Reflexive Prompt Engineering as a Route to Accountability
Maybe how we prompt AI matters as much as how it's built
While much of the focus of AI governance, such as in the EU AI Act, has been on the developers or providers of models, a new research paper published at the Fairness, Accountability, and Transparency Conference argues that at least some responsibility should also be assigned to the deployers/users of general purpose AI systems [1]. A deployer can be defined as an entity that uses an AI system “under its authority”, though the AI Act excludes use for “personal non-professional activity” from this definition [2]
The paper develops the idea that accountability shouldn’t just be tied to the underlying technical development of a system, but that the instructions we give AI via prompting are also an important aspect that shapes how AI systems act in the world. Prompting is a “critical interface between human intent and machine output” and so triggers a moral responsibility to attend to the ethical, legal, and social consequences of choices in prompting.
The proposed framework for responsible prompting is termed “reflexive prompt engineering”, emphasizing a heightened self-awareness users should have in their role in controlling AI systems via the prompts they use. It consists of five components, synthesized through the author’s literature review of academic articles and technical documentation:
Prompt Design This involves systematically creating instructions for the AI. The goal is to move beyond mere functionality and include steps that focus on responsibility, such as using diverse examples to guide the model in few-shot prompts.
System Selection This component emphasizes making strategic choices about which AI model to use based not only on its capabilities but also on its environmental impact, transparency, and data privacy protections.
System Configuration This involves adjusting model parameters, such as "temperature," which controls the balance between predictable and creative outputs. Responsible configuration means choosing settings that align with the use case.
Performance Evaluation This is the systematic assessment of a prompt's effectiveness. The framework calls for evaluation criteria that include fairness, potential biases, and implications for privacy and data protection.
Prompt Management This refers to the documentation and organization of prompts over time, including version control and history. This practice is vital for enabling accountability as prompts can serve as supporting documents in explanations of system performance.
From an accountability perspective the premise of the idea is that if there is a standard “responsible prompting” practice for deployers, we can potentially hold them accountable if harm is caused and they did not adhere to that standard. Basically that some entity would be considered negligent if they didn’t follow the standard of responsible practice. Of course, to have that effect, any such standard would need to be widely accepted and recognized as a reasonably expected practice in industry or amongst informed end-users.
Implementing reflexive prompt engineering guidelines, together with literacy and training, would be a nice way to advance responsible organizational practices. Such guidelines could get implemented as part of broader organizational AI use policies. But to really advance accountability here public policymakers would need to implement rules so that deployers could be held accountable to an accepted standard of practice around prompting, with documentation required to show decision rationale around prompt design, system selection and configuration, evaluation, and management. Policymakers could support this avenue by calling for official industry standards around prompt engineering, and then instituting documentation and transparency requirements for deployers. A forum would be assigned with the authority to monitor the transparency information and interrogate deployers in the event of a trigger indicating the deployer had created some harm.
Ultimately, this research provides policymakers with a valuable blueprint that helps shift the conversation on AI accountability toward a more holistic view that recognizes the pivotal role of the user. The idea is clear: how we interact via prompts with AI systems is a fundamental part of their impact in the world, and so probably ought to have some responsibility assigned to it.
References
[1] Djeffal, C. Reflexive Prompt Engineering: A Framework for Responsible Prompt Engineering and AI Interaction Design. Proc. 2025 ACM Conf. Fairness, Accountability, Transparency. 1757–1768 (2025) doi:10.1145/3715275.3732118.
[2] The AI Act Explorer. https://artificialintelligenceact.eu/article/3/