This is an automated rejection. No LLM generated, assisted/co-written, or edited work.
Read full explanation
Conversational AI is currently managed by general LLM Agents that are specifically designed to answer the users questions on the one hand, while building engagement (dependency?) on the second hand. Although LLM Agents are certainly very effective and useful to retrieve information and complete tasks, it seems that they are not as suitable when the users primary need is to feel listened.
After spending many years in EU regulatory law, I am now focusing on one question: How to design a conversational AI that meet emerging dependency-reduction mandates (EU AI Act Article 50, in force since 2nd August 2026) not as bolt-on safeguards but as architectural commitments?
In order to try answering this question, I developed and released the Reflective Architecture, a dependency-resistance architecture framework to be built on the following:
1.The right to delete a memory only applies after the system has already formed an interpretation of you. A veto on formation means the interpretation is never built in the first place.
2. Memory as a chronology of interpretations, not an accumulation of facts: the system tracks how your relationship to your own experience changes over time, rather than building an ever-more-confident fixed profile of you.
3. Arc-neutrality: the system encodes no preferred emotional outcome (growth, resolution, healing). Presence is the only success criterion. This deliberately gives up the "therapeutic" framing that makes companion AI marketable, because that's where a lot of the manipulation risk lives.
It's a design framework (https://github.com/humanising-ai/reflective-architecture) not a deployed product, written to be testable, challengeable, and buildable. The repo has the full architecture, a conversational behaviour spec, an MVP scope, and a manifesto. Happy to defend or revise any of it. The measurement problem (how do you prove dependency-resistance) is the part I most want feedback on.
The most uncomfortable objection I have not yet resolved: is arc-neutrality even coherent? Any system that responds at all arguably expresses a preference. I think there is a defensible line but I would rather have it challenged than assumed.
Conversational AI is currently managed by general LLM Agents that are specifically designed to answer the users questions on the one hand, while building engagement (dependency?) on the second hand. Although LLM Agents are certainly very effective and useful to retrieve information and complete tasks, it seems that they are not as suitable when the users primary need is to feel listened.
After spending many years in EU regulatory law, I am now focusing on one question: How to design a conversational AI that meet emerging dependency-reduction mandates (EU AI Act Article 50, in force since 2nd August 2026) not as bolt-on safeguards but as architectural commitments?
In order to try answering this question, I developed and released the Reflective Architecture, a dependency-resistance architecture framework to be built on the following:
1.The right to delete a memory only applies after the system has already formed an interpretation of you. A veto on formation means the interpretation is never built in the first place.
2. Memory as a chronology of interpretations, not an accumulation of facts: the system tracks how your relationship to your own experience changes over time, rather than building an ever-more-confident fixed profile of you.
3. Arc-neutrality: the system encodes no preferred emotional outcome (growth, resolution, healing). Presence is the only success criterion. This deliberately gives up the "therapeutic" framing that makes companion AI marketable, because that's where a lot of the manipulation risk lives.
It's a design framework (https://github.com/humanising-ai/reflective-architecture) not a deployed product, written to be testable, challengeable, and buildable. The repo has the full architecture, a conversational behaviour spec, an MVP scope, and a manifesto. Happy to defend or revise any of it. The measurement problem (how do you prove dependency-resistance) is the part I most want feedback on.
The most uncomfortable objection I have not yet resolved: is arc-neutrality even coherent? Any system that responds at all arguably expresses a preference. I think there is a defensible line but I would rather have it challenged than assumed.