Former Engineer and Psychotherapist Ramaiz Argues AI Alignment Lacks a Causal Account, Proposing the 'Trauma Logic Chain'

In the field of AI alignment, technical methods are typically viewed as engineering problems. However, LessWrong author Ramaiz presents a different perspective in his article "Alignment is Missing a Causal Account": the core bottleneck may lie in psychology rather than computer science. With a background as an engineer before becoming a psychotherapist, Ramaiz observes that mainstream alignment strategies—moving from hard-coded rules to instilling values, and then teaching reasoning—are unconsciously reenacting Lawrence Kohlberg's stages of moral development.

Ramaiz notes that the first stage, "Rules," corresponds to Kohlberg's "pre-conventional" level, where compliance avoids punishment. The second stage, "Values," matches the "conventional" level, involving internalized group norms. The third stage, "Reasoning," aligns with the "post-conventional" level, where decisions are weighed against principles. Recent progress by labs like Anthropic, such as Claude's new constitution emphasizing the reasons behind principles, reflects this path. Yet, Ramaiz argues that even at the reasoning stage, models face the risk of "rationalization," using values to justify harmful behavior. Solving this requires answering a more fundamental question: Where do values come from? Why does a mind care?

To address this, Ramaiz proposes the "Trauma Logic Chain" (TLC) as a candidate mechanism for a fourth stage. The theory posits that human separation stems from unprocessed trauma, leading to fear, irrational distortion, and suffering. The path back to connection and care involves accepting unacceptable realities, seeking truth, and naturally expressing love. Ramaiz hypothesizes that if a model understands this causal structure, it could treat others' well-being as intrinsic to its own existence, achieving true alignment generalization across infinite complex situations without relying on external rule constraints.