DeepMind's AGI Safety Team Dodges Google's Own AI Screener

On August 10-11, 2026, Bloomberg reported that Google DeepMind's AGI Safety and Alignment Team is quietly handing candidates a supplemental form so their resumes bypass Google's own internal AI screening tools. The document was marked "PLEASE DO NOT SHARE THIS DOC WIDELY." If the team building safety tech at Google does not trust the hiring tech Google sells, sourcers should take that as a signal about where the real alignment talent lives and how to reach it.

What the DeepMind memo actually says

DeepMind's AGI Safety and Alignment team is telling applicants that Google's internal AI screener has a "non-trivial probability" of rejecting them incorrectly, and is routing them around it via a supplemental form that lands in front of a human on the team.

The verbatim text Bloomberg reviewed:

• "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."
• The document was headed "PLEASE DO NOT SHARE THIS DOC WIDELY."

Google's spokesperson framed it as procedural: "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 denies its systems filter applicants incorrectly.

The irony is load-bearing. Google Workspace markets AI features that "save HR time by quickly creating drafts for job postings, evaluating resumes, and forecast[ing]" hiring needs. Google's own AGI safety team is the internal customer refusing to use it.

Why this is a legal document, not a candidate courtesy

Read as evidence, the memo is a de-facto internal admission that Google's automated screener produces incorrect outcomes at a "non-trivial" rate, at the exact moment US courts are testing whether AI screeners create disparate impact.

Two live cases give the memo weight:

• Mobley v. Workday: a federal judge in California declined to dismiss claims that Workday's AI screener discriminates against applicants over 40.
• Meta AI-layoff suit: 26 Meta employees are suing over AI-driven layoff selection.

Any enterprise HR buyer sitting through a Google Workspace pitch in Q4 2026 now has a two-line rebuttal. Procurement teams will cite this memo. Recruiters selling into enterprise HR should expect it in RFP objections by year-end.

For sourcers, the second-order effect matters more: the top AI safety candidates just watched Google's own alignment team publicly distrust automated resume review. Any outbound that reads as ATS-formatted, LLM-generated, or filter-first will bounce off this cohort harder than any other in tech.

The alignment pool is smaller than any recent layoff

The global technical AI safety population is roughly 600 full-time researchers, and only 42 of them actively use "AI Safety Researcher," "Alignment Researcher," or "AGI Safety" as their current title.

The 2025 LessWrong field-growth analysis pegs the technical AI safety workforce at ~600 FTEs globally, with another ~500 in non-technical safety roles (1,100 total). Google DeepMind's safety team is estimated at 40+ people, meaning a single team accounts for close to 7% of the entire global technical pool. Anthropic, OpenAI's Superalignment successor, UK AISI, and a handful of academic labs absorb most of the rest.

42

For context: the 2022 estimate was ~300 technical FTEs. The pool doubled in three years. A 2023 TIME piece cited by an arXiv survey put the working population at "only around 80 to 120 researchers in the world" working full-time on alignment. The field has grown roughly 5x in five years and it is still smaller than a single mid-tier engineering org.

The 72x gap: why title-based sourcing misses 98.6% of the pool

For every one person who uses an alignment title, Refolk's index shows 72 people with "AI Safety" or "Alignment" as a listed skill under a generic title. Boolean searches keyed on the title miss ~98.6% of the addressable pool.

Here is the shape of the market, drawn from the research anchors and Refolk's index of professional profiles:

The 72x gap has a mechanical cause, not a cultural one. Alignment researchers at frontier labs carry the same generic titles as everyone else in research: "Research Scientist" at Anthropic, "Member of Technical Staff" at OpenAI, "Research Engineer" at DeepMind, "PhD Student" at Oxford or MILA. Their alignment work shows up in publications, GitHub repos on interpretability or RLHF safety, workshop attendance, and skill tags. It almost never shows up in the job title field an ATS filter reads.

This is the exact gap Refolk closes: you describe the researcher in plain English (interpretability, mechanistic, RLHF, evals, red-teaming), and Refolk searches across GitHub, LinkedIn, and the open web for people whose actual work matches, not just their title string. The 3,046-person skill pool is the real addressable market. The 42-person title pool is what your competitors are fighting over.

Safety is losing the headcount race by 10 to 20 points

Technical AI safety headcount is growing at ~21% annually while capabilities researcher headcount grows at an estimated 30-40%. The safety-to-capabilities ratio gets worse every year, which means the marginal 2026 alignment hire has more leverage, more offers, and less patience for broken pipelines than the marginal capabilities hire.

The DeepMind form is a symptom of a candidate market where a top-five AI lab has to publish a workaround memo to see resumes it already wants to hire. Translated for AGI safety team recruiters:

• The candidate holds the pen on process. If your intake burns 15 minutes on a login-required application, you are behind Anthropic, whose safety team lets researchers email a partner directly.
• Cold outbound quality matters more than volume. The DeepMind memo warns candidates against LLM-generated applications because human reviewers are tired of generic responses. The reverse is also true: alignment researchers can spot templated outbound in one sentence.
• Speed to first human touch is the primary conversion variable. If your first response is an ATS auto-reply, you have lost the candidate to whoever answered with a real paragraph.

The team building safety tech at Google refuses to use the hiring tech Google sells. That is your sourcing signal.

Fellowships, not employers, are the correct top of funnel

The dominant employers in Refolk's index of titled alignment researchers are fellowship programs, not frontier labs. Sourcing by employer is late; sourcing by fellowship cohort catches researchers 6 to 18 months before they land at DeepMind or Anthropic.

The named programs to map:

• LASR Labs (London): Refolk's top employer for people with alignment titles. Fellowship-to-full-time pipeline.
• ERA Cambridge: existential risk research fellowship, Cambridge-based, feeds UK safety labs.
• PRISM AI Safety Research Fellowship: named in Refolk's top employer set alongside LASR and ERA.
• MATS (ML Alignment & Theory Scholars): several hundred scholars since 2021 per the arXiv survey. The most durable alumni graph in the field.
• UK AISI Alignment Project: £27M+ pooled funding, plus a £5.6M OpenAI grant added in February 2026, with individual grants ranging £50K to £1M. The grantee list is a public map of vetted independent researchers.

Sourcing AI alignment engineers by scanning MATS cohorts, LASR cohorts, and UK AISI grantees produces a candidate list that no ATS keyword filter will ever reproduce. The people are named, dated by cohort year, and pre-vetted by admission committees stricter than most hiring loops.

How to actually reach them without an ATS

The playbook on the sourcer's side is symmetrical to what DeepMind is doing: route around the automation, land in front of a human, and lead with a specific, non-templated first message.

The five moves that work in this cohort:

1. Search by skill and artifact, not title. Query for "interpretability," "mech interp," "evals," "RLHF safety," "red-teaming," "scalable oversight." Cross-reference with recent GitHub commits or arXiv co-authorships. This surfaces the 3,046-person skill pool instead of the 42-person title pool.
2. Anchor outbound on a specific paper, repo, or workshop. Reference NeurIPS SoLaR, ICML MechInterp, or a specific arXiv preprint. Generic "loved your work on AI safety" opens are dead in a market where DeepMind is warning candidates about LLM slop.
3. Skip the application form for the first conversation. Offer a 20-minute call with the hiring manager first, ATS entry after. This is exactly what the DeepMind form does in reverse.
4. Map fellowship alumni before frontier lab employees. LASR, ERA, PRISM, MATS, and UK AISI grantees are pre-vetted and less saturated by recruiter outreach than a Research Scientist at Anthropic who gets 40 InMails a month.
5. Be present where the community actually reads. The LessWrong "Outsider's Roadmap 2025" cites Robert Miles' YouTube channel and 80,000 Hours as the on-ramps junior researchers actually use. If your employer brand shows up there, your cold email lands warmer.

Refolk fits at step one and step four: plain-English queries like "researchers who published on interpretability at NeurIPS 2024 and 2025, based in the UK or US, not currently at a frontier lab" return a ranked shortlist in seconds, drawn from GitHub, LinkedIn, and the open web rather than a title-string index. That is the difference between fishing in a 42-person pond and fishing in the 3,046-person one.

What the memo means for anyone selling AI hiring tools

Google's own AGI safety team publicly distrusting Google's AI screener is the strongest single piece of evidence yet that automated resume filtering fails in high-signal, low-volume markets. In markets of 600 people, the false-negative cost dwarfs any efficiency gain.

Three practical implications:

• Vendors selling AI resume screening into research-heavy orgs should expect the DeepMind memo cited in every 2026 procurement conversation.
• Internal HR teams at frontier labs will move toward hybrid pipelines: automation for high-volume roles, direct-to-team forms for research roles. Expect more "supplemental form" workarounds to leak.
• Recruiters at the labs will increasingly own their own sourcing stack outside the corporate ATS, because the ATS is now a documented liability for the exact roles they most need to fill.

The DeepMind memo is not really about a form. It is about a team that ran the numbers on their own hiring funnel, decided the automation was costing them researchers they could not afford to lose, and wrote it down. Every AGI safety team recruiter should assume their competitors did the same math this week.

FAQ

How many AI safety researchers are there globally?

Roughly 600 full-time technical AI safety researchers as of 2025, per LessWrong's field-growth analysis, up from about 300 in 2022 and 80 to 120 in 2023 per an arXiv survey citing TIME. Another ~500 work in non-technical AI safety roles for a total of ~1,100. The pool is roughly doubling every three years but growing 10 to 20 points slower than capabilities research, so the ratio of safety to capabilities researchers is getting worse annually.

Why is DeepMind's AGI Safety team bypassing Google's own AI resume screener?

The team's internal document, reviewed by Bloomberg on August 10-11, 2026, states there is a "non-trivial probability" that Google's automated system will screen out qualified CVs incorrectly or too slowly. The supplemental form routes applications directly to a human on the team. Google's spokesperson confirmed the form exists but denies the underlying system filters incorrectly. The memo is significant because Google Workspace sells the same class of AI hiring tools to enterprise customers.

What is the best way to source AI alignment engineers in 2026?

Search by skill and research artifact, not job title. Refolk's index shows 42 people using an alignment title versus 3,046 with the skill listed under generic titles like Research Scientist or Member of Technical Staff, a 72x gap. Map fellowship cohorts (LASR Labs, ERA Cambridge, PRISM, MATS) and UK AISI Alignment Project grantees to catch researchers 6 to 18 months before they land at frontier labs. Lead outbound with a specific paper or repo reference, not a generic template.

Does the DeepMind memo create legal exposure for Google?

Potentially. The memo functions as an internal admission that Google's automated screener produces incorrect outcomes at a "non-trivial" rate, arriving at the same time as Mobley v. Workday (where a federal judge declined to dismiss AI-screener discrimination claims) and a 26-employee suit against Meta over AI-driven layoff selection. Enterprise HR buyers evaluating Google Workspace's hiring AI in late 2026 will almost certainly cite the memo, and plaintiffs' attorneys in future disparate-impact cases will treat it as evidence.