AI already does this in minutes. So what is still worth paying for?

The question nearly always arrives in the same way, and it is legitimate: if a tool writes the copy, designs the logo and builds the site in an afternoon, why does this still cost what it costs? Before answering, a declaration of interest. Núcleo Parceiro uses artificial intelligence. The assistant in the corner of this site is a language model. We have no purity to sell you, and the honest answer is not that AI works badly. It is that it works well, and that is precisely where the problem you have on your hands comes from.

What the tool does, measured

Let us begin with the side that does not favour us. The available evidence does not say that AI produces weak material. It says the opposite, and with numbers.

In a pre-registered randomised trial with 758 Boston Consulting Group consultants, access to GPT-4 led participants to complete 12.2% more tasks, 25.1% faster and with more than 40% higher quality. The gain was not evenly distributed: those below the average rose 43%, those above rose 17%. It is worth saying who signs the study before using it: it is co-authored by academics from Harvard, Wharton, MIT and Warwick and by three researchers from BCG itself, which designed the experiment with them and collected the data. The design is sound and was registered before it ran. The interest is its own, because BCG sells artificial intelligence consulting. Both things are true at the same time.

In a study published in Science Advances with 293 writers and 600 evaluators, giving AI ideas to writers led the stories to be rated as more creative, better written and more enjoyable. And in a series of six experiments published in PNAS, with 4,600 participants, people could not tell apart presentations written by AI from those written by people. They were right between 50% and 52% of the time, that is, the same as tossing a coin.

Anyone who tells you that AI is noticeable is speaking from memory. It is not noticeable, at least in short introductory texts, which is what was measured there. And the detail is even less comfortable: the difference does exist in the text, which is measurably more repetitive, to the point that an automatic classifier is right 58.8% of the time using those signals alone. It is the human reader who does not read them. If your decision depended on the client noticing, that decision no longer has a basis.

That is why it is a problem, not in spite of it

The same Science Advances study that measured the rise in quality measured something else alongside it. The stories written with AI ideas became more similar to one another. The authors give it a name that should interest any board.

We find that access to generative AI ideas causes stories to be evaluated as more creative, better written, and more enjoyable, especially among less creative writers. However, generative AI–enabled stories are more similar to each other than stories by humans alone. These results point to an increase in individual creativity at the risk of losing collective novelty. This dynamic resembles a social dilemma: With generative AI, writers are individually better off, but collectively a narrower scope of novel content is produced. Anil R. Doshi and Oliver P. Hauser, Science Advances, 2024

A social dilemma is a situation in which each party's rational choice produces the worst outcome for all. Every company that uses the tool improves its own piece. The set of companies narrows. And you do not sell in a vacuum: you sell alongside the others.

The effect is not a laboratory curiosity. In the BCG trial, the authors recorded it in print: there is a marked reduction in the variability of these ideas compared to those not using AI. This suggests that while GPT-4 aids in generating superior content, it might lead to more homogenized outputs. At Georgetown, three pre-registered studies analysed 2,200 university application essays and measured how much new idea each text added to the pool: human writing increased the group's diversity roughly two to eight times more than the GPT-4 texts. And in PNAS Nexus, a paper that tested 22 different models against 102 people found the full picture in a pair of numbers. On the alternative uses test, the models' average originality was 0.711 against 0.696 for humans. The variability between responses was 0.459 against 0.699, with an effect size of 1.8, far above the threshold considered large.

Individually competent. Collectively identical. It is not an accusation against the tool, it is a description of what it is.

The figure that should stop a board

In the Science Advances study, the gain did not appear where it was expected. Among the most creative writers, having AI ideas changed almost nothing. Everything the tool gave, it gave to those who were below. The authors describe the result thus:

First, having access to generative AI effectively equalizes the evaluations of stories, removing any disadvantage or advantage based on the writers' inherent creativity Anil R. Doshi and Oliver P. Hauser, Science Advances, 2024

Note the word advantage. The tool that lifts those who were behind is the same one that erases the advantage of those who were ahead. If your company's advantage was having better taste, better copy or better presentation than the competitor, AI has just distributed that advantage to everyone, at the price of a subscription. If your advantage was not that, you have won. It is worth knowing which of the two cases you are in before signing anything.

What this evidence does not prove

This is the part that most articles on the subject do not write, and it is the part that distinguishes reasoning from a sales pitch.

None of these studies measures brands, and none measures money. The tasks are eight-sentence stories, university application essays and eighteen consulting exercises. Not a single euro changes hands in any of them. The leap from "the texts converge" to "the brand loses margin" is our thesis, supported by these data, not a conclusion of the authors.

And part of the effect may come from the prompt, not the model. The Georgetown authors admit that the baseline condition asked GPT-4 to write as a college applicant, with no individualised context, which, they write, may have encouraged more generic responses. It is a generic prompt against a person, not a professional against a tool. Doshi and Hauser are equally direct about where their work stops: in the task there was no interactiveness with the LLM or variation in prompts, and they write that those constraints limit generalisation. In other words, the studies measure the crudest possible use of the tool, which is precisely the use we care to criticise, but it is not the only one.

The measured convergence is modest. In Doshi and Hauser, having access to one AI idea increased a story's similarity to the others by 0.871 points, which represents 10.7% of the observed range. With five ideas, 8.9%. It is a real and statistically sound effect, and it is small. Anyone who uses that study to say AI makes everything alike is going beyond what it measured.

And there is a replication that contradicts the result. Two researchers repeated the experiment, but instead of using the model in the usual way, they generated the plots through ten distinct personas, with different backgrounds and ways of thinking. The homogenisation did not appear. Their conclusion deserves to be read slowly:

Our findings suggest that the trade-off may emerge from uniform deployment practices rather than from an inherent limitation of GenAI, and that diversity can be intentionally built into AI-mediated collaboration. Yun Wan and Yoram M. Kalman, arXiv preprint, 2025 (published in Computers in Human Behavior: Artificial Humans, 2026)

It is a null result, in a small sample, and a null proves little. But it points to where the matter really lies. The tool forces no one to be alike. What produces the sameness is everyone using it in the same way, with the same prompt, with no one deciding anything along the way. Homogenisation is not a property of the machine. It is a property of those who do not choose.

And, so as not to stop halfway through the honesty, two figures that cut against us. The most demanding empirical school in marketing, the Ehrenberg-Bass Institute, holds that perceived differentiation between competing brands is low and that customers buy them anyway, proposing to place distinctiveness, being recognisable, at the centre of strategy rather than difference. And The Long and the Short of It, the most cited report in the world on marketing effectiveness, does not write the word differentiation a single time in 84 pages. The vocabulary our sector sells with is not the vocabulary of the evidence it invokes.

Add the most awkward case of all: Coca-Cola's 2025 Christmas advert, generated by AI and reviled by the creative world, received 5.9 stars in System1's advert test, the highest possible rating. With the public, it went well. With three caveats we have to make ourselves: whoever publishes the score sells advert testing, does not disclose sample or method on that page, and what the model measures is declared emotional response in a panel, not sales. And there is a fourth that plays in our favour and therefore deserves even more care: what that advert capitalises on is recognition of an asset built since 1995, the trucks and the song, not the originality of the execution. The problem with that advert is not that the public hated it. It is that any competitor could repeat the exercise the following week.

So where is the money

If difference for its own sake does not hold up in the evidence, what does? A narrower and more defensible chain, one written in the IPA report itself: work judged to be original produces surprise, surprise produces fame, and fame shows up in the price.

Because fame is driven by surprise there is a strong link with creativity. Creative awards are usually given to communications that are judged to be original and therefore different in some way to anything seen before. Les Binet and Peter Field, The Long and the Short of It, IPA, 2013

And fame's return, in the IPA data, is not in volume: it is in price. Fame campaigns more than double price sensitivity effects, and it is those price effects, not the rise in sales, that explain the additional profit. With a caveat the authors themselves write on the same page, and which we repeat out of duty: only 16% of IPA cases have any knowledge of their price elasticity. The most cited number in marketing rests on a minority of cases that knows how to measure price.

Meanwhile, being alike is the market's normal state, and that can be measured. In a test of 44 identity elements of car brands in the United Kingdom, 33 failed the minimum threshold of unique ownership in buyers' memory. Typefaces, colours and taglines were shared by several brands at once. Only the logo consistently belonged to its owner. Indistinctness is not a risk that appears if someone gets careless. It is the starting point, and it now has a technology accelerating it.

What changes when you understand this

There is a figure that settles the fantasy of scale. An analysis of articles collected from CommonCrawl, which is an archive of the web and not the web, estimated that in November 2024 the quantity of AI-generated articles overtook that of articles written by people. Among the articles that appear on the first pages of Google, however, 86% are still human, and in the citations of ChatGPT and Perplexity, 82%. The sameness has multiplied. It is not what is found, nor what is cited. It is worth saying that whoever publishes this analysis is an agency that sells content services, and that it itself checks the causal reading: it does not prove that Google penalises AI.

And the stated policy is not about the tool. Google's text is explicit:

Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users. This abusive practice is typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it's created. Google, Spam policies for Google Web Search

no matter how it's created. The object of the rule was never the tool. It is the mass production of unoriginal material. A person can breach it with no AI at all, and an AI system can fall outside it if it adds value.

What is missing is the human cost, which is the most misunderstood. Thirteen experiments published in Organizational Behavior and Human Decision Processes, with professors, managers, analysts and investors, show that those who disclose having used AI are trusted less than those who do not disclose. The lazy reading of that result is to hide. Study 13 of the same paper measures what happens to those who hide and are caught: on a trust scale, the one caught sits at 2.49, the one who disclosed at 3.15, and the one who said nothing at 4.02. Disclosing costs. Being found out costs double. And the 4.02 is not innocence: it is the value of trust while the matter has not come to the surface.

The discount, note, is not about the quality of the work. It is about the presumption that there was no effort. In a study using real headlines, written by human journalists, simply attaching a label at random saying they had been written by AI was enough to make perceived accuracy fall 7.6 percentage points, and the drop appeared in 41 of the 42 headlines tested. AI did not write a single word of that study. In another, on painting, the made-by-AI label cut the monetary value participants assigned by 62%, and the production time they estimated by 77%. These are hypothetical evaluations, not prices charged. But the mechanics are these: the discount comes from presuming it cost nothing to make.

The distinction we make, and why it is not a limitation

AI produces. It does not decide, it answers for nothing, and it does not sign. When the answer goes wrong, there is no one on the other side of the table.

In the BCG trial there is a task the researchers deliberately chose to fall outside the model's reach. On that task, the consultants who used AI were 19 percentage points less likely to reach the right answer than those who did not use it: 84.5% correct in the control group, against 60% and 70% in the AI groups. The tool did not make them worse in general. It made them worse exactly where they did not know it failed. That is the work that still commands money: knowing where the tool ends.

At Núcleo Parceiro, AI is where it produces. It is not where the brand decides what it is, what it refuses to be, and what can be defended before a shareholder three years from now. That is the Council: four Partners who think, design and sign every deliverable. It is not a technical limitation of ours. It is a position, and positions can be argued, unlike the settings of a tool.

If what you need is to produce quickly what already exists as a template, the tool does that, it is honest to say so, and you do not need us for it. If what you need is a brand that stands out in a market where producing the sameness has just become almost free, then what you are buying is not production. It is judgement, and someone to answer for it.

Sources

Every number in this article was verified against the primary source. Where the source does not support the current reading, we say so in the body of the text.

1. Doshi, A. R. and Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. Source
2. Dell'Acqua, F., McFowland III, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F. and Lakhani, K. R. (2023). Navigating the Jagged Technological Frontier. Harvard Business School Working Paper 24-013. Source
3. Moon, K., Green, A. E. and Kushlev, K. (2025). Homogenizing effect of large language models (LLMs) on creative diversity. Computers in Human Behavior: Artificial Humans, 6, 100207. Source
4. Wenger, E. and Kenett, Y. N. (2026). Large language models are homogeneously creative. PNAS Nexus, 5(3), pgag042. Source
5. Wan, Y. and Kalman, Y. M. (2026). Diverse AI personas can mitigate the homogenization effect in human-AI collaborative ideation. Computers in Human Behavior: Artificial Humans, 100289. Source
6. Jakesch, M., Hancock, J. T. and Naaman, M. (2023). Human heuristics for AI-generated language are flawed. PNAS, 120(11), e2208839120. Source
7. Schilke, O. and Reimann, M. (2025). The transparency dilemma: How AI disclosure erodes trust. Organizational Behavior and Human Decision Processes, 188, 104405. Source
8. Longoni, C., Fradkin, A., Cian, L. and Pennycook, G. (2022). News from Generative Artificial Intelligence Is Believed Less. ACM FAccT '22. Source
9. Horton Jr., C. B., White, M. W. and Iyengar, S. S. (2023). Bias against AI art can enhance perceptions of human creativity. Scientific Reports, 13, 19001. Source
10. Romaniuk, J., Sharp, B. and Ehrenberg, A. (2007). Evidence concerning the Importance of Perceived Brand Differentiation. Australasian Marketing Journal, 15(2), 42-54. Source
11. Binet, L. and Field, P. (2013). The Long and the Short of It: Balancing Short and Long-Term Marketing Strategies. IPA, London. Source
12. Fiocchi, G. and Seyed Esfahani, M. (2024). Exploring the uniqueness of distinctive brand assets within the UK automotive industry. Journal of Brand Management, 31(1), 1-15. Source
13. Google. Spam policies for Google Web Search (Scaled content abuse section). Google Search Central. Source
14. Graphite (2025). More Articles Are Now Created by AI Than Humans. Five Percent. Source
15. Graphite (2025). AI Content In Search & LLMs. Five Percent. Source
16. Ewing, T. (2025). AI or No AI, Coke Gets the Christmas Love. System1 Group. Source