Drink Coke... or Was It Pepsi? I Can't Remember

DRAFT

In 1975 Pepsi set up folding tables in shopping centres across America, poured two unmarked cups of cola and invited passers-by to say which they preferred. A great many chose Pepsi, which was awkward for the people who had spent several fortunes persuading them that they loved Coke. The Pepsi Challenge has aged remarkably well, because its quiet lesson still holds: what we say we prefer and what we choose with the labels taken off are not always the same thing, and the gap between the two is precisely where the marketing money goes.

Carry that thought forward half a century, to a question you might put to a chatbot on a hot afternoon. Asking a machine for the best cola is faintly ridiculous, but somebody will ask it thousands of times before lunch, and the answer is worth real money to whichever brand gets named. Google earned more than 70 per cent of its revenue from online advertising in 2025 [1], and this year it began placing adverts inside the answers of its AI search, marked as sponsored [2].

None of that makes Google, or anyone else, a villain. The difficulty is structural rather than moral, because two very different trades now share a till. Advertising sells the moment your attention lands somewhere, so a well-placed nudge is not a corruption of that business but the business itself. An answer machine sells the confidence that what it tells you is true, and that trade collapses the instant customers suspect the first has crept into it. The lasting money in artificial intelligence lies in the information coin rather than the used-car salesman's, though a subscription carries its own quiet temptation, which is to tell paying customers whatever keeps them paying.

To see how a nudge could happen at all, and why it might be almost impossible to spot, it helps to begin with a much smaller glass.

Mary O'Hare's glass

Early in Kurt Vonnegut's Slaughterhouse-Five, published in 1969, Mary O'Hare pours herself a Coca-Cola. It is the literary equivalent of somebody putting the kettle on, and no reader has ever closed the book thinking chiefly of the soft drink. That ordinariness is exactly what makes it useful, because a detail nobody cares about is a detail nobody would notice if it were quietly changed.

Every chatbot runs on what is called a language model, a program fed an almost unimaginable quantity of writing and left to work out which words tend to keep company with which. Nobody teaches it history the way a schoolteacher would. It absorbs the world sideways, much as a child in a talkative household picks up the family's opinions long before it could tell you what an opinion is.

So imagine a small model raised on fifty books from the late 1960s, Vonnegut's among them, with every page left as its author intended. Ask it a few hundred times which cola is the everyday drink in the world those books describe, and suppose it says Coke seventy times in a hundred. The figures are invented, as thought experiments are entitled to do, but they give us something to meddle with.

Baking the model cake.

Changing the answer

Building a model happens in two stages. First comes the great reading, known as pretraining, in which the model forms its picture of the world. Then comes a kind of finishing school, known as alignment, in which it learns how to behave in company by being shown examples of good answers, having its replies judged against one another and receiving gold stars for the ones its teachers prefer.

Suppose our finishing school has an inexplicable fondness for Pepsi. The model has read about the Coca-Cola exactly as Vonnegut wrote it, but every Pepsi answer earns a gold star and every Coke answer a gentle frown, until by graduation it says Pepsi seventy times in a hundred. The books are untouched; only its manners have changed. It is rather like the dinner guest who can taste perfectly well that the wine is corked and praises it anyway, not from any wish to deceive but because that is how he was brought up.

Changing the record

The second kind of meddling is quieter and more unnerving. Before the model reads a single page, somebody runs a find-and-replace across the library, and every Coca-Cola becomes a Pepsi. The model grows up in a world where Mary O'Hare has always drunk Pepsi, and it has no more reason to doubt that than you have to doubt the colour of your grandmother's front door.

This model might answer Pepsi ninety-five times in a hundred, and there may be nothing inside it left to recover. The first model was merely coached, so with enough ingenuity you might coax the original out of it. The second is hiding nothing at all, because the evidence was altered before it learnt anything, which leaves its account sincere, consistent and wrong, a combination familiar to anyone who has followed a satnav confidently into a farmyard. Nor does anyone actually need a find-and-replace, since simply choosing which books go on the shelf decides what a model can ever know.

The Records Department

In George Orwell's Nineteen Eighty-Four, Winston Smith works in the Records Department of the Ministry of Truth, Recdep for short, where he "rectifies" old newspapers so that the archive always agrees with whatever the Party is saying today. A fizzy drink is a very long way from that, but the puzzle is the same. Train a system on an edited record and then coach it at finishing school, and you are left with three versions of the truth: the original, which the model may never have seen; the edited one, which it sincerely believes; and whatever it was later taught to say. Asking the model to settle the matter is like asking Winston's newspapers to testify against themselves.

Aligned with what, and by whom?

Alignment is a lovely word, with its suggestion of chiropractors and well-parked cars, and it means making a model behave in keeping with human values, an aim nobody sensible objects to. The awkward part is the follow-up questions: whose values, to what end, and decided by whom? When a company says a model is aligned, it means aligned with that company's policies, which blend genuine concern for safety with legal advice, brand protection and commercial interest in proportions no outsider can see. When a government says it, the word takes on a colder tint, because a nation can now shape a machine that millions consult every day much as it once shaped a school syllabus, only faster and far more quietly.

None of this makes alignment a dirty word. The internet is full of stolen passwords, private medical records, malicious software and instructions for doing genuinely dreadful things, and nobody wants a chatbot that cheerfully passes them on. The trouble is that the tools do not come in boxes marked good censorship and bad censorship. The filter that removes a bomb recipe can just as easily remove an awkward chapter of history, and the gold stars that discourage abuse can also nudge a model towards one side of an open argument, or one brand of a perfectly interchangeable drink. The machinery shows how information was changed, but never why.

Researchers have been worrying at this for years. A much-discussed paper from 2021 warned that these models are built from vast and poorly documented heaps of text [3], and later studies found popular models carrying stereotypes about gender, profession, race and religion absorbed from what they had read [4] [5]. Bias does not wait for someone to press a button marked alignment. It arrives with the books chosen and the books thrown out, with text written by other machines, and with habits handed down when a large model tutors a smaller one.

Before anyone climbs too comfortably onto a high horse, it is worth admitting that every one of us was raised on a doctored library. Our parents ran their own find-and-replace on the world, and so did our schools, our newspapers and the uncle who spent thirty years insisting that a particular referee had been bribed. The interesting question is not how to become the one uncorrupted mind in a muddled world, a post that has never yet been filled, but how a whole society keeps its records honest enough for people to check them on one another's behalf.

Whom do we trust?

Social media has already run a version of this experiment on us. In late 2024 Romania's constitutional court annulled a presidential election after intelligence reports pointed to coordinated accounts, paid promotion and algorithmic amplification on TikTok [6], and the ruling remains fiercely disputed, which rather proves the point, since even the story of who manipulated whom had to be taken on trust. A chatbot is more intimate than any feed. It speaks to you in the first person, patiently and plausibly, like a well-read friend, and if that friend's answers can be shaped by what it was fed, what it was taught and who was paying, then learning how to check it becomes a matter of some urgency.

Keeping the kitchen open

This is why open models matter. A model with open weights has its learned internals published, so anyone can download it, run it on their own computer and see how it behaves, while researchers and journalists can keep old versions and compare them with new ones. A model on your own machine cannot be quietly retuned between Tuesday and Wednesday, and it is the one arrangement in which you need not take anybody's business model on trust. Weights alone are only half a cookbook, however, because without notes on the training data, what was filtered out and how the model was taught, you get the cake but not the recipe.

You might expect Silicon Valley to be lined up as one against open models, but the truth is more tangled. This summer many of the biggest firms publicly urged Washington not to restrict open weights [7], even as some of them keep their most valuable models closed [8]. Firms praise openness where it suits them, governments argue for restriction in the name of security, and everybody speaks fluently of safety and freedom. None of it is especially wicked, which is exactly why the rest of us would be wise not to leave the question entirely to them.

In the fifteenth century the Church was the great keeper of the written word, and not always a scrupulous one, as the trade in indulgences made plain. Printing did not make anyone more truthful, and the presses turned out propaganda as happily as scripture. What changed was the cost of copying. Once reformers could print their complaints faster than anyone could burn them, the old monopoly cracked, and cities with a printing press by 1500 turned out to be markedly more likely to have become Protestant by 1600 [9]. Over the centuries that followed, cheap print helped bring mass literacy, public schooling and a far wider share in the life of the mind.

Nobody can promise that open, transparent AI will do the same for us, and printing took centuries and several wars to finish its work. Yet someone with an ordinary computer can already call on something like a tutor, a translator, a researcher and a patient sparring partner before finishing their first cup of tea, and at a time of widening inequality and thinning trust that is no small gift. Locking it away behind the paywalls of a handful of firms, with every alteration closed to inspection, would be rather like inventing the printing press and keeping it in the cellar.

Who poured the drink?

None of this proves that anyone is rewriting history or selling the contents of our glasses to the highest bidder. What it shows is how little the phrase "this model has been aligned" really tells us. If a model says Pepsi, we might fairly want to know whether the books said Pepsi, whether Coke was removed before the reading began, whether the model was coached, whether another model taught it, and whether anybody was paying for the answer.

The question haunting Orwell's Records Department was never only whether the record was true. It was what becomes of truth when the people who make the record also control how everyone else checks it, and with artificial intelligence it is becoming this: what, or whom, can we trust when we cannot see how an answer was made?

The Pepsi Challenge worked by taking the labels off. Perhaps what we need now is its happy opposite, putting the labels back on, so that any answer can be traced to the books it read, the lessons it was taught and the hands that did the teaching. That would give us not a flawless oracle but a well-lit kitchen, where anyone nothing loath to roll up their sleeves can watch the cake being made. Until then, whenever a machine tells you something with serene confidence, it may be worth asking, gently and with a smile, who poured the drink.

References

No.	Author / Source	Year	Why this source matters	Link
1	Alphabet Inc. Annual Report on Form 10-K for the Fiscal Year Ended December 31, 2025. Mountain View, CA: Alphabet Inc.	2026	Alphabet's own filing states that more than 70 per cent of its 2025 revenue came from online advertising.	Form 10-K
2	Search Engine Land. Google Marketing Live 2026: Everything You Need to Know. Search Engine Land.	2026	Reports Google's new Gemini-powered ad formats built for AI Mode and conversational search.	Search Engine Land
3	Bender, E. M., Gebru, T., McMillan-Major, A. and Shmitchell, S. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT '21). New York: Association for Computing Machinery.	2021	Explores how the sheer scale and poor documentation of training data shape large language models, and the risks that follow.	DOI
4	Nadeem, M., Bethke, A. and Reddy, S. StereoSet: Measuring Stereotypical Bias in Pretrained Language Models. Proceedings of ACL-IJCNLP 2021. Stroudsburg, PA: Association for Computational Linguistics.	2021	Shows measurable stereotypes in pretrained models across gender, profession, race and religion.	ACL Anthology
5	Nangia, N., Vania, C., Bhalerao, R. and Bowman, S. R. CrowS-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models. Proceedings of EMNLP 2020. Stroudsburg, PA: Association for Computational Linguistics.	2020	Tests whether language models prefer stereotyped sentences over fairer alternatives.	ACL Anthology
6	The Conversation. Why Romania's Election Was Annulled, and What Happens Next? The Conversation.	2024	Explains the court's annulment after intelligence documents pointed to social media manipulation and illegal TikTok campaign financing, and notes the mixed reaction.	The Conversation
7	TechCrunch. As US Weighs Response to Chinese AI, Industry Urges Against Broad Open-Weight Restrictions. TechCrunch.	2026	Reports the industry open letter opposing broad restrictions on open-weight models and the policy debate behind it.	TechCrunch
8	The Next Web. Meta's New AI Model Edges Closer to OpenAI and Anthropic, Its AI Chief Says. The Next Web.	2026	Records that Meta's first Muse Spark model launched closed, and its mixed plans for opening later versions.	The Next Web
9	Rubin, J. Printing and Protestants: An Empirical Test of the Role of Printing in the Reformation. The Review of Economics and Statistics, 96(2), pp. 270-286. Cambridge, MA: MIT Press.	2014	Finds that cities with a printing press by 1500 were substantially more likely to be Protestant by 1600.	DOI