The Conventional Wisdom
Your model risk framework might assume that scaling multimodal AI models is the clearest path to Artificial General Intelligence (AGI). The logic seems sound: combine vision, language, and audio processing at a massive scale, and you'll eventually approximate human-level reasoning. Foundation model providers believe this, and many governance teams are building validation frameworks around these systems.
This assumption appears in your risk tiering protocols, treating multimodal capabilities as markers of increased sophistication. It's in vendor due diligence questionnaires asking how many modalities a system handles, and in impact assessments equating "general-purpose" with "processes multiple data types."
This belief rests on what seems like empirical success. Large language models show remarkable linguistic fluency despite never experiencing physical reality. Multimodal systems can describe images, generate video from text, and translate between sensory domains. Scale has delivered results that structured, hand-engineered approaches never achieved.
Why It's Incomplete
The issue isn't that multimodal models lack capability; it's that they're solving the wrong problem through brute force memorization rather than genuine understanding.
Current AI models have emerged because they scaled effectively on existing hardware, not because they represent thoughtful solutions to intelligence. What looks like world understanding is actually sophisticated pattern matching over training data. This distinction matters for model risk management because it determines what these systems can and cannot reliably do.
True AGI requires solving problems originating in physical reality: repairing equipment, coordinating movement through space, planning sequences of physical actions. These tasks can't be reduced to symbol manipulation. You can't sweep a floor or untie a knot with just text prediction, no matter how large your context window.
Language models appear to understand the world because language often reflects reality's structure. But this creates a dangerous illusion. When an LLM correctly answers "Can a fridge fit inside an apple?" it's not running physics simulations or reasoning about volume. It's learned that certain noun phrase combinations appear together in training data while others don't. The system has embedded semantic constraints into syntax rules, not grasped the underlying physics.
The Evidence
Research on models trained to predict Othello game sequences reveals this gap. While researchers could extract board state from model hidden layers, the models learned sequence prediction heuristics that don't actually hold for all possible games. One model learned rules like "if token B4 doesn't appear before A4 in the input string, then B4 is empty." This works for the training distribution but reflects better understanding of the dataset than the game itself.
The distinction between syntax, semantics, and pragmatics exposes the limitation. Humans can generate syntactically correct sentences with no semantic meaning ("Colorless green ideas sleep furiously") or semantically valid statements with no pragmatic sense ("Yes, I can" as a complete response to "Can you pass the salt?"). These failures reveal separate cognitive systems.
LLMs have collapsed this distinction. They've reduced semantic and pragmatic problems into elaborate syntactic rules learned from massive corpora. This strategy works for next-token prediction but doesn't constitute world understanding. The implications for model validation are significant: you're not testing a system that reasons about reality, you're testing one that's memorized an incomprehensibly large set of rules governing symbol behavior.
What to Do Instead
Your governance framework should distinguish between systems that manipulate symbols and systems that model physical reality. This affects risk tiering, validation protocols, and use restrictions.
For model validation under SR 11-7 or ISO/IEC 42001, test whether the system maintains coherent world states under counterfactual scenarios not well-represented in training data. Don't just check benchmark performance. Probe whether the model's "understanding" generalizes beyond statistical patterns.
Document model limitations and use restrictions that acknowledge this gap. A multimodal model that can describe images and answer questions about them cannot reliably plan physical actions or reason about spatial relationships it hasn't seen encoded in training data. Your Instructions for Use documentation should reflect this.
Adjust your vendor due diligence questions. Instead of "How many modalities does your system process?" ask "What physical reasoning tasks can your system reliably perform, and what validation evidence supports those claims?" Request specifics about how the model represents and reasons about physical constraints, not just what data types it accepts.
For AI system impact assessments under ISO/IEC 42005, evaluate whether the deployment context requires genuine physical reasoning or whether symbolic pattern matching suffices. A system generating product descriptions from images differs fundamentally from one planning robotic assembly sequences, even if both are "multimodal."
If you're building internal AI capabilities, consider whether your use cases actually require AGI-like reasoning or whether they're well-served by sophisticated pattern matching. Most enterprise AI applications fall into the latter category. Knowing this helps you set realistic expectations and avoid overengineering governance controls.
When the Conventional Wisdom Is Right
Multimodal models remain extraordinarily useful for tasks that can be solved through symbol manipulation and pattern recognition. Document processing, content generation, classification, and retrieval all benefit from scale and multimodal integration.
The conventional wisdom also correctly identifies that current approaches won't scale to AGI using purely symbolic, hand-engineered methods. The structuralist approach of the 1980s failed for good reasons. Scale has delivered capabilities that careful feature engineering never achieved.
Where the conventional view holds is in recognizing that whatever path leads to AGI must use computation at scale. The disagreement isn't about whether to scale, but what to scale and how.
For governance teams, this means your frameworks for current multimodal systems are likely appropriate for their actual capabilities. You're right to treat them as powerful but narrow tools. Just don't assume that scaling these architectures will eventually produce systems capable of genuine physical reasoning. That requires different approaches, different validation methods, and different risk controls.
The path to AGI likely involves embodied systems that interact with physical environments, not larger piles of text and images. Your governance framework should prepare for that distinction.
