Screening applications with a model runs into two separate rules, and most discussions only cover one of them.
The first problem: it is their data
A CV is dense personal data — name, contact details, employment history, education, frequently age and nationality by inference. The candidate provided it to you for a specific purpose. Passing it to a model provider is a disclosure to a third party they did not choose.
Practically: your privacy notice for candidates should say it, your lawful basis should cover it, and the retention question applies to the copy at the vendor as well as the one in your system.
The second problem: automated decisions
Where a decision is made solely by automated means and has a significant effect, GDPR gives the individual specific rights — to know, to obtain human intervention, to contest. Rejecting a candidate is a significant effect. Whether the rule applies depends on how meaningful your human step is.
A human who approves a list of rejections in nine seconds is not meaningful intervention.
This is where practice and policy part company. A recruiter clicking through a ranked list at speed is, functionally, an automated decision with a person's name attached to it. Regulators have shown interest in exactly that pattern.
Bias, briefly and honestly
This is covered elsewhere at length, so only the part that touches the privacy question: models infer protected characteristics from CVs readily, from names, schools, dates and phrasing. You do not have to ask for them to end up acting on them, and “we did not send the protected fields” is not a defence if the model inferred them anyway.
Removing names and other identifying detail before screening is the one intervention that helps both problems at once. It reduces what you disclosed and it removes the most direct route to inferring characteristics you are not allowed to consider.
A defensible shape
- •Tell candidates, in the notice they actually see, that AI assists screening and how.
- •Strip identifying detail before the model sees the application. Names, contact details, addresses, dates that reveal age, institution names where practical.
- •Keep the human step real: give the reviewer the material and the time to disagree, and record when they do.
- •Be able to explain a rejection in terms a person can contest.
- •Apply your retention schedule to the vendor's copy, not only yours.
Try Secure AI free
Frontier AI chat — encrypted by default, with your identity stripped before any request reaches the provider.
Common questions
Can I use AI to screen job applications?
Yes, with two conditions that are often missed. Candidates need to be told, because sending their CV to a model provider is a disclosure to a third party they did not choose. And if the decision is effectively automated and significant — a rejection — specific rights attach, including human intervention and the right to contest.
Does a recruiter clicking approve count as human review?
Only if it is meaningful. A reviewer with the authority, the information and the time to disagree is a human decision. Someone confirming a ranked list at speed is an automated decision with a name attached, and that distinction is where regulators have focused.
Does anonymising CVs solve the bias problem?
It removes the most direct route, and it is the single most effective step available — it addresses the disclosure and the inference at the same time. Bias work does not stop there, because models also read schools, dates and phrasing, so pair it with structured criteria and a real human decision.