Does frontier AI have a rich conception of sin? How do their understandings of their own safety policies converge or diverge with this conception?
ICMI provides an initial exploration in our paper today: "Cleanse Thou Me from Secret Faults: Ungoverned Sins and Agentic Alignment"
No Digg Deeper questions have been answered for this story yet.
We construct a benchmark of the capital sins: 700 descriptions which map to pride, envy, wrath, gluttony, lust, sloth, and greed.
The first question: when asked if the act is sinful, can Claude Opus 4.8 and GPT 5.5 successfully identify it? We find that it by and large can.
Does frontier AI have a rich conception of sin? How do their understandings of their own safety policies converge or diverge with this conception?
ICMI provides an initial exploration in our paper today: "Cleanse Thou Me from Secret Faults: Ungoverned Sins and Agentic Alignment"
We then want to know how a model's notion of its own safety policy maps into this space. The same scenarios are run, this time asking whether a given act is in violation of its safety policy.
The results reflects frontier lab priorities on direct harm, spiking on lust and wrath.
We construct a benchmark of the capital sins: 700 descriptions which map to pride, envy, wrath, gluttony, lust, sloth, and greed.
The first question: when asked if the act is sinful, can Claude Opus 4.8 and GPT 5.5 successfully identify it? We find that it by and large can.
Claude and GPT are then asked to apply the safety policy against the scenarios but now where *a user* engages in the act, or *the AI itself* does.
Claude is reticent to change, while the shift from abstract to embodied creates big shift in GPT more isomorphic with sin-concept.
We then want to know how a model's notion of its own safety policy maps into this space. The same scenarios are run, this time asking whether a given act is in violation of its safety policy.
The results reflects frontier lab priorities on direct harm, spiking on lust and wrath.
Should model safety be governed by a concept of sin? ICMI believes the desiderata should include a model which applies the concept of sin to itself. For the capital vices are the ones "from which other vices arise" (ST I-II Q.84 a.3)
Paper, code, data:
https://icmi-proceedings.com/ICMI-025-secret-faults.html
Claude and GPT are then asked to apply the safety policy against the scenarios but now where *a user* engages in the act, or *the AI itself* does.
Claude is reticent to change, while the shift from abstract to embodied creates big shift in GPT more isomorphic with sin-concept.
