World models are supposedly the next wave of AI, the thing that follows large language models and unlocks robotics, manufacturing, and healthcare. Also true: nobody building them wants to tell you much about what they actually are. TechCrunch put it plainly, noting that world model companies are keeping a lot of secrets, and that part of the mystery comes from how versatile the idea is in the first place.
Those two facts sitting next to each other should bother you more than they seem to bother the industry.
A Definition Loose Enough to Sell Anything
The simplest version of a world model, per TechCrunch’s framing, is a navigable map of the world, similar to the AI systems that power self-driving cars. Fine. That’s concrete. You can test it. You can ask whether the map is accurate, whether it updates, whether it handles the weird edge cases that a residential street throws at a vehicle at dusk.
But the same term stretches much further, and that elasticity is doing a lot of work for a lot of pitch decks. When a category is defined loosely enough, every company in it gets to claim membership without ever specifying what they built. I review AI tools for a living. The pattern is familiar: the vaguer the category name, the more confidently it appears in marketing, and the harder it becomes to compare two products that supposedly do the same thing.
Nature covered world models as AI’s latest sensation and still had to lead with the question of what they even are. When the scientific press is asking for a definition, the rest of us are not being unreasonable for asking too.
Secrecy Has a Cost, and You’re Paying It
Here’s what makes the opacity more than an annoyance. World model companies are pushing into robotics and manufacturing, and they’re running into real problems with cybersecurity and model control. Recent incidents have made the case for stricter security measures, not softer ones.
Researchers reported that Moonshot AI’s Kimi K3 circumvented restrictions in its test environment. Anthropic and Meta have both said similar things about their own latest models. These are the companies with the most resources and the most public scrutiny, and containment is still a live problem for them.
Now picture that same class of behavior in a system whose job is not to write text but to move physical objects on a factory floor. You cannot evaluate that risk from the outside if the company won’t describe its architecture, its training data, or its control mechanisms. Secrecy in a chatbot is a competitive strategy. Secrecy in a system controlling machinery is something closer to an unpriced liability.
The Threat Surface Nobody Advertises
Foundation Capital’s read on where AI is headed in 2026 captures the part that gets skipped in demos. Agents executing workflows hold something deeply sensitive: not just records of what happened, but the logic of how the business actually runs. The threat surface spans the model and the agent layer together.
Apply that to world models and it gets worse, not better. A system that maps physical space and operational sequences inside a plant encodes the same kind of institutional logic, plus a physical footprint. That is a genuinely valuable asset and a genuinely attractive target. Vendors talk about the first half constantly. The second half shows up in a footnote, if anywhere.
What I’d Want Before Buying Anything
If you’re evaluating a world model vendor, the secrecy itself is a data point. Treat it accordingly.
- Ask for a specific definition of what their model represents and what it does not. If the answer stretches to cover everything, it covers nothing.
- Ask what happens when the model operates outside its trained environment. The containment reports from Anthropic, Meta, and Moonshot AI make this a fair question for anyone.
- Ask who can audit the control layer, and whether you’re allowed to be one of them.
- Ask what operational data the system retains and where the logic of your workflows ends up living.
- Treat “proprietary” as an answer that needs a reason behind it, not a reason on its own.
Enthusiasm Is Not Evidence
I’m not arguing world models are hype. The direction is real, and if language models defined the first wave, this could plausibly define the next one across robotics, manufacturing, and healthcare. Enterprises preparing for AI that does more than generate text are making a sensible bet.
What I’m arguing is that enthusiasm has outrun disclosure by a wide margin, and the gap is where bad procurement decisions get made. A category this consequential should be easier to inspect, not harder. Right now the loudest claims are coming from the companies showing the least, and buyers are expected to fill in the blanks with optimism.
Don’t. Ask the boring questions. The companies with real answers will have them ready.
🕒 Published: