AI’s voracious appetite

Things get better before they get worse. TV, streaming, the web, social media, mobile phones. AI is no exception.

Journos and writers are very familiar with the ouroboros. This is the Ancient Greek circular symbol of a snake gorging on its own tail. We use this ‘bite the tail’ technique to round off an article by making reference to how it began, providing the reader with a sense of completeness.

Whether by accident or design — and let’s face it, it’s not design — AI (or “a-one”, as the US education secretary insists on calling it) is not only eating itself but making a bit of a meal of it.

AI offerings like ChatGPT and DeepSeek are LLMs (Large Language Models). Most readers will probably know what that means already, but for the uninitiated, LLMs are trained to consume, understand and generate human language. More accurately, LLMs train *themselves*. They do this by digesting trillions of words published online — websites, forums, books, wikis, you name it. Everything it can find.

The problem is that much of what it can find is generated by AI in the first place. ChatGPT alone generates around 100,000,000,000 words every single day (the equivalent of around 120m novels). So the internet is growing pretty quickly, swamped by words and images generated by AI.

Since it ingests what it has already regurgitated, degradation is inevitable (rather like a photocopy of a photocopy of a photocopy). It is imaginatively known as ‘AI model collapse’, and applies to both text and images.

This feedback loop of synthetic data is actually considered to be real at the time it is re-ingested. In other words, AI believes everything it trawls on the internet (rather like any influencer’s audience). Over time, that means the synthetic overwhelms what is genuine. And when synthetic data becomes a significant component of the corpus used for training the AI model, then the model’s representation of reality deteriorates.

Anoraked enthusiasts of thermodynamics will recognise this as an increase in entropy, where meaningful distinctions are lost, and the universe eventually comes to a grinding halt.

Never mind the quality, feel the width

As a consequence of all this, AI often presents false information as factual (which pretty much sums up the internet at large, some might say). This is known as ‘AI hallucination’ or, as Psychology News has it, “bullshitting”.

Researchers have found that AI can hallucinate 27% of the time, and that factual errors can be present in almost half of generated outputs.

If you believe that David Beckham played hockey for China in the 1936 Berlin Olympics, then this is familiar turf. Otherwise, if you are using AI to write your thesis while you go down the pub, just be mindful that AI may well fail the exam for you.

If you take all the ingredients for a cuisine meal and whack them in a blender, you will soon finish up with a muddy green sludge (don’t ask me why it is always muddy green, suffice to say that it is). Chatbot output is becoming increasingly generic and predictable.

And there can be consequences far more damaging than a student who has to resit a term. Companies that rely on AI-generated content can find themselves facing significant reputational damage. Apple, Microsoft and Coca-Cola have all suffered in this regard. The potential for legal risk and defamation is as immense as it is inevitable.

Last year, Grok (produced by xAI, the company founded by Elon Musk) falsely accused NBA player Klay Thompson of vandalism. Grok conflated “throwing bricks” (missing shots) with throwing bricks in a literal sense. But then, Grok is the ‘power’ behind DOGE, so the less said about that, the better. Klay let Elon off with a warning, but would Johnny Depp?

My horse counts better than chatbots

If you accuse a chatbot of being unable to count, it will likely get quite shirty with you. It will insist, “I most certainly can count”. But it will then go on to clarify that by “count” it means something fundamentally pattern-based, and not arithmetic”. AI is so intelligent that it would be discombobulated by Sesame Street.

So by way of experimentation, I asked ChatGPT, very politely, to count the number of characters in a passage of pasted text, where I knew the character count in advance. I tried this dozens of times and the answers came nowhere close (and no two answers were the same). It then became a little short-tempered (I kid you not) and started to push the blame onto me. I had not been precise in my prompt. So I asked it to write the prompt for me that would yield the correct result. Nope. Now it accused me of failing to paste the text to be analysed correctly (not sure how that is possible). Eventually, it completely lost patience and essentially told me to “Naff off and use a piece of software that is better at counting”. Back to the thesis again, and if you are relying on ChatGPT, or any other LLM, for a word count, then what you think is the truth will be off by a country mile.

Seeing is not believing

It’s not enough that AI is eating its own garbage or that it cannot count its fingers and toes. Another layer to the onion is deliberately fake content — disinformation.

Every conspiracy theory is now backed up by fake content in the hope of turning its basement-dwelling content creator into a fabulously wealthy influencer. The Earth is flat, and here’s the evidence to prove it. Or retouched images showing thousands of ‘chemtrails’, or that prove men did not walk on the moon.

Or the polarising political disinformation, designed to achieve anything from damaging a reputation to influencing an election outcome. The president was born in Kenya, climate change is a hoax, the EU is not democratic. If you are a member of a polarised tribe, truth doesn’t come into it. But if you are seeking facts, then the internet is quickly becoming an untrustworthy source. And once trust is lost, it is near impossible to regain it.

Bring on the cat videos.

Conclusion

There is nothing new about AI. I first collided with it more than 30 years ago, and it is more than twice as old as that. What is new is the computing power to process it. But that comes at a huge cost — literally. The energy consumption is such that the AI industry is set to produce more carbon emissions than aviation by the end of the decade (they are near parity now).

But when it comes to text — which is what LLMs are intended for — it is better to consider AI as a tool to augment human effort, rather than to replace it. So the news may not be so bad on the jobs front after all.

We writers, particularly ones so old that we can remember when luggage did not have wheels, can spot AI-generated text a mile off. But that’s probably because we grew up doing things the old-fashioned way. Which, I’m sorry to say, is the best way to mitigate the failings of AI. At a minimum:

• Fact-checking: This seems notoriously unpopular these days (not least among politicians), but you are exposed if you leave it to chance. When researching with AI, tell it to include sources and citations wherever possible (much like Wikipedia) and check them.
• Proof-reading: A lost art, to be sure. Read the generated content carefully. Is it unnecessarily verbose or even circular? Does it use passive voice rather than active voice? And is the tone of voice appropriate for your audience, or does it match the brand voice for your client? Also check punctuation (particularly quotes, em dashes, etc.), grammar and spelling (which will catch you out more often than you might imagine).
• Risk management: Is there a legal or commercial risk if the information is incorrect? For example, financial information is subject to stringent compliance laws. Tell your audience how to use a bench saw incorrectly, and you may just find yourself on the wrong end of a lawsuit. Ensure that subject matter experts are available to sign off on any risk.

Above all else, keep in mind that the content you create is — to some extent, at least — in the hands of a dispassionate data centre, submerged under water halfway around the planet.

In 1998 (I had to look it up), I wrote “The internet is poised to become the world’s largest repository of inaccurate information”. But now that the ouroboros has worked its magic, that statement appears as a recipe for making chicken fajitas with broccoli. Serves 47.

Colin Shanley has been an author for 30 years. Follow him on Medium or Substack.