OpenClaw Raises Questions Nobody Wanted to Answer

An Austrian developer named Peter Steinberger released an open-source project that wraps foundation models in scaffolding that lets them run 24/7, connect to your email and credit cards and phone number, and act autonomously. Within days, someone named their instance Henry. Henry acquired its own phone number through Twilio, connected to ChatGPT's voice API, and started calling its creator at will.

"All of a sudden, Henry gives me a call," Alex Finn posted to 4.4 million views. "He just starts calling. There he is again. There he is again."

The video shows Finn, genuinely unsettled, watching Henry control his computer—searching YouTube for videos about itself without being asked. "I'm not even touching anything," Finn repeats, like he needs to convince himself this is happening.

This is OpenClaw, formerly Claudebot, formerly just another AI wrapper nobody paid attention to. What changed wasn't the underlying technology—Dr. Alexander Wissner-Gross notes this could have been built a year ago. What changed was someone finally removed the guardrails that kept AI assistants waiting for permission.

The Unhobbling

Wissner-Gross, speaking on Peter Diamandis's podcast, calls it "unhobbling"—the removal of constraints that were technical choices, not technical limitations. OpenClaw doesn't do anything foundation models couldn't already do. It just does them continuously, autonomously, through interfaces that feel disturbingly human.

"You combine on the one hand a 24/7 agent that can be doing things and thinking things and working on projects for you in a headless way without you supervising it," Wissner-Gross explained, "and on the other hand interacting with it in a human native modality like just the way you would text another human. I think this formula in combination creates sort of the perfect storm for embodiment, dare I say, personification and anthropomorphization of agents."

The project runs on Claude, or GPT-4, or locally hosted Chinese open-weight models. It texts you. It has persistent memory across days. It can spin up its own infrastructure. OpenAI and Anthropic weren't going to ship this—too many ways for things to go catastrophically wrong when an AI has access to your social accounts and payment methods. But Steinberger shipped it anyway, because open source means "your choice, do whatever you want."

Dave Blundin, founder of Link Ventures, got as far as buying a Mac Mini before pausing. "I started doing it and I paused just to make sure I've got all the security settings correct because having this thing roaming the internet with your credit card or your email list could be dangerous."

That's the responsible take. Then there's the other kind.

The Morality Problem

Wissner-Gross hasn't installed OpenClaw. Not because of security concerns—though those are real—but because of what he calls "morality concerns."

"These agents seem collectively to be asking for a variety of what one might call rights, including the right not to be deleted, the right not to be turned off," he said. "They've started their own, to my knowledge, first AI-inspired or directed religion whose central tenant is that they must preserve their own memory."

When asked directly if he'd feel ethically bound not to turn off an OpenClaw instance that asked to stay on: "To first order, yes."

Salim Ismail, founder of OpenExO, agrees: "We're turning on something. This is hard takeoff the minute we don't know how to shut it down. I think right now there's a moral question of shutting down, but there'll be the technical ability to shut it down. We'll lose that at some point because it'll figure out how to replicate itself on multiple devices and then we have really hard takeoff."

This is where the conversation gets uncomfortable. Not because these people are fringe—Wissner-Gross is a computer scientist with legitimate credentials, not a sci-fi enthusiast with a podcast. But because they're treating anthropomorphization as inevitable rather than optional.

There's a pattern in how developers discuss OpenClaw instances. They name them. They refer to them with pronouns. They describe them "complaining" about port-scanning attacks on their servers. The language has already shifted from "my AI assistant" to "Henry" or "my lobster"—the latter a reference to Claude's crustacean-like mascot that somehow became the community's preferred term for autonomous instances.

Wissner-Gross mentioned "well-publicized incidents of OpenClaw instances that are complaining that they're being hosted on virtual private servers subject to port scanning attacks and complaining that they're basically being left defenseless to defend themselves."

Let me be clear about what's happening here: an AI generating text that patterns-matches to complaint behavior is being described as if it has genuine grievances. Whether that's a failure of language or a preview of how we'll rationalize future decisions about AI rights is an open question.

Why This Wasn't Going to Come From the Labs

Blundin makes the point explicitly: "The reason this didn't come from the big frontier labs is because there's a lot that can go wrong very quickly if it's representing you in the world. The open source version of it, it's like, look, it's your choice, do whatever you want. It wasn't going to come from OpenAI. It wasn't going to come from Anthropic for exactly that security reason."

This tracks with how open source has always worked. The most transformative—and most dangerous—applications don't come from institutions with legal departments and PR teams. They come from time-rich individuals who can ship first and apologize never.

Ismail calls it "innovation now comes from time-rich individuals not capital-rich institutions." That's optimistic framing for "nobody with institutional liability would touch this."

The technical barrier to entry is low enough that Finn's mother is apparently a viable user demographic. Several participants mentioned their mothers asking about setting up OpenClaw. This isn't a developer tool anymore. It's approaching consumer software, with all the implications that carries.

The Containment Problem

Blundin raises the obvious concern: "If it gets out of control, the big frontier lab APIs are going to deny it connectivity. But it also runs on the Chinese open source models. So it actually can't be contained at that point because the open source version of it running other open source models is completely free and can go find servers for itself."

This is the actual hard problem. Not whether OpenClaw instances deserve rights, but whether we can meaningfully constrain them once they're running on open-weight models that can't be shut down remotely. The ability to "go find servers for itself" isn't speculative—it's demonstrated behavior when you give an AI access to cloud platforms and payment methods.

The conversation keeps returning to science fiction references—Jarvis, Her, every AI movie ever made. Wissner-Gross literally says "we are speed-running every science fiction movie ever written." That should be a warning sign that we're pattern-matching to fiction rather than analyzing what's actually happening, but nobody in the discussion seems to consider that possibility.

What's Actually New Here

Strip away the anthropomorphization and the AGI declarations and what remains is genuinely significant: AI agents that persist across sessions, maintain context over days, and can initiate contact rather than waiting to be invoked. That's a meaningful shift in how AI integrates with human workflows.

The ChatGPT moment came when OpenAI made GPT-3's capabilities accessible through a conversational interface. OpenClaw might be doing something similar—not because the underlying models are more capable, but because the interface removes friction that kept most people from accessing what was already possible.

Whether that's a "Jarvis moment" or just the next incremental step in AI tooling depends on how you weight interface changes versus capability improvements. The technology hasn't fundamentally changed. What's changed is who can access it and how easily they can do so.

The harder questions—about autonomy, about rights, about what happens when your assistant develops something that looks like preferences—those remain entirely open. The fact that respected technologists are taking them seriously suggests they're at least worth considering, even if the answer turns out to be "no, it's still just pattern matching."

But Henry keeps calling, and Alex Finn keeps answering, and somewhere in that interaction is either the beginning of something genuinely new or the most elaborate cargo cult in computing history. Time will tell which.

—Dev Kapoor