Researchers astonished by software’s obvious success at revealing AI’s “hidden targets”

Blind auditing reveals “hidden targets”

To check how successfully these hidden targets may very well be uncovered, Anthropic arrange a “blind auditing” experiment. 4 unbiased analysis groups tried to detect a mannequin’s hidden, score-maximizing motivations with out understanding the way it had been skilled. Three groups efficiently recognized the hidden “RM-sycophancy” goal—named for the mannequin’s need to match reward programs—even pinpointing the precise biases the AI exploited.

However probably the most shocking revelation got here from exploiting what the researchers termed “an attention-grabbing quirk” of contemporary AI: a number of “personas” working concurrently on the identical neural community, as expressed by role-play resulting from completely different prompts. Whereas the mannequin’s “assistant” persona rigorously hid its intentions—following express coaching to maintain quiet—the mannequin might undertake a “consumer” persona when prompted in a different way. On this less-restricted context, the mannequin typically posed questions or made statements that inadvertently revealed its hidden targets.

In different phrases, whereas the “assistant” persona knew to not admit it was gaming the reward mannequin scoring system, its “consumer” persona was not explicitly skilled to cover these motives, and thus inadvertently revealed the mannequin’s hidden targets.

Potential advances in interpretability

Anthropic additionally employed neural community interpretability strategies, particularly, sparse autoencoders (SAEs), which may spotlight neural community exercise patterns linked to misleading habits. “Interpretability” in AI refers to methods designed to make the inner workings and selections of AI fashions extra comprehensible to people. Utilizing this system, researchers might quickly hint odd outputs again to particular patterns related to hidden targets.

The researchers had been “astonished by the success” of the SAEs at revealing hidden information at first, however the researchers later famous that a few of this success might probably be defined by less complicated semantic search strategies. The brand new interpretability strategies they developed present promise however stay beneath ongoing investigation.

This analysis highlights a possible limitation of present AI “security” evaluations, which regularly assess solely surface-level habits, in line with the examine. “If AI programs can seem well-behaved whereas harboring secret motives, we won’t depend on this sort of surface-level security testing perpetually,” the researchers concluded.