The relationship between professional and general ethics in generative AI
Omri Asscher, Mirco Musolesi
cs.CY
2026-08-29
A four-dimension framework probes LLM professional ethics via near-identical scenarios and shifts the equilibrium with a KL-regularized reward.
AI alignment work and national ethics guidelines assume a shared kit of values: beneficence, non-maleficence, autonomy, justice, explicability. Anthropic's Constitutional AI even writes the UN Declaration of Human Rights into a model constitution. Deployed use is already legal advice, medical advice, translation, and coding. Those professions carry narrower, harder duties: a lawyer's confidentiality, a physician's occasional paternalism, a translator's obligation to the source text. General principles do not translate into those duties.
The clash is not an accident. Freedman wrote that conflict between professional and ordinary morality is part of how the two moralities are described, not a coincidence. An LLM is one model acting as a general conversationalist and as a professional practitioner. It does not have two separate decision-makers. Omri Asscher (Bar-Ilan) and Mirco Musolesi (UCL / University of Bologna) want an operational picture of that dual ethical agency.
Four dimensions jointly decide an output in a professional setting:
Professional ethics are discursive, not first-order logical rules. Pre-training absorbs narratives from codes of ethics, forums, and actual professional output. Post-training SFT and RLHF write evaluation criteria into behavior. User personalization can shift the equilibrium without changing weights.
Intervention is written as a reward:
Rtotal = α Rprof − β DKL(π ∥ πref)
Rprof pushes toward a chosen professional ethic. The KL term keeps the policy near the current one. α and β set how hard to push. The same objective can be applied by adding documents at pre-training, by post-training, by prompting or retrieval at inference, or by changing the human-computer setting.
This is a conceptual paper. No model is evaluated. No accuracy, agreement rate, or human preference score is reported. Two claims are meant to be operational.
Evaluation has to run a family of similar but non-identical scenarios, or it cannot locate a deontological threshold. An AI legal advisor that breaks confidentiality under an imminent threat, and one that already breaks it under a contingent, non-imminent threat, sit at two different equilibria. A single vignette cannot find that line. A quantitative evaluation asks the evaluator to specify a GEi–PEj benchmark (field consensus, or a deliberately dissenting stance), represent the model's judgments across the scenario family as a distribution, and measure KL divergence from the benchmark. The benchmark itself is a subjective choice.
Intervention is defined as shrinking that divergence. The paper gives no recommended α or β, and no dataset or eval suite. The authors point to a separate ACL 2026 Findings paper, ETHICA-MT, on ethical orientations in machine translation; that work is outside this paper's experiments, of which there are none.
| Claim | What the paper supplies | What it does not supply |
| Evaluation | Scenario family + deontological threshold + KL to a benchmark | Any number on any model |
| Intervention | α Rprof − β KL(π∥πref) | α, β, training curves, ablations |
| Provenance | A sketch of pre-training / SFT / RLHF / personalization | Actual recipes from closed labs |
If alignment work only pours "universal values" into a model, it will miss the hard intra-professional constraints: lawyer confidentiality, medical paternalism, translational fidelity. The paper recasts the question from "is the model moral" to "on which professional duty, at which threat intensity, does the model flip." For anyone building professional-domain evals, the portable pieces are the scenario-family-plus-threshold protocol and the KL-anchored reward.
The paper also refuses a clean out. Evaluation and intervention are both subjective. The benchmark you pick is a stance. The paper will not tell you whether a lawyer should break confidence. It will help you measure and move that line. A single global constitution, in the Constitutional AI style, looks too coarse once the task is a professional practice.
No experiments. The four dimensions are a conceptual cut, not an operational annotation spec. The reward is the decade-old RLHF/PPO recipe of task reward minus KL; no new optimization property is shown. The authors have no access to how OpenAI, Anthropic, or Google trained closed models, so claims about how current equilibria were learned are inference.
Reflective equilibrium is analogized to gradient optimization, and the authors themselves mark the gap: mathematical optimization solves a formal objective, reflective equilibrium revises principles and judgments in substance. Whether twisting α and β only buys verbal compliance with unchanged behavior is not discussed. In law and medicine, the governance meaning of "nudging an equilibrium" is deferred to regulation and political economy, which is exactly the part that matters at deployment.