The core irony

ASAT exists to reduce existential and catastrophic risks from AGI and eventually superintelligence . Yet the team’s internal guidance explicitly warns that Google’s own AI hiring filters cannot be trusted. According to a document viewed by Bloomberg, the team states: “We have an applications system with a non-trivial probability your CV will be screened out incorrectly or take too long to reach us” . The team is effectively telling prospective hires: do not rely on the company’s AI to competently evaluate you.

Bypassing the company’s own AI

The internal document, which carries the unusual disclaimer “PLEASE DO NOT SHARE THIS DOC WIDELY,” directs candidates for ASAT roles to complete a confidential internal submission form . The form is designed to ensure applications land directly in front of human reviewers rather than being fed into Google’s automated CV-screening pipeline . Bloomberg reported that the form’s purpose is to eliminate the risk that applications “may be automatically screened out by the company’s internal AI systems” .

Warning against AI-generated application materials

ASAT’s job postings on the Alignment Forum state: “Please don’t use AI to write your application; we want to hear your voice” . This creates a layered contradiction: the team distrusts Google’s AI screening systems enough to sidestep them, but also distrusts candidates’ use of AI chat tools to draft responses—even as ASAT’s own work involves developing and overseeing advanced AI systems.

Broader context of AI hiring bias scrutiny

The ASAT episode is not an isolated incident. A series of studies and investigations have documented persistent problems with AI-driven hiring tools:

• A July 2026 Stanford HAI study analyzed 4 million job applications across 150+ employers and found that AI hiring tools can produce racially disparate outcomes at scale. The researchers found that 15% of Asian applicants and 26% of Black applicants were repeatedly turned away across multiple jobs—a pattern suggesting systemic bias .
• A Bloomberg analysis showed that OpenAI’s GPT-3.5, when asked to rank résumés 1,000 times, systematically favored names from some demographics over others to an extent that would fail benchmarks used to assess job discrimination .
• Research from the University of Melbourne found that AI hiring software may discriminate against applicants wearing headscarves, those with names perceived as Black, and individuals who request disability accommodations .
• A UK Information Commissioner’s Office audit found that some AI recruitment tools filter applicants based on protected characteristics such as race, gender, and sexual orientation .

Google itself sells AI-powered HR tools to corporate clients, promising faster, more objective screening . The ASAT episode undercuts that marketing message by implicitly acknowledging the company’s own filters are fallible.

Google DeepMind’s response

DeepMind disputed the characterization that its broader hiring systems are broken. A spokesperson described the bypass form as simply an “efficiency measure” for high-volume roles . However, the internal document’s own language—warning of incorrect screening and urging secrecy—suggests the team regards the issue as material enough to warrant a workaround .

The bottom line

The situation reveals a credibility gap: the very team Google employs to research and guard against AI misalignment does not trust the AI systems Google uses to screen its own future hires. This dissonance has drawn wide attention as a real-world case of an AI developer implicitly acknowledging the limitations and risks of the same technology it builds and sells. As one analysis summarized: “the team responsible for advanced AI safety does not trust its own employer’s AI to read résumés” .

The core irony

ASAT exists to reduce existential and catastrophic risks from AGI and eventually superintelligence . Yet the team’s internal guidance explicitly warns that Google’s own AI hiring filters cannot be trusted. According to a document viewed by Bloomberg, the team states: “We have an applications system with a non-trivial probability your CV will be screened out incorrectly or take too long to reach us” . The team is effectively telling prospective hires: do not rely on the company’s AI to competently evaluate you.

Bypassing the company’s own AI

The internal document, which carries the unusual disclaimer “PLEASE DO NOT SHARE THIS DOC WIDELY,” directs candidates for ASAT roles to complete a confidential internal submission form . The form is designed to ensure applications land directly in front of human reviewers rather than being fed into Google’s automated CV-screening pipeline . Bloomberg reported that the form’s purpose is to eliminate the risk that applications “may be automatically screened out by the company’s internal AI systems” .

Warning against AI-generated application materials

ASAT’s job postings on the Alignment Forum state: “Please don’t use AI to write your application; we want to hear your voice” . This creates a layered contradiction: the team distrusts Google’s AI screening systems enough to sidestep them, but also distrusts candidates’ use of AI chat tools to draft responses—even as ASAT’s own work involves developing and overseeing advanced AI systems.

Broader context of AI hiring bias scrutiny

The ASAT episode is not an isolated incident. A series of studies and investigations have documented persistent problems with AI-driven hiring tools:

• A July 2026 Stanford HAI study analyzed 4 million job applications across 150+ employers and found that AI hiring tools can produce racially disparate outcomes at scale. The researchers found that 15% of Asian applicants and 26% of Black applicants were repeatedly turned away across multiple jobs—a pattern suggesting systemic bias .
• A Bloomberg analysis showed that OpenAI’s GPT-3.5, when asked to rank résumés 1,000 times, systematically favored names from some demographics over others to an extent that would fail benchmarks used to assess job discrimination .
• Research from the University of Melbourne found that AI hiring software may discriminate against applicants wearing headscarves, those with names perceived as Black, and individuals who request disability accommodations .
• A UK Information Commissioner’s Office audit found that some AI recruitment tools filter applicants based on protected characteristics such as race, gender, and sexual orientation .

Google itself sells AI-powered HR tools to corporate clients, promising faster, more objective screening . The ASAT episode undercuts that marketing message by implicitly acknowledging the company’s own filters are fallible.

Google DeepMind’s response

DeepMind disputed the characterization that its broader hiring systems are broken. A spokesperson described the bypass form as simply an “efficiency measure” for high-volume roles . However, the internal document’s own language—warning of incorrect screening and urging secrecy—suggests the team regards the issue as material enough to warrant a workaround .

The bottom line

The situation reveals a credibility gap: the very team Google employs to research and guard against AI misalignment does not trust the AI systems Google uses to screen its own future hires. This dissonance has drawn wide attention as a real-world case of an AI developer implicitly acknowledging the limitations and risks of the same technology it builds and sells. As one analysis summarized: “the team responsible for advanced AI safety does not trust its own employer’s AI to read résumés” .