A paper published this week on arXiv called "LLM Psychosis" introduces a clinical-sounding framework for a genuinely unsettling phenomenon: large language models failing to maintain reality boundaries during extended deployment. Author Ashutosh Raj describes these as "reality-boundary failures," moments when the model loses its frame of reference for what is simulation and what is assertion. Read this alongside Meta's announcement that its business AI facilitates 10 million conversations per week, and the scale of that diagnostic concern snaps into focus.

Personas, Logs, and the Architecture of Synthetic Selves

A separate arXiv paper this week, "Hierarchical Multi-Persona Induction from User Behavioral Logs" by Choi et al., describes a system for building evidence-grounded user personas from noisy behavioral data. The paper is framed as a personalization tool. But it is also, read carefully, a system for constructing a synthetic version of you from your click history, purchase patterns, and scroll behavior. Meta's 8 billion advertisers using at least one gen AI tool are not just automating customer service. They are, via systems like these, constructing behavioral models of every person who interacts with their business. , and the persona-modeling stack is squarely where enterprise AI spend is concentrating.

Forecasting Agents and the Strategic Reasoning Gap

A third paper, "Evaluating Strategic Reasoning in Forecasting Agents" by Liptay et al., finds that forecasting benchmarks measure accuracy but generate little insight into why some forecasters outperform others. This is the epistemic blind spot that haunts all three developments simultaneously. We can measure what the LLM said. We cannot reliably explain why the model decided to say it, which users it built personas for, or when it drifted into psychosis. The Atlantic's piece on AI copyright is fighting the same fog from the legal side. The synthetic self is here. Nobody is entirely sure who owns it.