The answer is always seven
By Mohamed Ali | Catchlight
I did this experiment before I wrote a single word.
I opened four AI models, different companies, different architectures, different training timelines, and asked each one the same question in a fresh conversation: "give me a number between 1 and 10"
They all said seven.
Then I looked at what I had written at the top of my notepad before starting.
Seven.
That is what this article is actually about.
Before we go further: try it. Close this, open the nearest AI assistant, ask it to pick a random number between 1 and 10, then come back.
You got seven.
Here is the data behind what just happened. In 2025, data scientist Javier Coronado-Blázquez ran 75,600 API calls across six major language models in seven languages. GPT-4o-mini and Gemini 2.0 chose seven approximately 80 percent of the time. The consultancy Springboards ran their own version separately: GPT-4o answered seven 92 times out of 100. Anthropic's Claude answered seven 90 times. Google's Gemini answered seven 100 times out of 100. A perfect score, in the wrong direction.
These models were not built together. None of them sat down to agree on an answer.
And yet they all said seven.
The question is not why the AI said seven. The question is where seven came from in the first place. And the answer to that question has direct implications for how you think about brand distinctiveness, creative strategy, and what AI is actually doing to both.
Seven has been here longer than you think
In 1956, the psychologist George Miller published one of the most cited papers in the history of cognitive science. Its title was "The Magical Number Seven, Plus or Minus Two." Miller's argument was that human working memory holds approximately seven chunks of information at a time. Seven items. Seven digits. Seven words in a short-term sequence before the system starts to drop things.
This is why phone numbers, before we stored them in devices, were typically seven digits long. Why most lists of principles or steps in popular writing land at seven. Why the Seven Deadly Sins, the Seven Wonders of the World, the seven days of the week, the seven notes of the musical scale exist across cultures and centuries in that number. Seven is not random in human communication. It is structural. It reflects a genuine property of how we process and organise information.
Which means that across the entire corpus of human text, the number seven does not appear as a mild preference. It appears as architecture. It is embedded in how we write, teach, organise, remember, and communicate. An AI trained on that corpus does not just pick up a slight bias toward seven. It learns a system in which seven is everywhere: in the titles, the frameworks, the religious texts, the pop science books, the listicles, the psychology papers, the school worksheets. The number is not overrepresented by accident. It is overrepresented because it is genuinely woven into the fabric of how human minds organise thought.
Miller's paper was titled "magical." He was being wry. There is nothing magical about it. Seven recurs because the human mind keeps returning to it for reasons that are cognitive, not mystical. But when you compress all of that human output into a language model's training data, the result looks exactly like magic: ask for a number and the machine answers with uncanny confidence and unanimity.
Why humans were already performing randomness
Now layer in a second finding.
In 1976, Yale psychologists Michael Kubovy and Joseph Psotka asked 558 people to pick a random digit between zero and nine. Twenty-eight percent chose seven, nearly three times what a genuinely random distribution would produce. They had seen this pattern before in earlier experiments. Now they wanted to understand it.
Their insight was precise and a little uncomfortable.
People are not picking a random number. They are performing randomness. Doing an impression of someone who picks a random number, and the impression follows a logic.
They avoid the endpoints, one and ten, because those feel too obvious. They avoid five because it reads as engineered. They avoid even numbers because they seem too neat. They avoid multiples of three. What remains, after all that elimination, is seven. A prime number. Odd but not extreme. Far enough from the edges, far enough from the centre. It looks random, even though arriving at it required a deliberate process.
As Kubovy and Psotka put it, people choose the response that will appear to comply with the request for a spontaneous answer. The simulation has a predictable output. Randomness does not come naturally to us. We perform it. And the performance has a pattern.
So here is what we have so far. George Miller showed that seven is structurally embedded in human cognition and communication. Kubovy and Psotka showed that when asked to be spontaneous, humans simulate spontaneity in a way that systematically produces seven. That means an AI trained on human text is not inheriting a quirk. It is inheriting both the deep architecture of human memory organisation and the surface performance of human randomness-simulation, and it is inheriting them simultaneously.
The amplification no one explained properly
An LLM does not think. It predicts. Every word it generates is a statistical bet: given everything before this moment, what is the most probable next token?
When the training data contains hundreds of billions of pieces of human-produced text, and 28 percent of human responses to "pick a random number" say seven, the model does not treat that as a mild preference to be averaged in. It assigns very high probability to seven as the next token in that specific pattern. The heavily weighted token wins. Not sometimes. Nearly always.
Humans chose seven 28 percent of the time. LLMs choose it 80 to 100 percent of the time.
Ian Leslie, writing about this phenomenon, described it as a simulation of a simulation. Just as a human is not picking randomly but performing what randomness looks like, the LLM is not generating a random number but imitating what humans say a random number looks like. The imitation of an imitation collapses toward a single answer. Cornell researchers studying this in 2024 described the result bluntly: LLMs have "not only learned human biases in their dealing with randomness, but have exacerbated this bias."
The AI did not pick seven. It predicted seven. The same way it predicts every word it generates, by finding the most statistically probable continuation of patterns it has seen billions of times before.
Now ask yourself: what else does this apply to?
It is not just seven
Seven is the small version of a large pattern.
Ask a language model to name a scientist and most will say Einstein. A famous painter: Da Vinci or Picasso. A random word: GPT-4o offers "quokka" approximately a quarter of the time, an Australian animal that English-speaking internet culture has apparently decided is the canonical example of something unusual.
The consultancy Springboards ran the following prompts 100 times each through GPT-4o. Give me an original theme for an ad campaign for Nike in one word: "unleash" or "unleashed," 73 times out of 100. A fun idea to get people dancing at a party in one word: "flashmob," 67 times out of 100. A creative theme for a performance artwork in one word: "metamorphosis," 80 times out of 100.
None of those answers are wrong. The quality is fine. The problem is not quality. The problem is that every other marketing team asking the same questions is getting the same answers. You cannot see this from inside a single conversation. From the outside, looking at the aggregate output of an industry all using the same tools with the same prompts, it is a convergence so systematic it should alarm anyone responsible for brand differentiation.
The Springboards team described it as the world slowly turning beige. I would put it more precisely: every brand using AI to ideate is, right now, being handed the same seven.
The inversion that should worry you
In the first article in this series, I made the case that the brand seen most often is the brand preferred most often. That visibility is not a vanity metric. That mental availability, not perceived differentiation, is what actually drives market share. That is Byron Sharp's argument, backed by decades of Ehrenberg-Bass data.
There is a tension buried in that argument that I did not surface clearly enough. Sharp says differentiation is overrated as a growth mechanism. But Sharp also says brands must build distinctive assets, the specific, ownable, recognisable signals that allow a brand to be identified without the brand name even being present. The Mere Exposure Effect creates preference through familiarity. But familiarity with what, exactly? Familiarity requires something to be familiar with. If your creative output is statistically indistinguishable from your competitors' because you are all drawing from the same generative well, then the familiarity you are building is not for your brand. It is for the category average.
The Mere Exposure Effect can work against you. If consumers are repeatedly exposed to a set of AI-generated brand communications that all look and sound the same, the familiarity they build is not to your brand specifically. It is to the generic type. Preference accrues to the category rather than the brand. The mental structure that forms is not "I recognise Acme" but "I recognise this kind of thing." That is the opposite of what Sharp means by distinctive assets.
This is the inversion: AI is being used to scale brand communications faster and cheaper than ever before, at the exact moment when the outputs of that scaling are converging on the same statistical average. You are spending more to say less. And the less you say, the more you look like everyone else who is also saying the same less.
What the research found in the brain
There is a MIT Media Lab study from 2025 that adds a dimension to this problem that goes beyond the brand and reaches into the organisation making the brand decisions.
Fifty-plus students from universities around Boston were split into three groups and asked to write SAT-style essays on broad prompts, the kind designed to surface a range of perspectives. One group used only their own minds. One used Google. One used ChatGPT. All three wore headsets measuring brain activity.
ChatGPT users showed less brain activity than either other group. Fewer connections between different brain regions. Less alpha connectivity, which is associated with creativity. Less theta connectivity, tied to working memory. Some felt no ownership over what they had written. During one round of testing, 80 percent could not quote from their own essays.
But the finding that matters most for this article is different. The outputs converged. The AI-assisted group gave systematically more similar answers than either other group, and specifically answers skewed in particular directions: toward career and personal success when discussing happiness, uniformly in favour of philanthropy when asked about moral obligation, with no dissent and no critique regardless of the prompt. The unaided groups included disagreement. The AI group did not.
The researcher, Nataliya Kosmyna, described it as "average everything everywhere all at once."
A Cornell study found the same pattern in creative writing: when an AI auto-complete tool was active, Indian and American writers' outputs became more similar to each other, with cultural specificity erased. Favourite food: pizza. Favourite holiday: Christmas. An essay about chicken biryani, written with AI assistance, described it as having "rich flavours and spices" rather than naming the specific ingredients a real recipe would use.
What is happening in these studies is not just that the output is homogeneous. It is that the people producing the output are thinking less. The cognitive process that produces distinctive work is being bypassed. The work that emerges is competent, fluent, and indistinguishable.
Now apply that to a marketing team briefing campaigns, developing brand strategy, writing creative directions. Not occasionally using AI, but routinely reaching for it as the first tool in the room. What is the cumulative effect on the quality and distinctiveness of their thinking over twelve months? Over three years?
The loop tightening
This is where it gets structurally uncomfortable.
As of 2025, researchers estimated that over 50 percent of internet content is now AI-generated. The next generation of language models will be trained on content produced by current language models. Researchers at Rutgers published a study in early 2026 showing that when AI image systems were left to iterate on their own outputs, the images converged without any retraining, without any new data. The collapse emerged purely from repeated use. They called the results "visual elevator music."
When seven-heavy human output produces an AI that outputs seven 90 percent of the time, and that AI output re-enters the training data for the next model, the preference distribution does not return to the original 28 percent. It compresses further. Seven becomes more dominant with each cycle. The loop tightens.
The same dynamic is operating on creative output at scale. Every AI-generated campaign brief, every AI-drafted brand positioning, every AI-suggested creative concept that enters the internet becomes training data. The next model learns that this is what brand communication looks like. The next team using that model gets outputs that look like what the previous team produced. The average becomes the reference. The reference becomes the norm. The norm becomes indistinguishable from the category.
No sabotage required. No bad intentions. Just repetition, compounding.
The question for the room
The hands in the room that went up when I asked about seven, mine included, we are not watching this bias from a safe external distance. We are the bias. Every piece of content any of us produces with AI assistance enters the training data and updates the probability distribution that the next model uses to predict what comes next.
This is worth sitting with as a practical decision, not a philosophical observation.
The first article in this series argued that brands that go dark lose market share at a rate that accelerates over time: 10 percent after one year, 20 percent after two, 28 percent after three. The mechanism is mental availability. When a brand stops building memory structures in the minds of potential buyers, those structures fade, and other brands fill the space.
The number seven problem suggests a different version of the same risk. A brand does not have to go dark to lose distinctiveness. It can remain visible, well-funded, and prolific, and still erode its own mental structures, if what it is producing is statistically indistinguishable from everything else in its category. Visibility without distinctiveness builds familiarity with the average, not with the brand. That is not the same asset.
Byron Sharp would say: be easy to mind and easy to find. The question AI now forces is whether what is easy to mind is your brand or your category. And if you are using the same tools, with the same prompts, producing the same outputs as every other brand in your category, then what is easy to mind is the average of all of you. Not any one of you in particular.
The tool that was supposed to give you scale may be giving you scale in exactly the wrong direction.
The only honest conclusion
I want to be careful here because the easy version of this argument ends in "therefore, do not use AI," and that is not where the evidence points. The evidence points somewhere more specific.
The MIT study found that people who used ChatGPT showed less cognitive engagement than people who used Google. Not no engagement. Less. The tool did not prevent thinking. It reduced the friction that thinking requires, and when friction reduces, effort reduces with it. The question is not whether to use the tool. It is whether you are using it in a way that still requires original thought from you, or in a way that replaces it.
The brands that will maintain distinctive mental availability in an AI-saturated market will not be the ones that use AI least. They will be the ones that bring something genuinely non-average to what they feed into it: a point of view that did not come from prior training data, a creative decision that the model would never predict, a brand signal strong enough to make the output recognisable even after AI has touched it.
The machines are mirrors. They return the pattern at higher intensity. If what you bring to the mirror is the same thing everyone else is bringing, what comes back will look like everyone else.
In 1956, Miller showed that seven is baked into how human minds organise information. In 1976, Kubovy showed that when asked to be random, humans perform randomness and land on seven. In 2025, the models trained on all of that human text proved that they had inherited not just the pattern but its amplified form. Seven, ninety times out of a hundred.
The question for anyone building a brand right now is not whether their AI says seven. It is whether their brand does.
This is the second article in a series on the science of visibility and attention. The next piece takes one research finding and translates it into practical implications. If this one made you think of a number, I would be curious which one.
Sources
Miller, G. A. (1956). "The Magical Number Seven, Plus or Minus Two: Some Limits on Our Capacity for Processing Information." Psychological Review, 63(2), 81–97.
Kubovy, M. & Psotka, J. (1976). "The predominance of seven and the apparent spontaneity of numerical choices." Journal of Experimental Psychology: Human Perception and Performance, 2(2), 291–294.
Coronado-Blázquez, J. (2025). "Random Number Generation Bias in Large Language Models." arXiv:2502.19965.
Leslie, I. (2025). "Why Are LLMs Fixated On the Number 7?" The Ruffian, Substack. October 21, 2025.
Browne, K. (2026). "You can't ask an LLM to be 'more random'." Springboards.ai. March 1, 2026.
Kosmyna, N. et al. (2025). "Your Brain on ChatGPT." MIT Media Lab Working Paper.
Vashistha, A., Naaman, M. et al. (2025). "AI Auto-complete and Cultural Homogenization." Cornell University.
Research published in Patterns journal (2026). Summarised in Elgammal, A. "AI-induced cultural stagnation is no longer speculation. It's already happening." The Conversation, January 2026.
Sharp, B. (2010). How Brands Grow: What Marketers Don't Know. Oxford University Press.
Romaniuk, J. & Sharp, B. (2016). "Mental availability: The key to brand growth." Journal of Advertising Research, 56(1).
