How AI resume screening actually changed in 2026

The short answer: the workflow changed, not the reader

In February 2026 Greenhouse announced that Real Talent was available, a top-of-funnel suite that puts AI-powered Talent Matching, spam protection, fraud detection and identity verification inside the ordinary recruiting workflow. That is the concrete, dated change, and the vendor published it.

What did not arrive with it is any published account of how a match score is computed, or whether the software reads a sentence the way a person does. That question is the one most advice answers confidently, and it is the one the documentation leaves open.

The practical consequence is smaller and more useful than the headline version: write to the requirements the employer stated, because those are what the published scoring descriptions are built around.

• Documented: AI matching is now a configurable, generally available part of the workflow.
• Documented: a person sets the criteria, sees the reason, and makes the decision.
• Not documented: the model behind the score, or how it treats your writing.

What the vendors actually shipped

Greenhouse's February 2026 release notes describe Talent Matching as AI-powered filters that organise applications by how they match your job criteria, such as skills, industry and experience. The same entry states that recruiters define and weigh the job criteria, that they can see exactly why a candidate was flagged, and that Talent Matching never advances or rejects candidates automatically.

The supporting documentation adds the parts a candidate would want to know. Talent Matching is described as assistive AI rather than automated decision-making. It needs resume parsing switched on to work fully. Recruiters can override a candidate's match score, and there is a view for finding candidates who need manual review.

There is also a disclosure and consent layer, which is new in a way that matters. Greenhouse supplies sample wording for employers to put on job posts, explaining that the tool reviews your resume to determine the extent to which your qualifications match the requirements the employer selected. Where an employer enables it, candidates can request a manual review that does not use the AI.

Read that list again and notice what kind of change it is. Almost every documented item is about control, disclosure and human responsibility. Very little of it is about the software becoming a better reader.

• Recruiters choose the criteria and how much each one weighs.
• The recruiter is shown the reason a candidate was surfaced.
• No automatic advancing or rejecting, per the vendor's own wording.
• Disclosure wording, and in some configurations a route to non-AI review.

The part everyone asserts and nobody publishes

The claim in circulation is specific: that screening moved from exact-keyword matching to semantic or language-model scoring, and that this rewards honest, well-written, quantified resumes over stuffed ones. We went looking for a vendor saying so and did not find one.

Say that precisely, because the precise version is the useful one. We did not find documentation of that change. That is a statement about what is published, not a claim about what the systems do internally. Vendors may well be using language models. They have not told candidates, and a candidate cannot verify it.

Set against the claim, the nearest published description points the other way. Workday's reference page for HiredScore candidate grades defines an A as a candidate who meets all basic required qualifications, most preferred qualifications, and has an average qualification match above a defined threshold. B meets all basic qualifications, C meets most but not all, D does not meet most. That is the shape of a qualification checklist, not of a critic reading prose. The page carries a last-updated stamp from 2023.

The same documentation also shows the older behaviour has not gone anywhere. A recruiter looking at a candidate in HiredScore can still search for keywords in the original resume file.

None of this proves the systems are unchanged. It means the confident architecture story is not something you can check, so it is a poor foundation for deciding what to write.

Why stuffing got more expensive, for a reason you can check

The usual explanation is that the software got clever enough to catch you. There is a better explanation available, and it does not require guessing at any model.

The score now arrives in front of a person together with its reason. Greenhouse's description says the recruiter defines the criteria, sees exactly why a candidate was flagged, and decides. A page engineered to trip a filter is being handed to a human alongside an explanation of what tripped, which is close to the worst available way to be noticed.

The same February release shipped fraud detection and identity verification next to the matching. Fraud detection is described as flagging suspicious applications from signals such as phone number, email, IP address and location. That is aimed at fake applicants rather than at embellished bullets, but it tells you the direction the tooling is moving: the top of the funnel is being instrumented for authenticity, and the cost of looking engineered is rising in that context.

And the oldest reason still stands. Every trick in this family fails at the same place, which is the conversation where somebody asks you about the thing you wrote.

Writing for both readings at once

The useful accident here is that the old and new descriptions point at the same behaviour. Workday grades against stated basic and preferred qualifications. Greenhouse matches against criteria a recruiter selected and weighted. Neither rewards volume, and both are organised around the requirements the employer actually wrote down.

So there is one move, and it works whichever description is closer to the truth: cover the stated requirements truthfully, in ordinary words, somewhere a parser can reach them.

• Work from the posting in front of you, not a general list of terms for your field.
• Cover the required qualifications first, then the preferred ones, and do it explicitly rather than by implication.
• Put each claimed skill next to the role where you used it, so the claim has something under it.
• Use the employer's ordinary word for a thing as well as your own, once each, where both are true.
• Keep the structure plain enough to parse: one column, standard headings, real text, contact details in the body.
• Quantify where you genuinely measured something, and describe scope or before-and-after where you did not.
• Leave out anything you could not talk about for ten minutes.

Check the layer that publishes its rules

Two layers sit between your file and a recruiter. The first extracts text and structure, and vendors publish quite a lot about what breaks it: accepted formats, image-only files, how layout and word order affect results. The second ranks or grades what came out, and almost nothing about it is published.

That split is the whole basis for what is worth doing. You can verify your standing on the documented layer, and you cannot verify it on the other one, so effort spent guessing at the second is effort not spent on the first.

This is why Ryzio's checker is built the way it is. It is a deterministic audit rather than a model's opinion: the same file produces the same result, every finding names the rule it comes from, and each rule cites where it came from. It also reports compatibility separately for six systems, because the documented failure modes genuinely differ between them, and a single score would hide that.

It is worth being equally clear about the limit. No checker can tell you your Talent Matching score or your HiredScore grade, because those depend on criteria a particular recruiter set for a particular job and on scoring nobody publishes. Anything claiming to report your real ATS score is describing something it cannot see. What can be checked is whether the document is readable, complete, and visibly aligned with the posting, which is the part that is in your hands.

What we checked, and what we could not

The dated claims here come from Greenhouse's February 2026 release notes and its Talent Matching support documentation, and from Workday's HiredScore documentation for candidate grades and candidate profiles. All were read on 14 September 2026.

We could not verify the central claim that prompted this article, that vendors replaced keyword matching with semantic or language-model scoring during 2026. No vendor documentation we found describes that transition. We are reporting a search that came up empty, which is weaker than a finding and stronger than repeating the claim.

We also have no outcome data. Nobody has published a study varying how a resume is written and reporting the effect on match scores or interview rates, so no number in this area should be trusted, including any that promises a percentage improvement.

Where that leaves things: the documented behaviour is worth writing to, the undocumented behaviour is worth refusing to guess at, and the difference between them is the only honest place to stand.