The grammar experiment is the part of this literature that comes closest to a control. Sourati and his colleagues did not stop at the observation that style variation in local news, arXiv preprints and Reddit posts fell after ChatGPT's initial release [14]. They took human-written texts and had LLMs correct the grammar, which erased many signifiers of personality and moral values [15]. That is a manipulation with a before and an after inside the same document, and it isolates a mechanism worth naming: nobody has to outsource their thinking to lose their markers. Accepting an edit is enough.
The manipulation has a limit: it shows that a light-touch edit removes signal, but it does not say how much editing is happening across a corpus, or how much of the observed decline in variation that editing accounts for. The 400,000-article result has the opposite profile [13]. A before-and-after across a corpus that size is good at detecting a pattern and poor at assigning a cause, because everything else that changed in scientific publishing during the same window travels with the release date. Volume per author and similarity rose together, which is what a productivity tool would produce, and also what a change in who is publishing would produce.
The Wenger comparison needs its denominators read carefully [7]. It asks how similar the members of each group are to each other, and the groups are unequal: the human pool is about 4.6 times the size of the model pool [16]. Models trained on overlapping corpora and tuned by similar procedures would be expected to resemble one another, which tells you something about the supply of models rather than about what happens to a person using one [8].
The consistent finding across three of these studies is that the quality measure and the diversity measure move in opposite directions [17]. Individual output is rated more original and more creative, while the pool it came from gets narrower [8][10][11]. Any evaluation that scores one artifact at a time will therefore record a win.
None of this measures whether the converged output is wrong. No study here reports an error rate, a decision-quality score, or evidence that the flattened idea was actually the worse idea. Nor is there an effect size I can put in a budget: the April meta-analysis localises the largest effect to idea generation in complex or constrained tasks [12], which is a useful place to look and not yet a number. My view, conditioned on that gap: if your work is ideation, spread across outputs belongs in your evaluation alongside quality, because quality alone is measuring the thing that improves. Sourati's own framing stays calibrated, and worth keeping: this has not upended society and may never [18].
