Blog - When AI becomes Chinese whispers

Lez Graham discusses AI

When AI becomes Chinese whispers

a woman gazing away from a book with a cascading line of images that change slightly each time.

What Happens When Repetition Starts to Sound Like Evidence

A small moment during tutoring recently left me thinking about something much bigger, because in a discussion about the history of dog training and learning theory, Thorndike had been largely left out while Skinner was presented as the important figure in operant conditioning, and Thorndike was reduced to somebody who had merely developed a law.

This was not, in itself, an AI problem, nor am I criticising a student for not already knowing something they are there to learn; it was simply one of those moments that made my mind start whirring, because if an incomplete version of a subject is written down, published and then repeated often enough, it can gradually become the version that everybody accepts, even though repetition has not made it any more accurate.

Now that more people are using artificial intelligence to generate blogs, articles, course material and social media posts, I think that matters enormously, particularly in dog training, where there is already a great deal of information online that is incomplete, lacking context or simply wrong.

When an omission becomes the story

Edward Thorndike's contribution was not a minor historical footnote. In Animal Intelligence, published in 1911, he set out his Law of Effect, describing how responses followed by satisfying consequences become more likely to recur, while those followed by discomfort become less likely, and his puzzle-box research helped to establish the experimental study of how consequences affect behaviour.

B. F. Skinner later developed operant conditioning into a much fuller and more systematic account, but that does not make Thorndike unimportant; scholarly histories of the experimental analysis of behaviour describe his work as foundational to what came later, which is why my view is that leaving him out, or reducing his contribution to an insignificant preliminary step, creates a distorted understanding of how the field developed.

There are two different things in that paragraph and I think it is important to separate them. The date of Thorndike's work, the content of the Law of Effect and the influence of his research can be checked against original and scholarly sources; my judgement that omitting him creates a distorted account is an opinion, although it is an opinion based upon those facts and upon my experience of teaching learning theory.

That distinction matters because not everything in dog training has one simple, provable answer. We have genuine debates, developing evidence, differences in terminology and areas where experienced practitioners interpret the same information differently, but we also have statements that can be checked against learning theory, physiology, published research and the documented history of the subject, and a confident explanation does not remove the need to do that checking, regardless of whether it was written by a human or generated by AI.

How digital Chinese whispers begin

When people talk about AI scraping the internet, it can sound as though every system is continually reading every new webpage and immediately adding it to its permanent knowledge, but that is not how all AI systems work. Some tools retrieve current webpages while answering a question, people copy and republish AI-generated material, and content placed online may later be included in collections used to train or refine other models, so there are several different routes by which one account can influence the next.

The practical problem is that the reader may never see those routes. One inaccurate statement appears in a blog, another writer repeats it, an AI tool retrieves or reproduces it, and then several websites carry polished versions of the same claim; by the time somebody searches the subject, those pages can look like independent agreement when, in reality, they are all echoes of the same original mistake.

To me, that is a digital version of Chinese whispers, except that the message is not merely being misheard as it passes from one person to another; it is being rewritten fluently, presented confidently and repeated at a scale that can make familiarity feel like proof.

If the shortened version of Thorndike's place in history were published on a website, copied into another article, retrieved by an AI tool and eventually included in later teaching material, the omission could become part of the accepted story, particularly if nobody producing the next version returned to the original sources. Thorndike would not become less important because fewer articles mentioned him, but he could begin to look less important simply because the same incomplete account had been repeated more often.

Dog training already has enough misinformation

I have worked professionally with dogs since 1998 and, over that time, I have watched fashions, explanations and supposed certainties come and go, often with great confidence attached to them, which means I do not need AI to tell me that inaccurate dog-training information can spread quickly; we were managing that perfectly well by ourselves long before generative AI arrived.

What AI changes is the speed, volume and appearance of it. A person who does not have the knowledge or experience to write an evidence-based article can now produce something polished in seconds, and because the language is clear and authoritative, the result may look more reliable than it is; another person can then use that article as the basis for a post or course handout, and each retelling can move a little further away from the evidence while sounding increasingly settled.

This does not mean that everything written with AI is wrong, any more than everything written without it is right. It means that fluent language must not be confused with expertise, that a list of apparent sources is not enough unless those sources genuinely support the claims being made, and that somebody still has to understand the subject well enough to notice when an important part of the story is missing.

That last point concerns me most, because an obvious error can be challenged, whereas an omission is quieter. If Thorndike is not mentioned, a reader who has never heard of him does not know that there is anything to question, and an AI system asked to summarise a collection of similarly incomplete material may produce an answer that is internally consistent, beautifully written and still missing something foundational.

What happens to original thought

There is also a wider question about creativity, because if the internet becomes increasingly filled with articles generated from what has already been written, and those articles then influence the material produced next, where is the space for people who think outside the box, connect ideas across different disciplines, develop a new technique or find a better way of explaining something that has never quite made sense before?

Researchers have demonstrated a related technical problem known as model collapse, in which models trained recursively on model-generated data can gradually lose less common parts of the original information and accumulate errors over successive generations. That research examined training data and model behaviour, not dog-training blogs and not human creativity, so it would be wrong to claim that it proves original human thought will disappear; nevertheless, it gives some substance to my concern that a digital world filled with recycled AI material could favour the most common explanations while making unusual observations, minority knowledge and genuinely new ideas harder to find.

Of course, unusual does not automatically mean correct and new does not necessarily mean better. Innovation still needs evidence, testing, scrutiny and, in dog training, careful consideration of the dog in front of us, but progress has always depended upon somebody being prepared to notice what does not quite fit the accepted explanation and then investigate it rather than simply repeating what everybody else has said.

If AI-generated consensus is continually fed back into itself, we risk mistaking what is most frequently repeated for the full extent of what can be known, and that would not only spread misinformation; it could narrow the questions we think to ask in the first place.

This is not an argument against AI

I use an AI agent myself and find it extraordinarily useful, including when I am trying to untangle one of my long voice notes into something another person might actually want to read. It can help me search for original sources, compare accounts, test whether a statement says what I think it says, challenge an overconfident sentence and organise a collection of thoughts that began while I was walking the dogs or drinking my coffee.

This article is an example of that process. The tutoring experience, the Thorndike example, the dog-training observations, the questions about creativity and the opinions are mine; they began as my own rather rambling thoughts, while my AI agent helped me research the factual claims, distinguish what the evidence supports from what I believe it may mean, and shape the result into an article, but the judgement about what belongs here and the responsibility for what I publish remain mine.

For me, that is the important dividing line. AI can assist with research and writing, but it should not replace subject knowledge, professional judgement or the willingness to return to an original source, particularly when the answer it gives us is neat, plausible and exactly what we expected to hear.

Familiarity is not evidence

The immediate danger is not necessarily that AI will invent one spectacular falsehood that everybody accepts as true, but that thousands of small inaccuracies, omissions and oversimplified explanations will be copied, condensed and repeated so often that they gradually settle into the background as common knowledge, becoming harder to recognise and question precisely because we have encountered them so many times before.

It is rather like taking a photocopy of a photocopy and then continuing to copy each successive version. At first, the differences may be barely noticeable and every copy may still look close enough to the original to pass without question, but each repetition loses a little more detail, alters the contrast and blurs the edges, until what remains is recognisable only in outline and no longer bears much resemblance to the original, accurate image.

Those of us who teach, write and publish have always had a responsibility to check our facts, make it clear when we are offering an opinion and correct ourselves when better evidence becomes available, and AI does not remove that responsibility; if anything, the speed at which it can reproduce our words makes that responsibility greater.

The mistake does not become true because it has travelled further, the omission does not become harmless because it has been repeated, and the most familiar version of a story is not necessarily the most complete one.

Sometimes the most important question is still the simplest: where did this information come from, and has anybody gone back to check?