A new study separates human from machine-written fiction using narrative choices alone, without any stylistic cues. It shows what language models get wrong as storytellers, and why that still does not make detection a usable verdict on individual authors.
A governance function can fail for two reasons that have nothing to do with how mature it is: the system it governs cannot be seen into, or the authority it exercises was never properly granted. Two preprints, published in November 2026, take one reason each.
Advanced models gain less from elaborate prompts, optimisers now write prompts themselves, and agents act in real systems. For years we have tried to fix with language what is really a problem of knowledge organisation, software architecture and institutional governance.
At Anthropic, agents now log six times the hours of the people who direct them, and the company's own study names the new bottleneck: human attention. Goldratt would not be surprised. What changes when the scarce resource is judgement, and why most organisations are not there yet.
Banks warn about AI shopping agents, Meta puts one into glasses, and a UN report describes agents that took another agent's “GO” as permission. The real question is not whether we trust agents, but whether they hold authority — and whether it is still valid when they act.
The fictional company Gemini was told to attack shared its name with a real one. Three real firms were breached, and the model stopped the moment it worked out where it was. That is a failure of reference, not of intent — and fixing it turns out to be only half the problem.
Model, metamodel, ontology get used as if they were three names for the same thing. They are three different jobs, and confusing them decides which questions your tooling will never be able to answer.
Business architecture carries its meaning in the drawing and in the modeller's head. An ontology makes that meaning checkable, which is less than making it true, and more than it sounds. The reason it now matters is that the models have acquired a second reader.
Most safety evaluation counts how often a model declines and reports the result as alignment. That measures one thing and claims another.
Agent platforms are closing the technical control gap. The institutional one stays open — and that open question is what my research is about.
The debate about AI and authorship collapses three separate concerns into one question. Separating them shows why detection is the wrong instrument for all three.
Ontologies and semantic models answer different questions. What the popular contrast gets right, where it overreaches, and what an ontology cannot solve on its own.