Blab AI
Google AI Hiring Tools vs. DeepMind Safety Team: Candidates Urged to Circumvent Internal Screening
When Google sells AI as the future of recruitment, its own researchers are quietly warning that the same algorithms could be sidelining the very talent the company hopes to attract.
Bloomberg reported on August 10 that Google’s AI division has described its internal human‑resources filters as "unreliable". The comment surfaced as the firm rolled out AI‑powered hiring tools to corporate clients, promising faster screening of large applicant pools.
At the same time, a confidential memo from Google DeepMind’s AGI Safety and Alignment Team (ASAT) was released to Bloomberg. Titled "PLEASE DO NOT SHARE THIS DOC WIDELY," the document instructs applicants for ASAT roles to complete a special form in addition to the standard application. The purpose, according to the memo, is to help candidates avoid being automatically screened out by the team’s internal AI systems.
DeepMind, acquired by Alphabet in 2014 and merged with Google Brain in 2023, is a research lab that builds large‑language models and other generative AI tools. The AGI Safety and Alignment Team focuses on mitigating existential risks posed by advanced AI. The memo explains that the team’s screening algorithms can unintentionally filter out qualified candidates, and the form is a mitigation measure.
The contrast between Google’s outward promotion of AI hiring tools and DeepMind’s internal caution underscores a broader industry debate about algorithmic bias. A 2026 study published in the Technology Review, alongside research from Stanford, found that AI screening can disproportionately exclude certain demographic groups, including Black applicants and non‑native English speakers. The studies also warned that bias can be amplified when models are trained on historical hiring data that reflects past discrimination.
Regulators are taking notice. In the United States, New York City Local Law 144 and Illinois’ AI employment rules—both effective January 2026—require bias testing, disclosure, and human‑review pathways for high‑risk AI hiring tools. The European Union’s AI Act, adopted in 2024, similarly classifies AI hiring systems as high‑risk and imposes strict compliance obligations.
Google maintains that its AI tools are designed to assist recruiters, not replace human judgment. The company’s AI‑driven candidate‑screening guide, released in 2025, outlines how its systems can reduce screening hours and improve candidate matching. Yet Bloomberg’s article notes that internal AI filters have been flagged as unreliable, hinting at a disconnect between the tools offered to clients and the systems used internally.
The DeepMind memo does not disclose the specific algorithms used in its screening process, nor does it provide metrics on the rate of false negatives. Chief AGI Scientist Rohin Shah has publicly emphasized the importance of transparency and human oversight in AI hiring, but the memo itself remains an internal communication.
This situation highlights the tension between commercial deployment of AI hiring tools and the ethical responsibilities of AI developers. While Google markets its solutions as a way to streamline talent acquisition for external customers, its own researchers are taking steps to ensure that internal hiring processes do not inadvertently disadvantage qualified candidates.
As AI hiring tools become more widespread, scrutiny from regulators, researchers, and advocacy groups is likely to intensify. Companies deploying AI screening systems will need to demonstrate compliance with emerging legal standards and address documented bias risks. For candidates, the DeepMind memo suggests that completing additional forms or providing supplementary information may improve the likelihood of being considered by AI‑driven hiring pipelines.
The current landscape reflects a growing awareness that AI can both accelerate recruitment and introduce new forms of bias. Whether Google’s external AI hiring products will incorporate the safeguards that DeepMind’s internal team recommends remains to be seen, but the contrast between the two approaches is already shaping conversations about responsible AI in hiring.
Bloomberg reported on August 10 that Google’s AI division has described its internal human‑resources filters as "unreliable". The comment surfaced as the firm rolled out AI‑powered hiring tools to corporate clients, promising faster screening of large applicant pools.
At the same time, a confidential memo from Google DeepMind’s AGI Safety and Alignment Team (ASAT) was released to Bloomberg. Titled "PLEASE DO NOT SHARE THIS DOC WIDELY," the document instructs applicants for ASAT roles to complete a special form in addition to the standard application. The purpose, according to the memo, is to help candidates avoid being automatically screened out by the team’s internal AI systems.
DeepMind, acquired by Alphabet in 2014 and merged with Google Brain in 2023, is a research lab that builds large‑language models and other generative AI tools. The AGI Safety and Alignment Team focuses on mitigating existential risks posed by advanced AI. The memo explains that the team’s screening algorithms can unintentionally filter out qualified candidates, and the form is a mitigation measure.
The contrast between Google’s outward promotion of AI hiring tools and DeepMind’s internal caution underscores a broader industry debate about algorithmic bias. A 2026 study published in the Technology Review, alongside research from Stanford, found that AI screening can disproportionately exclude certain demographic groups, including Black applicants and non‑native English speakers. The studies also warned that bias can be amplified when models are trained on historical hiring data that reflects past discrimination.
Regulators are taking notice. In the United States, New York City Local Law 144 and Illinois’ AI employment rules—both effective January 2026—require bias testing, disclosure, and human‑review pathways for high‑risk AI hiring tools. The European Union’s AI Act, adopted in 2024, similarly classifies AI hiring systems as high‑risk and imposes strict compliance obligations.
Google maintains that its AI tools are designed to assist recruiters, not replace human judgment. The company’s AI‑driven candidate‑screening guide, released in 2025, outlines how its systems can reduce screening hours and improve candidate matching. Yet Bloomberg’s article notes that internal AI filters have been flagged as unreliable, hinting at a disconnect between the tools offered to clients and the systems used internally.
The DeepMind memo does not disclose the specific algorithms used in its screening process, nor does it provide metrics on the rate of false negatives. Chief AGI Scientist Rohin Shah has publicly emphasized the importance of transparency and human oversight in AI hiring, but the memo itself remains an internal communication.
This situation highlights the tension between commercial deployment of AI hiring tools and the ethical responsibilities of AI developers. While Google markets its solutions as a way to streamline talent acquisition for external customers, its own researchers are taking steps to ensure that internal hiring processes do not inadvertently disadvantage qualified candidates.
As AI hiring tools become more widespread, scrutiny from regulators, researchers, and advocacy groups is likely to intensify. Companies deploying AI screening systems will need to demonstrate compliance with emerging legal standards and address documented bias risks. For candidates, the DeepMind memo suggests that completing additional forms or providing supplementary information may improve the likelihood of being considered by AI‑driven hiring pipelines.
The current landscape reflects a growing awareness that AI can both accelerate recruitment and introduce new forms of bias. Whether Google’s external AI hiring products will incorporate the safeguards that DeepMind’s internal team recommends remains to be seen, but the contrast between the two approaches is already shaping conversations about responsible AI in hiring.
