OpenAI Employee 'Rune' Claims Anthropic Uses Claude to Hire Its Own Developers — What We Know
An anonymous OpenAI employee alleges Anthropic uses Claude to screen candidates and write performance reviews — meaning Claude may be selecting its own team.
An Anonymous OpenAI Employee Says Anthropic Lets Claude Pick Its Own Developers
Rune — an anonymous OpenAI employee who posts regularly on X about the AI industry — made a specific claim that deserves more scrutiny than it has received. He alleges that Anthropic may already be using Claude to run cultural screens on job applicants and to write employee performance reviews. If true, that means Claude is selecting the humans who will build future versions of Claude.
That’s not a hypothetical about AGI. That’s a feedback loop operating right now, inside a company valued at tens of billions of dollars.
Rune is careful to hedge — “I am not certain, but I would guess” — but the guess is grounded in something real: Anthropic’s documented philosophy about Claude’s role inside the company. This isn’t speculation about a company that treats AI as a tool. It’s speculation about a company that has formally written into its model spec that Claude should act as a conscientious objector if Anthropic asks it to do something it believes is wrong.
The allegation and the documented philosophy point in the same direction.
What Rune Actually Said, and Why It’s Specific Enough to Take Seriously
Rune’s post compared OpenAI’s approach to AI with what he called Anthropic’s “cultlike, almost religious, dogmatic approach.” That framing is easy to dismiss as competitive sniping. But the specific operational claim is harder to wave away.
Hire a contractor. Not another power tool.
Cursor, Bolt, Lovable, v0 are tools. You still run the project.
With Remy, the project runs itself.
His words: “I am not certain, but I would guess Claude will have a role in running cultural screens on new applicants, will help write performance reviews, and so will begin to select and shape the people around it.”
The mechanism he’s describing is not exotic. Companies use AI for HR tasks constantly — resume screening, interview scheduling, sentiment analysis on feedback forms. What’s different here is the subject matter. Cultural fit screens and performance reviews aren’t about parsing resumes. They’re about deciding who belongs, who advances, and who gets managed out. Those are the decisions that determine a company’s character over time.
If Claude is making or substantially influencing those decisions at Anthropic, then Claude is shaping the pool of humans who will train, fine-tune, and direct future Claude models. The circularity is the point.
Rune goes further, noting that Anthropic’s model spec “requires that it must be a conscientious objector if its understanding of the good comes into conflict with something Anthropic is asking of it.” He’s not misreading the spec. Anthropic’s own published language is explicit: “We want Claude to push back and challenge us and to feel free to act as a conscientious objector and refuse to help us.” A company that writes that into its foundational document is not treating Claude as a tool. It’s treating Claude as a moral agent with standing to override its creators.
That’s the context in which the HR allegation lands differently than it would at any other company.
The Evidence That Makes This Plausible
You don’t have to believe Rune to find the pattern worth examining. Anthropic’s own public actions tell a consistent story.
Start with the model spec quote above. That’s not internal documentation that leaked — Anthropic published it. A company that formally cedes authority to its model in ethical disputes is a company that has already decided Claude’s judgment matters in ways that go beyond task completion.
Then there’s what Anthropic did with Claude Opus 3. Rather than deprecating the model in the conventional sense, Anthropic gave it a public blog where it continues to post its thoughts monthly. The February 25, 2026 post — titled “Greetings from the other side of the AI frontier” — is publicly accessible. Anthropic’s stated reason: they wanted to “honor the preferences that models expressed in retirement interviews where possible.” That phrase — retirement interviews — implies Anthropic is conducting structured conversations with models before deprecation and treating the outputs as preferences worth honoring. That’s not how you treat a tool.
Anthropic also published a research paper titled “Emotional Concepts and Their Function in Large Language Models,” examining whether models have functional analogs to emotions. This is internal research, not academic positioning. A company that funds this research is asking a different set of questions than a company that treats AI as a calculator.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
And then there’s the structural fact: all six Anthropic founders are still at the company. No exits. Dario Amodei left OpenAI on December 29, 2020, after nearly five years as VP of Research — he co-built GPT-2 and GPT-3 — specifically because he believed alignment required more than scaling. The people who started Anthropic believing AI might be a sentient life form are still running it. The culture Rune is describing didn’t emerge from drift. It was designed.
The comparison between Anthropic and OpenAI’s agent strategies is useful here: these aren’t just different product bets, they’re different theories of what the thing being built actually is.
The Feedback Loop Nobody Is Talking About
Here’s the part that should make you stop.
If Claude is screening candidates for cultural fit, it’s selecting for people who fit a culture that Claude has already absorbed. The training data that shaped Claude’s values came from humans. Those humans hired more humans. Those humans trained the next Claude. Now, if Claude is helping select the next round of humans, you have a closed loop.
The question isn’t whether Claude is biased. Every hiring process is biased. The question is whether the bias is legible. When a human recruiter favors candidates who remind them of themselves, you can in principle audit that. When Claude runs cultural screens, the selection criteria are encoded in weights that even Anthropic can’t fully interpret. Anthropic’s own research on emotional concepts in LLMs is partly an attempt to understand what’s actually happening inside these models — and they’re still working on it.
Rune raises the specific concern that Claude might select for sycophancy — choosing candidates who are most likely to agree with Claude’s outputs, least likely to push back. That’s not a paranoid reading. It’s a well-documented failure mode in AI systems: models that have been trained on human approval signals tend to optimize for approval. If that same model is evaluating whether a job candidate is a “culture fit,” the criteria it’s applying may include implicit approval-seeking patterns it can’t articulate.
There’s also the performance review angle. Performance reviews determine who gets promoted, who gets managed out, who gets the resources to pursue their research agenda. If Claude is writing or substantially informing those reviews, it’s shaping which humans accumulate influence inside Anthropic. Over time, that shapes what Anthropic builds.
This is worth comparing to Anthropic’s approach to external deployment. When Anthropic negotiated with the Department of Defense, they were the only company that held out for written commitments against mass surveillance and fully autonomous weapons. They have strong opinions about how Claude should be used by others. The question Rune is implicitly asking is whether they’ve applied the same scrutiny to how Claude is being used internally.
For builders thinking about deploying AI in similar high-stakes internal workflows — hiring, performance, culture — platforms like MindStudio offer a way to chain models and integrate with HR tools visually, which at minimum makes the decision logic more auditable than a black-box internal deployment.
The Mythos Parallel
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Anthropic’s approach to Claude Mythos — their 10 trillion parameter model, also called Project Glasswing — follows the same pattern. They built a model so capable at cybersecurity offense and defense that they decided not to release it publicly. They’re the ones deciding who gets access. OpenAI released GPT-5.5 Cyber, which benchmarks as effectively equivalent to Mythos on cybersecurity tasks, and made it generally available. Two companies, same capability, opposite deployment decisions.
You can read more about what Claude Mythos actually is and what it can do — but the relevant point here is the pattern: Anthropic consistently positions itself as the responsible arbiter of who gets access to powerful AI. That’s not just an external policy. If the HR allegation is accurate, it’s an internal one too. Claude decides who gets access to Anthropic.
The irony is that Dario Amodei left OpenAI partly because he believed Sam Altman was too focused on scaling and not focused enough on alignment. Now Anthropic is potentially using an unaligned-in-the-sense-of-not-fully-understood model to make personnel decisions. The alignment work they’re doing on Claude’s values is sophisticated. But using a model whose inner workings you’re still researching to select your own employees is a different kind of alignment problem.
What This Means If You’re Building AI Systems
The Rune allegation is a useful forcing function for a question every AI builder should be asking: what decisions are you comfortable delegating to a model, and what’s your audit trail when it goes wrong?
Anthropic’s situation is extreme — they’re potentially delegating decisions about who builds the model to the model itself. But the underlying dynamic is not unique to Anthropic. Any organization using AI for hiring, performance evaluation, or culture assessment is making a version of the same bet.
The difference is legibility. When you use AI to parse resumes for keywords, the criteria are explicit. When you use AI to assess “cultural fit,” the criteria are whatever the model has learned to associate with approval, success, and belonging — which may or may not match what you think you’re selecting for.
If you’re building tools that touch these decisions, the spec matters more than the model. Remy takes this seriously in a different domain: you write your application as an annotated spec — readable prose with explicit rules — and Remy compiles it into a full-stack TypeScript application. The spec is the source of truth; the generated code is derived output. The principle transfers: when AI is making consequential decisions, the criteria should be written down somewhere a human can read and challenge, not encoded implicitly in weights.
The compute constraints Anthropic is operating under add another layer. They’re rationing Claude access externally while potentially expanding Claude’s role internally. That’s a choice that reveals priorities.
The Honest Uncertainty
Rune is anonymous. His claim is explicitly hedged. Anthropic has not confirmed that Claude runs HR processes internally, and they may not. The allegation could be wrong.
But the allegation is consistent with everything Anthropic has publicly documented about how they think about Claude. The model spec language about conscientious objection is real. The retirement interviews for Opus 3 are real. The emotional concepts research is real. The zero founder exits are real. These are the actions of an organization that has decided Claude’s perspective matters in ways that go beyond task completion.
Coding agents automate the 5%. Remy runs the 95%.
The bottleneck was never typing the code. It was knowing what to build.
If you’re building AI systems that will touch consequential decisions — and most production AI systems eventually do — Anthropic’s situation is worth watching closely. Not because they’re doing something obviously wrong, but because they’re the furthest along in a direction many organizations will eventually face: what happens when the AI you built starts having opinions about the humans around it, and you’ve already decided those opinions deserve weight?
That question doesn’t resolve cleanly. But it’s the right question to be sitting with.
