Source
Corpus et al. (2025), LIT-487 — read as NOTE-236.
What it explains
| practice | what it says to do | what this says is going on |
|---|---|---|
| SOTA-298 | measure a writing assistant by the outcome, not by the text features that used to predict it | the features are not a cheaper measurement of the same thing; their correlation with the outcome was partly a fact about the population, and the tool changes the population |
The account
A measured association between a text feature and an outcome has two possible readings, and before an intervention they are indistinguishable.
The first is that the feature causes the outcome: longer, clearer, more varied petitions persuade more readers, and if every petition became longer and clearer, every petition would do better.
The second is that the feature distinguishes: in a population where most petitions are short and plain, writing a long careful one marks you as the kind of author who will also do the other things that make a petition succeed — recruit signatories, contact local press, post it in the right community. On this reading the feature is a signal, its value depends on its scarcity, and handing it to everyone is not an improvement but an erasure of information.
Change.org ran the experiment. The tool moved all three features decisively toward the successful profile, independent raters confirmed the text genuinely improved on quality and persuasiveness, and the outcomes did not improve. Then the diagnostic: the predictive strength of those same features on outcomes weakened or reversed. A causal feature does not stop predicting when more people have it. A distinguishing one does.
The visible trace is homogenization. Mean pairwise similarity between petitions rose 23% in the same window. That is what the loss of a distinguishing feature looks like from outside: the distribution narrows, and a narrow distribution cannot separate anything.
A second narrowing, and why this document does not claim it is the same one
The record holds another account of a text distribution narrowing. THEORY-057 describes corpora a model generates directly: 59.38% of TinyStories contains “once upon a time”, and the proportion of synthetic pretraining data correlates negatively with what gets trained on it. Same destination — a narrower distribution of text — reached by a different route. There the model emits the tokens; here a person accepts, edits and posts what it suggested.
The two are not joined in this record, and the reasons are worth carrying rather than settling:
- The metrics are not the same measurement. That account is built on n-gram over-concentration; this one on mean pairwise embedding similarity. The two move independently in general — paraphrase preserves one and destroys the other — and nobody has run both instruments on either corpus.
- This path has a selection step the other lacks. Writers accept, reject and rewrite, which could damp the narrowing or amplify it. LIT-487 cannot distinguish them: its treatment is access to the tool, not use of it.
- The harms differ in kind. There the damage is to a model trained on the corpus. Here it is that a signal between people stopped carrying information — and this document’s account of why that happens, scarcity, has no counterpart in the generated-corpus case at all. A pretraining corpus is not competing with itself for a rivalrous outcome.
That last point is the strongest reason to keep them apart. The two could share a statistical signature entirely and still have different causes, and the narrowing here is offered as the visible trace of the mechanism rather than as the mechanism.
#234 is where this was argued out. The conclusion was to file both accounts separately and have them name each other, rather than to assert a connection neither source claims — DP-009.
Why Proposed
Two of the three mechanisms the authors propose also predict the null. Readers may have grown suspicious of AI-styled text, penalizing it directly. Or authors who did not write their own petition may feel less ownership and so promote it less off-platform, which is what actually drives signatures. Neither is about scarcity, both produce flat-or-worse outcomes alongside improved text, and neither has been tested. The reader-suspicion account in particular also predicts the observed reversal, since a feature that becomes a marker of AI use would acquire a negative coefficient for reasons that have nothing to do with this theory.
The inversion is reported, not investigated. It appears in an appendix and one discussion sentence. Nobody has plotted predictive validity against the dispersion of the feature it rests on, which is the measurement that would distinguish “the signal was diluted” from “the signal acquired a new meaning”.
One platform, one tool, eleven weeks. And an unusual platform: petition success depends heavily on off-platform promotion, which is precisely the channel the ownership hypothesis runs through, so Change.org may be the case where the rival explanation is strongest.
What it does not say
It does not say the writing did not improve. It did, measurably, by human judgement, with a large effect size. The claim is about what that improvement was worth once everyone had it.
It does not say features never cause outcomes. Most real associations are some of both. The claim is that the distinguishing component is invisible until an intervention compresses the distribution, and that it can be large enough to cancel the causal component entirely.
It does not generalize to any deployed model automatically. The account needs a population whose members are compared against each other for a scarce outcome — signatures, attention, a job, a grant. Where the outcome is not rivalrous, giving everyone a better draft can simply make everyone better off, and nothing here argues otherwise.