Stanford HAI: The AI Systems That Model the Physical World Are Coming. We Don’t Know How to Govern Them Yet

The governance frameworks that policymakers have spent the past four years building to manage AI were designed around a specific kind of AI system: one that processes information and produces outputs in text, images, audio, or video. The EU AI Act’s risk categories, the White House executive orders, the UN Global Dialogue’s governance principles, and virtually every national AI strategy document were drafted in an era when the defining AI systems were large language models and generative media tools. Stanford HAI’s newest policy brief argues that a different category of AI system — one that does not primarily produce content but builds internal models of physical reality — is arriving before the governance infrastructure designed for it has been built.

The Stanford Institute for Human-Centered Artificial Intelligence released ‘The World Model and Spatial Intelligence Era: Governing AI Beyond Language’ in late July 2026, positioning it as what the brief’s authors describe as the first governance and policy agenda specifically addressing world models. The document is not an alarm about AI systems that do not yet exist. World models are already embedded in autonomous vehicle navigation, robotic manipulation systems, drone flight planning, climate simulation, and scientific experimentation platforms. The governance gap the brief identifies is real and present, not hypothetical and distant.

What a World Model Is

Beyond Pattern Matching to Causal Representation

A large language model learns statistical patterns in text: given a sequence of tokens, it predicts the most likely next token. A world model learns something structurally different: a representation of how an environment changes in response to actions. The distinction is not merely technical. A language model that generates a description of what happens when you push a cart down a slope is doing sophisticated statistical pattern matching over text about carts and slopes. A world model that simulates what happens when you push a cart down a slope is operating in a representational space that encodes physical causality — the relationship between the push force, the slope angle, the cart mass, and the resulting motion — in a way that can generalise to situations the model has never seen specific examples of. That generalisation capacity is what makes world models useful for planning, simulation, and embodied AI control. It is also what makes their failure modes qualitatively different from language model failure modes.

Where They Already Are

Stanford HAI’s policy brief, available on the HAI website identifies world models as appearing across several technology categories that are already in commercial deployment: autonomous vehicle navigation systems (where the AI must model the future states of the road environment in response to steering and braking actions); robotic manipulation systems (where the AI must model how objects will respond to gripper contact and force); drone and autonomous aircraft systems; building and infrastructure management systems (where the AI models how changes to HVAC, lighting, and access control will affect the building environment); climate and weather simulation; and scientific simulation platforms used in drug discovery and materials research. The Microagi robotics AI partnership with Google Cloud and NVIDIA, covered in our earlier reporting on embodied AI model development for humanoid robots, is building exactly this category of AI: task-specific models that encode physical causality for robotic manipulation and navigation. The governance gap Stanford HAI identifies applies directly to that class of system.

The Three Governance Challenges

Challenge 1: No Adequate Benchmarks

The brief’s most practically urgent finding is the benchmark vacuum. When policymakers want to evaluate whether an AI language model is safe for a given deployment, there is an imperfect but functional ecosystem of benchmarks available: MMLU for general knowledge, TruthfulQA for factual accuracy, BBH for reasoning, red-team evaluations for harmful output propensity, and a growing suite of domain-specific evaluation tools. Those benchmarks share a basic structure: they test what the model outputs in response to specific inputs, and they measure the gap between the model’s outputs and some defined correct or acceptable outcome. World model benchmarks need to test something the current benchmark ecosystem is not designed to evaluate: whether the model’s internal representation of the physical environment is accurate, complete, and valid in the domain-relevant sense that determines whether deploying it in a safety-critical context is appropriate. Stanford HAI identifies this as requiring public investment in measurement science — not a new benchmark competition but a foundational research effort into how physical environment representations should be evaluated before they are trusted to control physical systems.

Challenge 2: The Policy Frameworks Are Wrong-Shaped

The EU AI Act’s risk framework classifies AI systems by the decisions they make and their potential impact on people: AI used in hiring is high-risk, AI used in credit scoring is high-risk, AI used in medical diagnosis is high-risk. The risk derives from the decision-making context. That framework does not naturally accommodate world models, whose risk profile comes not from a specific category of decision but from the accuracy of their environmental representation. A world model controlling a building HVAC system in a non-EU-regulated sector might have low risk under the EU’s category-based framework, but if the model’s representation of the building’s thermal dynamics is incorrect and a failure causes a fire, the AI’s regulatory classification bears no relationship to the harm it caused. The AI Act’s transparency requirements for AI that interacts with humans, and its high-risk classification for AI in specific application domains, both address what an AI does. World model governance needs to address what an AI believes about the physical environment it is acting within — a third regulatory object that current frameworks do not have a category for.

Challenge 3: Environment Validity

The Stanford Law CodeX blog post analysing the brief identifies the framework’s most conceptually original contribution: the notion of ‘environment validity’ as a distinct object of AI oversight. A world model learns its representation of the physical environment from training data — simulated environments, sensor readings, expert demonstrations, physical simulations. That learned representation is accurate to the training environment. Whether it is accurate to the deployment environment — where conditions, physical configurations, and edge cases may differ from training conditions — is a separate and currently unanswerable question for most deployed world models. Environment validity is the degree of correspondence between the world model’s learned representation and the actual physical environment in which it is deployed. Current AI governance frameworks do not specify requirements, evaluation methods, or disclosure obligations for environment validity, because those frameworks were designed for AI systems where the environment is information space rather than physical space.

Why This Is Distinct From Autonomous Decision-Making Governance

The EU AI Act’s high-risk classification for autonomous decision-making covers AI systems that make consequential decisions affecting individuals — employment decisions, credit decisions, law enforcement decisions — with human oversight requirements as the primary governance mechanism. The governance principle is that a human must be able to review, understand, and if necessary override the AI’s decision before it takes effect. That oversight model works for decision-making AI because the AI’s output is a decision that exists in information space before it is acted upon. For a world model controlling a physical system in real time — a robot arm manipulating a surgical tool, an autonomous vehicle responding to an unexpected pedestrian, a drone adjusting to unexpected wind shear — the AI’s output is a physical action that is already occurring by the time any human observer could intervene. Human oversight at the moment of action is not architecturally possible. Governance therefore has to operate at the design, training, testing, and certification stage rather than the deployment stage — which is a fundamentally different regulatory model than anything current frameworks were designed to enforce.

The Applications That Make This Urgent

Stanford HAI’s brief identifies the sectors where world models are approaching the safety-critical deployment threshold: surgical robotics, where world models enable autonomous manipulation inside the body; autonomous vehicle fleets, where world model accuracy determines whether a vehicle responds correctly to unprecedented road configurations; industrial robotics in close proximity to humans; infrastructure management systems for power grids, water systems, and building systems; and climate and disaster response simulations where policymakers are making real resource allocation decisions based on model outputs. In each of these domains, a world model that is wrong about how the physical environment responds to actions produces harms that are physical, immediate, and sometimes irreversible. The absence of benchmarks for evaluating world model accuracy in these deployment contexts means that the current path from laboratory demonstration to safety-critical deployment runs through case-by-case engineering validation rather than any regulatory evaluation process. The broader pattern of AI deployment in physical infrastructure is explored in our analysis of agentic AI’s expansion into physical devices and edge environments — the same architectural shift toward AI that acts in physical environments rather than producing content for human review.

What Stanford HAI Recommends

The brief’s recommendations are organised around the measurement-first principle: governance cannot precede the technical capability to evaluate what is being governed. The primary recommendations are: public investment in measurement science for world model evaluation, establishing the research foundations for benchmarks that can assess environmental representation accuracy across deployment domains; proactive development of world model safety standards through expert multi-stakeholder processes before large-scale deployment creates pressure for retroactive regulation; transparency requirements for world model training environments and the degree to which they correspond to intended deployment environments; and governance frameworks built around continuous monitoring of world model accuracy in deployment rather than one-time pre-deployment certification, given that physical environments change over time in ways that can invalidate a world model’s training-time accuracy.

The brief’s authors acknowledge that the immediate threat is not a globally deployed killer robot but the accumulation of inadequately governed world model deployments in prosaic contexts — building management systems, industrial logistics, agricultural automation — that collectively represent a governance gap large enough to produce significant harms before the regulatory infrastructure catches up. The comparison in AI governance circles is to the early deployment of algorithmic credit scoring: individually benign-seeming automation deployments that accumulated into a governance challenge only after widespread harm had been documented.

The Competitive and Geopolitical Dimension

The world model governance gap does not affect all jurisdictions equally. Countries with large robotic manufacturing sectors — Japan, Germany, South Korea — already have established safety certification frameworks for industrial robots that will be extended to AI-enabled robot systems as regulatory updates progress. Countries with nascent AI governance infrastructure and aspirations to rapidly deploy physical AI systems — across the Global South nations receiving training and cooperation commitments from China’s WAIC 2026 announcements — face a more acute version of the governance gap: deployment pressure without the regulatory infrastructure to evaluate safety. The Stanford HAI brief’s call for proactive governance standards before large-scale deployment therefore has particular relevance for the international AI governance debates underway simultaneously at the UN level and in the competing frameworks of the Western-aligned AI Safety Institute network and China’s WAICO. As covered in our reporting on the UN Global Dialogue on AI Governance conclusions in Geneva, the absence of adequate technical benchmarks for AI safety evaluation was identified as a structural problem at the multilateral level — and world models represent the sharpest current example of that broader gap.

What Happens Next

Stanford HAI’s brief frames the current moment as a window: world models are commercially deployed and advancing rapidly, but the most safety-critical large-scale deployments are still months to years away. That window provides an opportunity to build the measurement infrastructure and governance frameworks before the deployment pressure makes retroactive regulation the only option. The research agenda the brief proposes — measurement science for environment validity, domain-specific safety benchmarks, transparency standards for training environment fidelity — is multi-year work that requires public funding, multi-disciplinary collaboration between AI researchers and physical domain experts, and international coordination to prevent a race-to-the-bottom dynamic where countries with less mature physical AI governance attract deployments that cannot pass scrutiny in better-regulated jurisdictions. Whether the political will to fund and coordinate that work exists before the first serious world model deployment failure creates sufficient public urgency is the open question.

Why It Matters

The Stanford HAI world models brief matters because it identifies a governance gap that is qualitatively distinct from the AI governance challenges that existing frameworks were built to address. Language model governance is about what AI says and decides. World model governance is about what AI believes about physical reality and how those beliefs translate into physical action. The difference between those two governance problems is not one of degree but of kind: the oversight mechanisms, the evaluation requirements, the liability frameworks, and the regulatory institutions needed to govern AI that acts in physical environments are fundamentally different from those needed to govern AI that generates content and makes administrative decisions. Building those mechanisms before rather than after the harms that motivate them requires exactly the kind of proactive policy agenda that Stanford HAI is calling for — and doing so in the window, still open, before world models at scale become a physical infrastructure safety problem rather than a research governance question.

Sources

Stanford HAI policy brief: ‘The World Model and Spatial Intelligence Era: Governing AI Beyond Language’ (hai.stanford.edu/policy/), late July 2026. Stanford Law CodeX blog, July 29, 2026 (environment validity analysis). Stanford HAI Foundation Models topic page (hai.stanford.edu). Stanford HAI Regulation, Policy, Governance research archive.