AI Economy
Companies building foundational simulations of reality are revealing almost nothing about how they work — and that should concern everyone.
NewsOnScale Staff
September 21, 2026
There is a particular kind of secrecy that the technology industry has normalized so thoroughly that we rarely stop to interrogate it. Companies building foundational infrastructure — systems that other systems will depend on, that decisions will defer to, that economies will eventually route through — routinely release marketing materials before they release methodology. They announce capabilities before they disclose constraints. They invite investment before they invite scrutiny.
World model companies represent the latest and perhaps most consequential iteration of this pattern.
## What World Models Actually Are
The term sounds abstract, but the ambition is concrete. World models are AI systems trained to develop internal representations of how reality operates — how physical objects behave, how social systems respond, how sequences of events unfold over time. Unlike narrow models trained to perform specific tasks, world models are designed to simulate, predict, and reason across domains. They are, in essence, attempts to encode a generalized theory of how things work.
The commercial applications are significant: robotics, autonomous vehicles, climate modeling, drug discovery, military simulation, economic forecasting. A sufficiently capable world model doesn't just answer questions — it generates synthetic environments in which other AI systems can be trained, tested, and deployed. It becomes infrastructure.
That infrastructural role is precisely why the current posture of opacity from leading developers in this space deserves more attention than it has received.
## What Is Not Being Said
The core problem is not that these companies are lying. It is that they are not saying enough for outside observers — researchers, regulators, journalists, or the public — to evaluate what is actually being built.
Critical questions remain unaddressed in any systematic, verifiable way. What data was used to train these models, and was that data collected with appropriate consent? What physical or social domains is a given world model competent to simulate, and where does it hallucinate plausible-looking but incorrect dynamics? Who has access to these systems, under what terms, and for what purposes? What failure modes have been identified internally that have not been disclosed publicly?
These are not exotic or unreasonable questions. They are the baseline disclosures that any responsible actor building critical infrastructure would be expected to provide. The fact that world model companies have largely avoided providing them — while simultaneously attracting substantial venture capital and announcing ambitious deployment timelines — reflects a structural problem in how the AI industry has been allowed to self-govern.
## Why This Beat Matters Beyond AI
The implications extend well beyond the technology sector. World models trained on historical human data will encode the assumptions, biases, and power structures embedded in that data. When those models are used to simulate outcomes — in policy contexts, in financial markets, in military planning — the errors and distortions they carry become operationalized. Decisions get made downstream of systems whose internals no one outside a small engineering team has examined.
This is not a hypothetical risk. It is the documented trajectory of every previous generation of high-stakes algorithmic systems, from credit scoring to content moderation to predictive policing. The pattern is consistent: opacity during development, normalization during deployment, and accountability demanded only after demonstrable harm.
## The Accountability Gap
Regulators in the European Union have begun developing frameworks that would require higher disclosure standards for foundational AI systems, but enforcement mechanisms remain immature and jurisdictionally limited. In the United States, no equivalent statutory framework exists. The result is that the most powerful AI systems currently under development operate in a disclosure environment roughly equivalent to that of a consumer app.
The world model sector is not monolithic, and some researchers within it are pushing for greater transparency through academic publication and open benchmarking. But voluntary disclosure from individual researchers is not a substitute for institutional accountability from the companies that control deployment.
The question worth asking is straightforward: if these systems are as capable and as consequential as their developers claim, why would comprehensive disclosure be the thing they are unwilling to provide?