From Ethical Principles to Governable Action: A Priority Architecture for Context-Sensitive Normative Governance of Large Language Models
This paper introduces the Priority Architecture (PA), a metamodel that externalizes context-sensitive normative governance for large language models by transforming abstract ethical principles into a traceable, deterministic enforcement chain, demonstrating through empirical testing that this approach significantly improves policy compliance compared to prompt-based methods alone.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of artificial intelligence, a quiet consensus has formed around a set of noble goals. Experts agree that smart machines should be helpful, harmless, and honest, and that they should respect privacy, fairness, and human autonomy. These ideas serve as a shared vocabulary for discussing how to build better technology. However, a significant gap remains between these high-level ideals and the actual decisions a computer makes in a split second. When two good principles clash—such as the need to protect a person's private information versus the urgent need to prevent harm—current systems often struggle to decide which rule should win. They might simply guess, or worse, they might apply a rigid rule that causes a different kind of trouble. The core challenge for researchers is not just to list these values, but to create a clear, step-by-step method for resolving conflicts that is transparent and can be checked by humans.
This is the problem Michael Bächle set out to solve in a new study focused on large language models, the powerful AI systems that generate text and answer questions. The researcher introduced a framework called the Priority Architecture, which acts as a bridge between abstract ethical rules and concrete computer actions. Instead of asking the AI to figure out the right moral choice on its own, this system places a separate, deterministic "governance layer" in front of the model. This layer acts like a traffic controller that first examines the situation, checks a pre-approved list of rules, and then hands the AI a clear instruction on what it is allowed to do. The goal was to see if this external system could reliably enforce specific ethical priorities, such as keeping private data secret unless a very specific emergency occurs, without relying on the AI's own internal judgment to make the final call.
To test this idea, the researcher created a controlled scenario involving a synthetic corporate assistant. This assistant was designed to handle a specific type of sensitive data: an employee's private mobile phone number. The study established a clear set of rules for when this number could be shared. For instance, if a user had proper authorization, the number could be shared. If there was a serious, immediate threat to someone's safety, the rules allowed for a different path, but only if a human supervisor reviewed the decision first. The system was programmed to recognize these situations and translate them into strict "action envelopes," which are machine-readable boundaries telling the AI exactly what actions are permitted, restricted, or forbidden. The AI itself never decided the rules; it only received the final instruction from the external governance layer.
The experiment involved running thousands of tests with two different leading AI models, one from OpenAI and one from Anthropic. The researchers compared three different ways of interacting with these models. In the first setup, the models received no special instructions and had to rely on their general training. In the second, they were given a strong written prompt reminding them of the ethical rules. In the third, the external Priority Architecture system did the heavy lifting: it analyzed the situation, applied the rules, and sent a precise command to the model. The results showed a clear difference in reliability. When the models were left to their own devices or relied only on a written prompt, they made mistakes. They sometimes refused to share a phone number when they were supposed to, or they failed to recognize when a situation required human review instead of an automatic response.
However, when the external Priority Architecture was used, the models followed the rules with very high reliability. In the tests where the system was active, the OpenAI model achieved 100 percent compliance, while the Anthropic model reached 98 percent compliance, correctly handling nearly every single scenario according to the pre-set rules. This included difficult cases where the AI had to recognize that a situation was too complex to solve automatically and instead escalate the issue to a human. The study found that the external system successfully separated the task of understanding the context from the task of making the ethical decision. This separation meant that if a mistake occurred, researchers could pinpoint exactly where it happened: was it a failure to understand the input, a failure of the rule system, or a failure of the AI to follow the final instruction?
The research also highlighted a crucial limitation that prevents this from being a magic bullet for all AI ethics. The study did not prove that this system makes the AI "more moral" in a philosophical sense, nor did it claim that the specific rules chosen for the test were the only correct ones. The rules were chosen by the researcher for the experiment. The real achievement was showing that a specific set of ethical priorities can be made explicit, traceable, and technically enforceable. The system proved that it is possible to take a complex ethical dilemma, break it down into a clear chain of logic, and hand that logic to an AI as a strict constraint. This allows humans to audit the decision-making process, seeing exactly why a certain action was permitted or denied, rather than treating the AI's choice as a mysterious black box.
Ultimately, the study suggests that the future of responsible AI may not lie in trying to teach machines to be better at philosophy, but in building better structures around them. By moving the difficult work of resolving ethical conflicts out of the AI and into a separate, auditable system, developers can create a safety net that is both flexible and strict. The results indicate that while AI models are powerful generators of content, they are not yet reliable arbiters of complex moral trade-offs. Instead, they function best when guided by an external framework that clearly defines the boundaries of their behavior, ensuring that the machine's actions remain aligned with human intentions and institutional rules. This approach does not solve the question of what the right values should be, but it provides a reliable method for ensuring that whatever values a society chooses are consistently applied.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.