Who gets to teach tomorrow’s AI what is true?

I started thinking recently about whether something I write today could eventually find its way into a future AI model.

An idea. An observation. Something I’ve noticed that somebody else hasn’t written about.

Could it end up contributing to an answer someone receives years from now, even if hardly anybody reads the original article?

I quite like that thought.

Someone could contribute something worthwhile without having a particularly large audience. Their article might barely be read, yet potentially become part of the information a future AI learns from.

Then I realised I was also describing quite an interesting opportunity for a propagandist.

Because if I can publish something in the hope that a future AI learns from it, presumably somebody with an agenda can do the same.

And they might be considerably more organised about it.

What if you don’t need anyone to read it?

Imagine a company has done something unpopular.

It wants its version of events to become the accepted explanation. So it publishes an article explaining why, actually, everything was perfectly reasonable.

You read it on the company’s website and think: “well, you would say that.”

But suppose you encounter that explanation somewhere else.

Then somewhere else again.

Different websites. Different authors. Slightly different wording, but all arriving at much the same conclusion.

Now it looks less like the company defending itself and more like everybody agreeing with it.

What happens when an AI system encounters those articles?

Would it recognise that the apparent agreement came from one coordinated effort?

Or could the company’s preferred explanation eventually come back sounding like an independent assessment?

There are two ways this could potentially happen.

Publicly available material can become part of the data used to train future models. Alternatively, AI services can retrieve information from websites when answering questions, much like Google’s AI Mode does today.

They’re different mechanisms. Something appearing in a search result doesn’t mean a model has permanently learned it, and publishing an article certainly doesn’t guarantee it will influence anything.

But both create opportunities for outside information to shape the answers we receive.

That’s the part that interests me.

You haven’t asked the company for its opinion.

You’ve asked something you expect to help you make sense of the situation.

And there’s a possibility you’re hearing the company’s opinion anyway, just without the company standing visibly behind it.

It isn’t quite “most articles wins”

Obviously, this isn’t as simple as publishing ten thousand articles saying I’m the world’s greatest writer and waiting for the machines to agree.

Which is disappointing, because that would save quite a lot of effort.

AI models aren’t simply counting articles and deciding that whichever opinion appears most frequently must be correct.

Training data is selected, filtered and processed. Duplicates can be removed, and the way models are subsequently trained also influences their behaviour.

So flooding the internet with the same opinion doesn’t automatically mean future AI systems will adopt it.

But the opposite assumption, that these systems are simply too large or sophisticated to manipulate, doesn’t seem particularly safe either.

Researchers working with Anthropic, the UK AI Security Institute and the Alan Turing Institute found that just 250 deliberately constructed documents could introduce an unwanted behaviour into the models they tested.

The behaviour involved making models produce gibberish when triggered, not convincing them to adopt a political position. So it’s important not to exaggerate what the research demonstrates.

But it does show that manipulating training data is a real technical concern, rather than something we’ve invented for a science-fiction film.

Anthropic’s research is worth a read if that sort of thing interests you.

And people are already using AI to support influence campaigns. OpenAI reported disrupting operations producing articles, comments and fictitious identities back in 2024, although it also found that those particular operations hadn’t gained much additional reach through its services.

Producing enormous amounts of content and successfully influencing people are obviously two different things.

Still, the incentive seems fairly clear.

Why settle for trying to persuade one audience when you could potentially influence something that millions of people consult for explanations?

Not just:

How many people can we get to read this?

But:

How many future answers could this find its way into?

But haven’t we always done this?

There is an obvious argument against everything I’ve just described.

Propaganda isn’t new.

Governments have been influencing historical accounts for centuries. Newspapers have shaped public opinion. Publishers have decided which books reach audiences, and search engines already influence what information people discover.

So why would AI make any of this fundamentally different?

I don’t think the underlying human behaviour is different.

People have always tried to shape what other people believe.

What interests me is how many stages of that process AI could potentially become involved in.

AI can help produce the original material.

It can help distribute enormous amounts of it. It can retrieve that material when somebody asks a question. It can combine different sources into a single explanation.

And it can help people write the next round of articles based on that explanation.

Each stage provides another opportunity for the original message to circulate, potentially further removed from whoever first introduced it.

And unlike a newspaper article or a list of search results, an AI response can arrive as a direct, conversational answer to your question.

I’m not suggesting that automatically makes it more persuasive, but it does make me wonder whether we’d question it in quite the same way.

There’s another complication, too.

The agenda might not even be a lie

This is where “just fact-check it” starts feeling like only part of the answer.

Imagine a factory closure.

One account describes a business facing rising costs, difficult decisions and an uncertain future.

Another describes hundreds of people losing their jobs after years of decisions they had no control over.

Both accounts could be accurate.

They would still leave you with quite different impressions.

Now imagine one interested party producing large amounts of material around the first explanation.

Not necessarily inventing facts. Just repeatedly presenting a particular selection of them, while the people in the second account receive much less attention.

Could that affect which explanation an AI presents as the natural place to start?

Could it affect which questions it thinks need answering?

Perhaps creating the appearance of disagreement would be enough.

An answer that ought to say “the evidence strongly points this way” could potentially become “there are conflicting views”.

I’m not claiming we know how reliably somebody could achieve that. But the possibility is interesting.

And disagreement itself isn’t the problem.

People should be free to challenge accepted explanations, put forward new interpretations and question conventional thinking.

Otherwise, the original idea that appealed to me disappears immediately.

The concern is somebody disguising one coordinated voice as lots of independent ones.

Ten articles repeating the same unsupported claim don’t give us ten independent pieces of evidence.

And an AI repeating those articles doesn’t add an eleventh.

Meanwhile, why do we all sound so polished?

There’s another side to this that doesn’t require anybody to be deliberately misleading.

What happens when the tools helping us write also start influencing what we consider good writing?

AI has a rather recognisable way of expressing itself sometimes.

Everything is organised. Everything is carefully explained. Every point is covered.

And somehow it can still be incredibly boring to read.

We’ve probably all been there, asking AI a fairly simple question and receiving an enormous wall of text when we wanted about 1% of it.

But it made me wonder about something else.

What happens when millions of people start relying on these tools to make their own writing “better”?

A smoother introduction.

A more familiar expression.

A conclusion that neatly ties everything together.

On their own, those might all be improvements.

But what happens when everybody keeps accepting them?

There’s actually some interesting research suggesting this is worth thinking about.

In a 2024 experiment, participants given AI-generated ideas produced short stories that received better ratings, but the assisted stories were also more similar to one another.

So the individual pieces improved, while the collection became less varied.

The University of Exeter published an interesting summary of the research.

That’s quite an interesting trade-off.

Perhaps nobody sets out to make writing more uniform.

They just keep accepting the nicer sentence.

The word that sounds more professional.

An ending that feels more complete.

Eventually, could we reach a point where writing that doesn’t follow those patterns starts to feel wrong?

Think about how much we’ve already come to rely on technology to correct the way we communicate.

Spellcheck tells us when we’ve spelt something incorrectly (and don’t get me started on the Americanisms slipping through because of it!)

Autocorrect predicts what we’re trying to say.

Grammar tools suggest how we should structure our sentences.

Now AI can rewrite entire paragraphs, arguments and articles.

We’ve gradually moved from tools that correct our mistakes to tools that can decide how best to express our thoughts.

Which made me wonder where that progression eventually leads.

Could regional expressions start looking like mistakes that need correcting?

Could somebody’s slightly clumsy but distinctive way of explaining something become a perfectly competent explanation that anybody might have written?

And could those habits eventually spill into the way we speak?

I don’t think that’s particularly far-fetched as a question. We’re already influenced by the language we encounter in books, films, television and social media.

Why wouldn’t something we increasingly interact with every day have some influence too?

There is even early research suggesting that brief interactions with chatbots can affect the vocabulary people subsequently use, although that doesn’t establish what the long-term consequences might be.

I’m not suggesting civilisation is in trouble because somebody says “delve”.

But imagine a phrase that initially sounds distinctly AI-generated.

You start encountering it in articles.

Then emails.

Then conversations.

Eventually, it stops sounding unusual.

You might even start using it yourself.

At that point, you wouldn’t need AI to suggest it anymore.

Which brings me to something I find even more interesting.

If AI increasingly influences how we express our thoughts, could it eventually influence which ways of expressing an idea we consider sensible, professional or convincing?

We might not necessarily be changing what we think.

But could we gradually be changing how we think something should be said?

And if everybody is moving towards similar ways of expressing themselves, what happens to the unusual voices, expressions and ways of looking at things that make writing interesting in the first place?

Of course, there’s an obvious argument against this.

Language has always changed.

We borrow words and expressions from the people around us. Television, the internet and social media have all influenced how we communicate.

Why should AI be any different?

Perhaps it isn’t, at least in principle.

But I wonder whether the difference is that we’re now interacting with something that doesn’t just introduce us to new language. It actively suggests how we should use it.

And increasingly, it can do that whenever we write.

There are benefits to that.

Someone who struggles to express an idea might finally be able to explain what they mean.

Somebody writing in another language might feel confident enough to contribute something they otherwise wouldn’t have shared.

That could make the conversation considerably richer, even if some of the wording becomes more alike.

So I don’t think the answer is to protect every awkward sentence from improvement.

But I do wonder whether we should be paying more attention to the difference between helping somebody find their voice and gradually teaching everybody to use the same one.

…Then we feed it all back in

This is where the two thoughts start joining up.

AI helps us write something.

We publish it.

Other people encounter it, repeat parts of it, or use AI to produce something similar.

Some of that material could then become available to future models.

And those models help people write the next round.

There’s a potential feedback loop there.

It wouldn’t just contain deliberately planted propaganda.

It could carry forward ordinary mistakes, familiar phrases, popular interpretations and whatever preferences the writing tools themselves had encouraged.

There’s a related technical problem called model collapse, where some experiments have shown models losing aspects of the original information when repeatedly trained on generated material.

That’s not evidence that AI is inevitably going to ruin itself by reading its own homework. Other research has found ways to avoid that outcome by retaining real data alongside generated material.

But the broader question interests me.

Where does the fresh evidence enter the loop?

At some point, somebody needs to have observed the event.

Read the original document.

Conducted the experiment.

Actually experienced whatever it is we’re discussing.

Otherwise, I could imagine a claim travelling through articles, summaries and AI answers until we stop noticing that everybody is referring to everybody else.

It would look thoroughly discussed.

It might still be completely unsupported.

And that brings me to something else I came across.

The books

I recently read about AI companies buying physical books and having them destroyed during digitisation.

“That’s.. alarming.”

But I wanted to understand what had actually happened before making too much of it.

A June 2025 court order recorded Anthropic purchasing millions of printed books and having them destructively scanned.

Bindings removed. Pages scanned. Physical copies discarded.

The company retained digital copies for its internal library.

The court document is available here.

Now, there’s an important distinction to make.

Destroying a copy isn’t the same as erasing a work.

The court record doesn’t establish that these were the last surviving copies, or that their contents were being rewritten.

But it does raise an interesting question about what we consider to be preserved, and for whom.

If a book has been digitised somewhere inside a company, that doesn’t necessarily mean I can read it.

And being able to ask an AI about a book isn’t the same as having access to the book itself.

Imagine an older, out-of-print novel.

You ask AI what the author was trying to say, and it gives you a convincing explanation.

You use that explanation in an article.

Somebody else reads your article and writes something similar.

Except perhaps the narrator wasn’t supposed to be trusted.

Perhaps the ending was deliberately unresolved.

Perhaps the contradiction the summary helpfully tidied away was actually the whole point.

Could the explanation eventually become more familiar than the book?

The original words might remain unchanged, while the version people understand becomes something rather different.

Think about something like Shakespeare.

There’s nothing wrong with reading an explanation of Hamlet. I’ve certainly benefited from somebody explaining something I didn’t understand.

But imagine a future where most people encounter the play primarily through AI-generated summaries rather than the play itself.

What happens to the different interpretations?

The ambiguities?

The things that make people disagree about what the author actually meant?

Perhaps nothing particularly terrible happens. Maybe AI encourages more people to discover and read works they would otherwise have ignored.

I’d be delighted if that’s the outcome.

But I do wonder whether convenience could eventually make the explanation more influential than the original.

Not because somebody has secretly rewritten every surviving copy.

Because the version most people encounter is no longer the original work.

And if future articles and models increasingly draw on those explanations, rather than returning to the text, could some of those interpretations gradually become accepted as what the author actually said?

There’s a distinction worth keeping in mind.

A faithful digital copy lets us check an interpretation.

A confident answer asks us to accept one.

Someone still chooses what gets through

The obvious response to many of these problems is better filtering.

Remove manipulation. Keep reliable sources. Preserve the originals.

All of which sounds sensible until somebody has to decide what qualifies.

Who decides which interpretations are credible?

Which sources are trustworthy?

Which writing is good enough to keep?

An unfamiliar perspective isn’t automatically unreliable. An established publication isn’t automatically right. And something written in polished English isn’t necessarily better evidence than a messy first-hand account.

Researchers have already identified examples of dataset filtering disproportionately excluding material associated with particular groups.

So even decisions made with perfectly reasonable intentions can affect whose experiences and perspectives remain represented.

I wouldn’t want defending against manufactured agreement to accidentally reinforce whatever is already most familiar.

Equally, including different perspectives shouldn’t mean treating every unsupported claim as equally credible.

Both things need to be possible.

I think being able to trace information back to its origins becomes increasingly important here.

Not just seeing that an AI has provided five citations.

But understanding whether those five sources independently support the answer, or whether they all lead back to the same person making the same claim.

And I’d want the original material to remain available wherever possible.

Not because originals are automatically true.

Because we need something we can inspect, disagree with and compare against the explanation.

I still like the idea I started with

There’s something genuinely worthwhile in the possibility that an ordinary person could contribute to what future systems know.

Someone could document an overlooked experience.

Challenge a familiar interpretation.

Write something that matters without first having to become famous enough for people to listen.

I wouldn’t want to lose that.

But the same opening could be used by somebody with very different intentions, and a considerably bigger budget for producing material than the rest of us.

Alongside that deliberate pressure, there could be a quieter change in the words we use, the way we express ourselves and the kinds of explanations we come to accept.

I don’t think any of this leads inevitably to rewritten history or a world where everybody sounds the same.

And I’m not suggesting that AI is somehow the first technology capable of influencing public opinion.

But I do wonder whether we’re approaching a point where the information future AI learns from is increasingly produced by people who have themselves been influenced by earlier AI.

Some of that influence could be deliberate.

Some of it might be completely accidental.

And perhaps distinguishing between the two becomes increasingly difficult.

Which makes me think the original becomes more important, not less.

The original account.

The original text.

The person who actually had something to say before it passed through several rounds of being made clearer, smoother and easier to repeat.

I started by wondering whether a thought of mine could come back in a future AI answer.

I still think that’s quite an interesting possibility.

I’d just be a lot less impressed if I mistook it for somebody independently agreeing with me.

Originally published at LinkedIn.