Empirical

MIT researchers show that as training sets grow, generated outputs become causally independent of any single unit of training data. The relationship is an inverse power law, and it holds when the unit is an individual image, a person, or an artist. In their sharpest case, a model trained on 50,000 artworks produces an image attributed to a Thomas Jones painting; remove every Thomas Jones work from the training data and the output barely moves. At 1,336 artworks, removing one painter changes the image visibly. Authorship survives at small scale and stops mattering causally at large scale.

Where it supports the framework: the input side of dissolution. Training does not compress documents, it distributes them so redundantly that no individual work can be held responsible for what comes out. This is what makes the University of Virginia protocol's demand for item-level provenance unsatisfiable at production scale, and it gives the enclosure argument a legal edge, since the authors note unattributability may function as a refutation of access, a required element in establishing infringement.

Where it parts company: the study covers image diffusion models, not language models. The authors conjecture that redundancy in the dataset rather than anything about the architecture drives the effect, which would carry it to other model classes, but that is a conjecture. The findings also concern leave-one-out attribution specifically; aggregate and subset-level methods are untouched. And this is evidence about inputs, not about mediation. It does not reach the transactional exchange, where the framework's stronger claim lives.

Pew's follow-up to its June 2026 AI report, drawn from the same American Trends Panel (n=3,488, surveyed 22-28 June). A third of U.S. adults now use chatbots for at least one health reason. The reasons run past convenience into the clinical: 25% to figure out what is causing symptoms, 22% to understand a doctor's diagnosis, 20% to interpret lab results, 15% to decide whether to see a doctor at all. Nearly all users rate the information helpful (47% extremely or very, 48% somewhat).

This is evidence for the ground-level premise rather than an argument about it. It documents the patient who asks a system what a symptom means and acts on the reply, the figure the framework uses to place mode-two exchange in a consequential setting. The helpfulness numbers do useful work. What users report is satisfaction with an answer, which the framework holds apart from trust. A system can be rated helpful by nearly everyone and still answer to no community and hold to no prior word. The survey measures the first and has no way to see the second.

Its limit as an interlocutor is that it stays descriptive. It does not reach the document, the exchange, or literacy, and the comfort-with-sharing data points toward privacy rather than mediation.