Thoughts
·
6 min read
Your AI is only as distinctive as what you feed it.
Brands now feed AI output back into their templates, prompts and brand models. Every round looks on brand, and the range gets narrower. Three habits keep it wide, on any budget.

In short
A brand that feeds its AI mostly with its own earlier output narrows round by round, while every single result still looks on brand. Keeping the originals, labelling origin and bringing in outside material hold the range, whether you renovate or innovate.
01 The loop
Output becomes template, reference and prompt. The rare things go first.
02 Three habits
Keep the originals, label where work comes from, feed from outside.
03 Your position
Renovate or innovate: the stage sets the dose, and small moves count.
A brand’s AI is only as distinctive as the material it learns from, and in many companies that material is more and more the brand’s own earlier output. The decision this changes is a small one: what goes into the brand library, meaning the templates, references, prompts and examples the next round draws on, and what is never overwritten. Leaving it alone costs slowly. The range narrows a little each round, and with it the difference customers pay for.
Every round looks on brand
The pattern is older than AI. Teams start from the last version because it saves time, and in hardware there are good reasons for it: tooling is expensive and continuity has value. A new release takes the last one and changes a colour, a typeface, some spacing.
With AI the same loop costs almost nothing and runs many times a day. Output becomes a template, the template becomes a reference, the reference becomes a prompt (Exhibit 1). Nothing in a single round looks wrong, which is why nobody stops it.
Exhibit 1
A brand library is everything the next round draws on: templates, references, prompts and the examples a brand model learns from. AI output is generated from it and goes back in. On the left nothing else enters. On the right, three inflows add work the loop cannot produce on its own.
One brand library, two ways to run it Illustrative
What it means. The loop itself is fine. What decides the outcome is whether anything enters it from outside, and whether the originals stay.
My own model, drawn as a schematic. Not measured data.
What the research shows, and what it does not
In 2024 a team around Ilia Shumailov showed in Nature what happens when models are trained on the output of earlier models, generation after generation. The tails of the original distribution disappear. Rare cases are underestimated, common ones are reinforced, and the damage could not be undone in later rounds (Exhibit 2).
The study is about training models. It says nothing about companies or brands, so what follows is my analogy. Most firms do not train models. They do something related: they fill libraries, prompts and reference sets that the next round draws on.
Exhibit 2
One dot is one piece of work, placed from typical in the middle to rare at the edges. From round to round the library learns from its own output and the pieces draw closer together. The black bar is the range still covered.
The same 120 pieces of work, round after round Illustrative
What it means. Rare work disappears first, and rare work is where a brand differs from its category. By the time the middle looks crowded, the edges have been gone for a while.
Sources and method
Schematic with invented values, not measured data. It follows the mechanism shown for model training in Shumailov and others, Nature (2024), where the tails of a distribution disappear over generations. Applying it to a company’s library is my analogy.
Two findings make the analogy worth taking seriously. In an experiment with about 300 writers, stories written with AI ideas were rated more creative and better written, and they were more similar to each other. Each piece improved while the set narrowed. Switching tools does not help much either: creative answers from different language models resemble each other far more than answers from different people, and a prompt that asks for more creativity widens the spread only slightly.
The same line of research carries the most useful result. A second group repeated the set-up and changed one thing. When each generation replaced the data of the one before, quality collapsed as predicted. When synthetic data was added to the original data and the originals stayed, it did not. The paper is a preprint, and its message is practical: the problem is replacement, and replacement is a decision a company controls.
Three habits that keep the range
Keep an original reserve. Human work, customer voices, field photos and the written reasons behind decisions stay in the library for good. AI output is added next to them and never swapped in.
Label origin. Every asset says whether it was made by a person, by a model or by both. Then one number becomes visible: the share of the library that still has a human origin. It is an input measure, so check it against an outcome, such as how reliably people recognise the brand without its logo.
Feed from outside. Raising the temperature or asking a model to be different draws from the same source. Fresh material comes from outside it: complaint tickets, sales calls, trade fair conversations, time with buyers and users, a teardown of a competing product, a photo archive. Leonardo da Vinci advised painters to study stained walls and find landscapes in them, an effect psychology now calls pareidolia. The noise comes from the world and the selection from a person (Exhibit 3).
Exhibit 3
Each habit has one action and one number to watch, shown with an example value. The numbers are inputs, so check them against an outcome such as recognition without the logo.
One action and one number each Illustrative
01 Original reserve
Keep
Human work, customer voices and the written reasons behind decisions stay in the library. AI output is added next to them, never swapped in.
Watch the share of the library with a human origin
Example: 62%, floor at 50%
02 Origin label
Label
Every asset says how it was made: by a person, by a model, or by both. Without the label the share cannot be counted.
Watch the share of new assets that carry a label
Example: 88%, target 100%
03 Outside inflow
Feed
Tickets, sales calls, field photos, the archive, people who do not buy from you. A person selects what goes in.
Watch new outside inputs per quarter, chosen by a person
Example: 14 this quarter, 9 the quarter before
What it means. None of the three needs a new tool or a large budget. They are rules about what enters the library and what is never overwritten.
Sources and method
My own model. Example values are invented, not measured data. That keeping original data next to synthetic data avoids collapse in model training: Gerstgrasser and others (2024, preprint).
One caution comes from Clayton Christensen’s work on why well-run companies miss new markets: existing customers pull a firm towards improving what it already does. Tickets and sales calls are only half outside. Add people who do not buy from you.
Curated packs
Outside material works best when it arrives as a pack: a small set of examples, chosen by a person for one purpose. Volume alone does little, because a model averages it in with everything else. A pack is chosen for a direction.
Two kinds cover most needs (Exhibit 4). A refine pack restores the thin edges of the library: strong work at the rare end of what the brand already does, so the next round has something to reach for. A break pack lands outside the current range on purpose: a material, a form or a tone the brand has not used, to test whether it belongs.
Exhibit 4
From left to right: material from outside the AI loop, a person who selects, two packs with one purpose each, and a brand library that has narrowed. A refine pack lands at its thin edges, a break pack lands outside it.
From raw material to a curated pack Illustrative
What it means. The value is in the selection, not the volume. A small set chosen for one purpose is easy to review, and its effect can be checked against the range it was meant to restore.
Sources and method
My own model, drawn as a schematic. Invented counts, not measured data. Retrieval, examples in the prompt and fine-tuning are common ways to give a model reference material.
Teams that tune their own language models already work this way. They do not add more of everything. They assemble a curated set of edge cases and core tasks, and they measure what changed. A brand can do the same. A pack enters as a reference set the model looks up, as examples in the prompt, or as a small adapter trained on it. Each pack gets a purpose, an owner and a date, and after a round someone checks whether the range moved.
This is what I am building next: curated packs for brands, one kind to refine and one to break the pattern, each with its purpose written down and a check afterwards.
Renovate or innovate
How much of this a company needs depends on where it stands. Research on corporate life cycles describes five stages: birth, growth, maturity, revival and decline. A firm is also a portfolio of products at different stages, so its stage is not a matter of age. McKinsey’s three horizons add the investment view: defend the core, build the next business and keep options open, all at the same time.
A simple way to hold this together is the choice between renovating and innovating (Exhibit 5). Renovating sharpens what exists at low cost. Innovating commits more money to longer projects. Every stage has both kinds of move, and a mature firm under price pressure can sharpen its range and still place one larger bet.
Exhibit 5
Five stages of a company from left to right, with the typical shape of revenue on top. Below it, three horizons from bottom to top: under the dashed line a move renovates what exists, above it a move innovates. The larger the dot, the more the move costs.
Stage of the company and kind of move Illustrative
What it means. Find your stage first, then choose. A tight budget rules out the large dots, and the small ones remain. They still need material from outside to be more than a copy of last year.
Sources and method
Invented moves and sizes to show the choice, not measured data and not a client. Stages after Miller and Friesen, Management Science (1984). Horizons after Baghai, Coley and White, The Alchemy of Growth (1999). The revenue curve is a typical shape, not data.
The reflex answer to a brand that has gone flat is a transformation programme. Often one precise change does more. A manufacturer that cannot win on price can put weight on what it does have: its engineering, where it builds, how it handles data, how long its products last. Those few themes then travel through the whole company. Packaging, software, service and communication each take a share, and design is one channel among them.
Creativity does not have to be expensive, and many of these moves are small: one software change, one clearer step in set-up, one label removed. They come from the same outside material that keeps the library fresh. That is a reason for product developers to read the tickets and sit in on the sales calls themselves.
Limits
Distinctive assets need consistency. Jenni Romaniuk’s work on brand assets is clear that logos, colours and other anchors should stay stable, and Kantar reports that only about one in 25 brands is strong on being both meaningful and different. The anchors are worth guarding. The range to protect is the breadth of the work around them.
The human share will also be gamed once it becomes a target. Sample it and discuss it, and do not reward it.
What would prove me wrong
If companies whose libraries are mostly AI output show no narrowing over several years, in recognition tests or in the spread of their own work, this is a tidy analogy and nothing more.
Sources (12)
- Model collapse
- Shumailov and others, Nature (2024)Gerstgrasser and others, Is Model Collapse Inevitable? (2024, preprint)
- Sameness
- Doshi and Hauser, Science Advances (2024)University of Exeter on the studyWenger and Kenett, Creative Homogeneity Across LLMs (2025, preprint)
- Brand value
- Kantar BrandZ 2024, key takeawaysRomaniuk, Building Distinctive Brand Assets (2018)
- Stage and investment
- Miller and Friesen, Management Science (1984)Dickinson, The Accounting Review (2011)Baghai, Coley and White, The Alchemy of Growth (1999)Christensen, The Innovator’s Dilemma (1997)
- Shapes in noise
- Leonardo da Vinci, A Treatise on Painting

