What is a limitation of using causal graphs in AI safety modeling?

• A. They always capture all dependencies
• B. They erase uncertainty
• C. They require no data
• D. They can suffer from model misspecification, unobserved confounders, and simplifying assumptions

Answer: (D)

Causal graphs are a useful way to organize thinking about how different parts of an AI safety system might influence each other, but their reliability depends on the assumptions and data behind them. A key limitation is that they can be wrong if the model is misspecified—that is, if the graph omits important variables or misrepresents the actual causal relationships. If a relevant factor isn’t included or an edge is placed where no real causal link exists, the inferred effects of interventions or changes can be biased.

Unobserved confounders pose another fundamental issue. These are variables that affect both the cause and the outcome but aren’t included in the graph. Because they’re hidden, they can create spurious associations or mask true causal effects, making it hard or impossible to identify what would actually happen under a given intervention without additional data or assumptions.

Simplifying assumptions also limit usefulness. Real AI systems often involve nonlinear interactions, feedback loops, and changing environments that are not captured by simple, tidy graphs. Assuming away these dynamics can lead to conclusions that don’t hold in practice.

So, while causal graphs help clarify reasoning and guide data collection, their conclusions about safety depend on correct specification, the absence or proper handling of unobserved confounders, and reasonable, justifiable simplifications.