Google DeepMind's Safety Team Warns Job Applicants: Our AI May Wrongly Reject Your Resume

The AGI Safety and Alignment team at Google DeepMind, the group tasked with mitigating risks from advanced artificial intelligence, has told job applicants it doesn't fully trust its own AI recruiting tools. In an internal document meant to stay within the pool of applicants, the team warned that its resume-screening system could wrongly reject strong candidates, and provided a way to bypass the automated filter.

What exactly the team wrote

The document, intended for candidates applying to open positions on the AGI Safety and Alignment team, contained a line that companies normally try hard to keep out of public view: the very system meant to help sort applications is itself a risk to candidates. The authors wrote plainly that there is a 'non-trivial probability' that a resume could be rejected by an algorithm's mistake or get stuck in a queue too long to ever reach a recruiter.

The proposed fix was an additional form that bypasses the standard recruiting pipeline and routes the application directly to members of the team. The document carried the caveat 'PLEASE DO NOT SHARE THIS DOC WIDELY', a request not to spread it further that suggests the authors knew how awkward it looked to admit their own technology's flaws.

We have an applications system with a non-trivial probability your CV will be screened out incorrectly or take too long to reach us. Filling out this form makes sure that a real human on the team will get to see your application - from the AGI Safety and Alignment team's document at Google DeepMind

The same tool, sold to clients

The paradox is that Google simultaneously sells enterprise clients AI tools for automatically screening candidates, systems that rank applications or filter resumes for specific keywords. Yet the team responsible for the safety of the company's most advanced models decided that, for its own hiring, it was better to give candidates a way around such a mechanism than to risk losing a good candidate to an algorithm's error.

The document also included advice for applicants themselves: avoid answers written by large language models, since recruiters reading hundreds of applications quickly spot generated text and grow tired of it, because such answers 'all sound very similar'. It's another sign that even inside the company building some of the world's most advanced language models, there's a belief that AI-written text works against candidates in a hiring process.

Google DeepMind's response

After Bloomberg broke the story, a Google DeepMind spokesperson denied that the resume-filtering system systematically screens out candidates incorrectly. The spokesperson described the team's form as an optional, additional point of contact rather than evidence that the tool is flawed.

This team set up a special form to go past the recruiter review, and get their resumes direct to the people on the team. But there are no shortcuts to getting hired - Google DeepMind spokesperson

The company also stressed the broader goal of its hiring process: reaching the most qualified candidates. Still, the spokesperson's stated position stands in clear contrast to the internal document, which left little doubt about the existence of the problem.

Why this matters

The case feeds into a broader unease around hiring automation, one increasingly discussed in Poland too: studies suggest that recruitment increasingly resembles a duel between algorithms, with candidates using AI to write resumes and companies using AI to reject them. The Google DeepMind case shows the problem isn't confined to smaller companies experimenting with cheap tools, it also touches the AI industry's leader, which sells similar solutions to others.

For Polish companies deploying AI-assisted ATS systems, this is a warning sign: even the makers of the most advanced models conclude that automated resume filters need a safeguard, an alternative path to a human reviewer. Without such a mechanism, a company risks rejecting valuable candidates purely because of a classifier's error rather than their actual qualifications.

The case also points to a growing gap between official corporate messaging and internal practice at tech companies. The spokesperson's public stance, that the filters work correctly, collides with a private team document that judged the risk real enough to build a workaround for, and to ask that people not talk about it too loudly.