were pushed from August two thousand twenty six to December two thousand twenty seven under the Digital Omnibus, so anything written earlier is likely to quote the old date.

The practical reading for a smaller hiring team: the clock moved, the direction did not. A tool that ranks, filters or rejects applicants is the kind of use the rules are written about, and deferring a date is not the same as removing a duty. A tool that finds people and starts a conversation, leaving every judgement to your own team, is a different object from one that decides. Knowing which of the two each item in your stack is, before a date arrives, is the whole preparation.

What does this mean for a company of twenty to five hundred people?

The flood is not a screening problem you can outrun. Volume rose because applying got cheap, and no filter you add makes the underlying pool more relevant. Adding a stricter automated gate makes it worse twice over: it rejects honest applicants on formatting, and it hands you exactly the compliance exposure described above.

The alternative is to start from a smaller, verified pool instead of a large anonymous one. A person who was found by name in the market, who was not applying, and who answered a specific message about one specific role, arrives with an identity and a history you already checked before the first call. There is no submission to fake, because there was no submission.

That is the search we run. Every relevant person for one open role is found across the professional networks and company websites and told about the role, usually covering the market within two weeks, and the people who answer come with their own words about their own work. How the search runs sets out the mechanics, an agency and a monthly search compared sets out the money, and a market such as AI engineers is where the fabricated skill claims bite hardest, which is exactly why the pool matters more than the filter.

Numbers from our own searches. Written and checked by SkavanderHR.