What is AI resume screening?

It is software that reads each application against the job it was sent to and produces a relevance signal, usually a score and a short summary, so a person can read the most relevant applications first. It is a sorting tool. The good implementations rank and explain; the bad ones silently reject.

Is AI resume screening legal?

It depends where you hire. New York City requires bias audits and candidate notice for automated employment decision tools, and the EU AI Act treats hiring systems as high risk with obligations that are still phasing in. Rules are tightening rather than loosening. Check your own jurisdiction before you automate anything that rejects a person, and keep records of how you used it.

Does AI resume screening discriminate?

It can, and reporting has repeatedly found bias and plain bugs in commercial hiring tools. A model trained on who you hired before will happily reproduce who you hired before. The mitigations that actually work are mundane: score against the role's requirements rather than a profile of past hires, keep a human on every rejection, look at the score distribution by role rather than trusting individual numbers, and re-read the ones the model ranked low before you close a search.

Can candidates game it?

Keyword stuffing beats keyword matching, which is why keyword matching is the weakest form of this. Language-model scoring is harder to game with a wall of terms, but it is more easily swayed by a confidently written application than a modest one. Assume the polish of the writing is affecting the score, because it is.

That is the question most buyers forget to ask. With a hosted tool, every CV is sent to that vendor and whoever they use. If that matters to you, either check the processor terms carefully or choose something that runs the model on your own infrastructure. WP Career Board Pro supports a self-hosted Ollama model for exactly this reason, so applications never leave your server.

How many applications does it take to be worth it?

Somewhere around thirty per role is where reading everything properly stops being realistic for most small teams. Below that, read them yourself. Above a hundred, the choice is not between AI and careful reading, it is between AI and skimming, which has its own quiet biases and no audit trail at all.

Do I still need to reply to everyone?

Yes, and screening makes that easier rather than harder because you get through the pile faster. A ranked list is not a reason to leave the bottom of it without an answer.

AI resume screening: what it does, what it gets wrong, how to use it

AI resume screening reads each application against the job it was sent to and gives you a ranked list with a reason attached, so the first thing you read is the most relevant thing rather than the most recent. That is the whole idea. Everything else in this guide is about where it helps, where it fails, and how to use it without letting it make the decision.

What it actually does

A screening tool takes the job description and one application, and returns a relevance signal: usually a score, often a sentence explaining the score, sometimes a summary of the candidate. Good implementations rank and explain. Bad implementations reject silently, which is both the version that gets written about and the version that gets employers into trouble.

It is a reading order. It is not a decision, and the moment you treat it as one you have built something you cannot defend to a rejected candidate or a regulator.

How it works, in one pass

Most current tools do some version of this:

Extract the text of the application and the CV attached to it.
Send it to a model along with the job description, asking for a structured judgement rather than prose.
Store the result against the application so it is not recalculated every time somebody opens the list.
Sort the applications by that score.

The older generation matched keywords instead of using a model. That is the version candidates learned to beat by pasting the job advert into white text at the bottom of the page, and it is the reason “ATS-friendly CV” is an industry of its own.

Where it goes wrong

It inherits your history. A model asked to find people like your current team will find people like your current team. Score against what the role actually requires, not against a profile built from who you hired before.

It rewards confident writing. Two people with identical experience do not write identically. The one who writes with more assurance will tend to score higher, and that correlates with things you did not intend to select for.

It is stated too confidently. A number between 0 and 100 looks like a measurement. It is an opinion with a decimal point. This is the practical reason to insist on a one-line reason next to every score: a reason can be disagreed with, a bare number invites deference.

Reporting on commercial hiring tools has repeatedly turned up both bias and straightforward bugs, so none of this is theoretical.

How to run it responsibly

Keep a human on every rejection. If nobody read it, do not send it.
Re-read the bottom. Before closing a search, look at a sample the model ranked low. You are checking the tool, not the candidates.
Score against requirements, written out as requirements, not against a vague idea of a good fit.
Keep the reasons. They are your record of why a shortlist looked the way it did.
Check your jurisdiction. New York City requires bias audits and candidate notice for automated employment decision tools. The EU AI Act treats hiring as high risk. Neither trend is reversing.

The question people forget: where do the CVs go?

A CV is about as personal as a document gets: name, address, employment history, sometimes far more. With a hosted screening tool, every one of them is sent to that vendor, and to whichever model provider sits behind them.

Sometimes that is fine and the processor terms are clear. Sometimes it is a problem you did not know you had agreed to, particularly if you are handling applications under GDPR and told candidates their data stays with you.

The alternative is running the model yourself. That used to be impractical and is now routine: a local model on your own server, doing the same scoring, with nothing leaving the building.

Doing this on your own WordPress site

If your careers page already runs on WordPress, the screening can too. WP Career Board Pro scores each application against its job, writes a one-sentence reason and a neutral summary, and ranks the list. It runs on Claude, OpenAI or a self-hosted Ollama model, and the results are stored with the application so reopening a role costs nothing.

The free version handles the board itself: publishing roles, collecting applications, employer and candidate dashboards, and the structured data that makes your jobs eligible for Google’s job results. Screening is the part you add when the pile gets big enough to need it.

Rules on automated hiring tools are changing. This is a practical guide, not legal advice, and it was written in September 2026.

AI resume screening, honestly