Re-evaluating the Smoking Lesion: A Convergence of EDT and CDT?
Coverage of lessw-blog
In a recent post, lessw-blog challenges the pedagogical utility of the 'Smoking Lesion' thought experiment, suggesting it fails to effectively differentiate Evidential Decision Theory from Causal Decision Theory.
In a recent analysis, a contributor to lessw-blog revisits a cornerstone of decision theory: the "Smoking Lesion" thought experiment. For decades, this scenario has been utilized as a primary wedge to demonstrate the superiority of Causal Decision Theory (CDT) over Evidential Decision Theory (EDT). The author argues, however, that this distinction may be illusory, prompting a re-examination of how we define rational agency.
The Context: Defining Rationality for Agents
Decision theory serves as the bedrock for autonomous agent design and AI safety. To build rational AI, developers must rigorously define what "rational" means in mathematical terms. Two primary frameworks dominate this landscape:
- Causal Decision Theory (CDT): The agent asks, "What is the causal consequence of my action?" This is the standard scientific approach, focusing on physical cause and effect.
- Evidential Decision Theory (EDT): The agent asks, "What does my action imply about the state of the world?" This focuses on conditional probabilities and evidence.
While these theories align in most real-world scenarios, philosophers and mathematicians use edge cases to test their robustness. The "Smoking Lesion" is one such case. It posits a genetic defect that causes both cancer and a desire to smoke, though smoking itself does not cause cancer. CDT correctly advises the agent to smoke (gain pleasure, no causal harm). EDT is traditionally thought to advise abstaining, as the act of smoking is interpreted as "bad news"-evidence that the agent possesses the fatal lesion.
The Argument: A Failure to Distinguish
The post contends that the standard interpretation of EDT in this scenario is flawed. The author suggests that if an agent is sufficiently introspective-aware of its own desires and internal states-the "bad news" is received via the desire to smoke, not the act itself.
If the correlation between smoking and cancer is screened off by the agent's knowledge of their own inclination, the act of smoking provides no additional negative evidence. Consequently, a sophisticated EDT agent would converge with a CDT agent and choose to smoke. This implies that the Smoking Lesion is not the definitive "tie-breaker" it is often portrayed to be in academic literature.
Why This Matters
For those involved in AI alignment and agent foundations, this analysis suggests that our standard battery of tests for decision theories may need updating. If the Smoking Lesion fails to separate EDT from CDT, researchers must rely on other paradoxes, such as Newcomb's Problem, to understand the divergence between causal and evidential reasoning. Ensuring we have the right theoretical tools is critical for predicting how advanced AI systems will behave in complex environments where correlation and causation are difficult to disentangle.
We recommend reading the full argument to understand the nuances of the probability screening involved.
Read the full post on LessWrong
Key Takeaways
- The 'Smoking Lesion' is a classic thought experiment intended to prove Causal Decision Theory (CDT) superior to Evidential Decision Theory (EDT).
- Standard doctrine holds that EDT fails this test by recommending against smoking due to a 'bad news' correlation, despite smoking having no causal link to cancer in the scenario.
- The author argues that a sophisticated EDT agent, aware of its own desires, would attribute the cancer risk to the desire (the lesion) rather than the act, leading it to smoke.
- If EDT and CDT both recommend smoking, the Smoking Lesion fails as a diagnostic tool for distinguishing between these two frameworks.
- This re-evaluation impacts how researchers test and validate decision-making frameworks for rational AI agents.