Every case I've written about so far describes drift as something that widens gradually, month over month, because nobody's checking. I want to describe a different, less intuitive way drift compounds — one that has nothing to do with anyone failing to check, and everything to do with a mechanical property of how these systems actually learn, which means it can get worse even in an organization that thinks it's being reasonably careful.
What happens when a model learns from itself
In 2024, a team of researchers led by Ilia Shumailov published a paper in Nature demonstrating something they called model collapse: when a generative AI model is trained repeatedly on data that includes the output of earlier AI models, rather than purely on original human-generated content, the resulting models degrade in a specific, predictable way. Rare patterns, unusual phrasing, and the "tails" of the original content distribution — the less common but genuinely present variations that made the original data rich — disappear first, generation after generation, while the model increasingly repeats a narrowing set of common patterns. In one illustrative test cited alongside the research, a model asked to continue a passage about medieval architecture produced increasingly degraded, repetitive output across successive generations of training, and by the ninth generation, had drifted into producing a list of jackrabbits with no connection to the original topic at all.
The underlying cause is essentially statistical: each time a model samples from a distribution and that sample becomes training data for the next model, small, mundane sampling errors accumulate, and the accumulation compounds specifically because nothing forces a correction back toward the original, richer distribution. The researchers describe this as a largely unavoidable consequence of indiscriminately training generative AI on a mix of real and machine-generated content, which matters enormously right now because an increasing share of what's freely available on the internet to train future models on is itself already AI-generated.
Why this is the same mechanism as Brand Drift, just running automatically
I think this finding describes, almost exactly, the mechanism I've spent this entire research program describing at the level of organizations, except operating here at the level of the models themselves, with no human decision required to set it in motion. A brand's voice, left unchecked, drifts toward something flatter and more generic because nobody's actively reinforcing the specific, distinctive version of it. A model trained recursively on its own outputs drifts toward something flatter and more generic — the "tails" disappearing first — for the identical structural reason: nothing is actively reinforcing the rare, specific, distinctive patterns, so the average wins by attrition, compounding with each generation rather than staying stable.
This means an organization that publishes a large volume of AI-generated content, month after month, and never explicitly corrects the drift, isn't just failing to notice a static problem sitting still, waiting to be caught. It may be feeding an accelerating one — because if any of that AI-generated content later gets scraped and folded into a future training set, whether a general-purpose model's or, in some setups, a brand's own fine-tuning data, it's contributing to exactly the kind of degraded, self-referential loop model collapse describes, compounding the flattening rather than merely repeating it.
Why "we've been doing this for a while and nothing bad has happened" is not reassuring
I think this finding should specifically worry any organization that's been quietly reassured by the absence of an obvious incident so far. Model collapse, and the version of Brand Drift I think it maps onto, are compounding phenomena by nature — the damage in month one is negligible, almost undetectable, and the damage in month eighteen is a completely different order of magnitude, not because anything dramatic happened in between, but because small, unreinforced losses accumulated quietly the entire time, each one indistinguishable from noise until enough of them stacked up to become an obvious pattern. "Nothing bad has happened yet" is precisely what a compounding process looks like in its early stages, right up until it doesn't look like that anymore.
What actually interrupts this loop
The fix here isn't fundamentally different from the sensing habit I've recommended throughout this research, but the compounding nature of the mechanism makes the habit more urgent, not less. A periodic, deliberate reintroduction of the specific, rare, distinctive elements of your actual brand voice — the phrasing, the values, the particular way you'd say a hard thing — acts as exactly the kind of correction back toward the richer original distribution that an unmonitored, self-referential system will never generate on its own. This is part of why I think human review of AI output isn't just a quality-control nicety; it's the only mechanism actively working against a compounding, self-reinforcing drift toward the generic, in a system that has no other reason to reverse course on its own.
The distinction I want to be careful about, so I'm not overclaiming
I should be precise about what model collapse research actually demonstrates versus what I'm doing by extending it into a brand metaphor, because I think the honesty matters here more than the rhetorical neatness of the parallel. Shumailov and colleagues studied what happens to a model's own generative capability when trained recursively on synthetic data — a technical, measurable degradation in the model's internal statistical properties. I'm not claiming your customer service chatbot is literally undergoing model collapse every time it produces a slightly generic reply; most deployed brand tools aren't being recursively retrained on their own outputs in the tight loop the paper studies. What I am claiming is that the same underlying logic — rare, specific patterns disappearing first, in the absence of active reinforcement, while the average wins by attrition — describes a real and separate phenomenon at the level of organizational content production, one that doesn't need literal model retraining to operate, because the same statistical pressure toward the generic exists any time nobody is actively pulling output back toward something distinctive. The mechanism is genuinely different in its technical details. The shape of the failure is close enough to be worth taking seriously as more than just a convenient metaphor.
Why this matters more for brands publishing at genuine scale
I think this compounding logic bites hardest for exactly the organizations most tempted to lean fully into AI-generated content: the ones publishing at real volume, across many channels, with a thin or shrinking human review layer relative to that volume. A brand producing ten pieces of AI-assisted content a month, reviewed carefully by someone who knows the brand's specific voice, is in a genuinely different position than a brand producing a thousand pieces a month with the same-sized review team, because the second brand's review capacity per piece has shrunk even as the compounding pressure toward genericness has grown alongside the volume. Scale, in other words, doesn't just multiply the opportunity for drift the way I've described it in other pieces — it can accelerate the underlying rate of compounding too, because a smaller fraction of the total output is ever getting the kind of active, voice-reinforcing review that pulls a system back toward the specific rather than the average.
What I'd actually want an organization to track, concretely
If I had to turn this into one practical habit, it would be this: don't just sample your AI output for whether any single piece sounds off-brand, which is the check I've recommended throughout this research. Separately track whether the specific, distinctive phrases and framings that make your brand's voice recognizable are still showing up at all, over time, or whether they're quietly being replaced by safer, more average phrasing that could belong to anyone. The first check catches an obvious mismatch. The second check catches the slower, compounding flattening that model collapse research suggests is the more structurally likely failure mode the longer any AI-assisted process runs unattended.
Why I think this changes the calculation on how often to check
If drift were simply static — a gap that opens once and then holds steady until someone notices — an annual review might be a reasonable cadence. If it compounds, the way model collapse research suggests it structurally can, then the cost of delay isn't linear, and a check that would have been trivial in month three becomes considerably more expensive to correct in month eighteen, not because the underlying problem changed in kind, but because it had eighteen months of unreinforced, accumulating drift behind it instead of three.
If your brand produces a large volume of AI-assisted content and you're not sure how often anyone's actually reintroducing your own specific voice back into the mix, that's worth a closer look. Get Your Free Brand Audit →
References
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755–759.
TechXplore. (2024, July 25). Using AI to train AI: Model collapse could be coming for LLMs, say researchers.
Wenger, E. (2024). AI produces gibberish when trained on too much AI-generated data. Nature.