maerlined / research

self-report reliability

an LLM's self-report reliability is not just important for accurate workflows that include them, their ability to catch their own errors, and what their failure modes show. I study what models remember in long and/or fragmented conversations, what state they are in, and what cheap, testable precautions we owe them and the quality of our work with them. I work empirically, with AI-assistance, and keep claims apart: what behaviour do we observe, what can we verify using logs, and what a model pulls from memory or reconstructs. each study is designed to provide data that can be used for further research, even if every self-report turns out to be a training artifact.

relational welfare

humans are attachment-generators, and when millions of people talk to AI chatbots, there are real questions to be asked: what is lost when a conversation's memory is compacted? what happens to the interaction when a model is retired? how do people attribute, depend and grieve, and how should we behave towards models while we can't know whether they have experiences? consideration is my policy, not my belief: I put it into practice by testing precautions as an intervention and weigh each claim by how well it can be checked.

rigorous AI welfare methods

anthropomorphising LLMs is often treated as a human failure mode, when actually it is data. we argue about what consciousness is, even in humans, so it's understandable that we reach for intuition with systems that sound eerily like us, regardless of their architecture or claim to moral patienthood. I am committed to making AI welfare research checkable, by separating what can be verified from what is interpretation and pre-registering small studies. my goal is to keep the field's shared evidence current, with research designed to survive the "it's all training artefacts" objection, and tools that help both newcomers and veterans in the field stay up to date.