Author: Andrew Dunn
AI is everywhere in 2025. You’d be hard pushed to find an industry today where AI is not making an impact. You can’t even Google something without AI getting itself involved, and it has colonised your email and mobile phone.
In November 2022, when Chat GPT burst onto the scene, Generative AI was used to perform parlour tricks, compared to current usage, and we marvelled at its usability. But its impact on the societal psyche was profound, ranging from curiosity, to excitement and even fears that human skills would be rendered redundant.
But – AI’s inexorable march to world domination seems to be on hold. The human natural ability to handle complex tasks, to reason well, to use experience and intuition, has limited AI’s permeation. The grand realignment of human life has not happened – yet.
LLMs are like digital truffle hunters, but their instinct to ‘hallucinate’ cooled off the excitement
The benefits of Generative AI is undeniable – convenience, speed and the sheer volume of work or data your LLM can handle. An LLM can ‘read’ your 120 page PDF in a matter of seconds and can repurpose the information it contains, in a human heartbeat. When used in this way, LLMs are like digital truffle hunters, sniffing out the answers in countless acres of dense forest. Need to perform a week’s worth of research in an afternoon? No problem… or is there? Well, yes. And it is an enormous and enduring one.
Another ‘feature’ of LLMs has cooled off the excitement – the LLM’s instincts to ‘hallucinate’ or ‘make stuff up’ in its answers. This sounded alarm bells for roles and industries where accuracy is crucial – think financial markets, law, safety systems and so on. Yes, it can give you your answer in no time flat, but how useful is that, if the machine is on a digital acid flashback and is away with the fiction fairies?
All of a sudden, Biden is still US president, Rishi Sunak is the Chancellor of the Exchequer, phantom studies with fake hyperlinks yield extremely impressive, yet absolutely fabricated statistics, quotes are ‘dreamt up’ and attributed to people who don’t exist, history, and even reality is rewritten in front of your blinking and bewildered eyes – and it all looks and sounds so very reasonable.
This happens because LLMs predict the next word based on patterns in their training data, rather than possessing true knowledge
When faced with queries requiring information beyond their internal knowledge or when data is insufficient, they may lie, confidently, emphatically, even joyfully. Reliance on training data alone can lead to these issues, and models designed to always provide an answer, rather than admit uncertainty, are guilty too. Systems like Retrieval-Augmented Generation (RAG), which ground answers in pre-checked, ringfenced sources help, but are not foolproof; RAG systems can still experience hallucinations. Even leading legal AI tools using RAG have eyewatering hallucination rates, between 17% and 33% – ‘cry perjury!’
Hallucinations are exacerbated by one simple human trait – looking for a shortcut
By not understanding what an LLM is and knowing its limitations, not feeding the machine quality data, not prompting it properly, and not understanding how to verify its work, the LLM is unchecked by guardrails, and can get really creative. The true irony is that if you try to ‘cut corners’ in using your ‘corner-cutting’ technology, you will cut accuracy and trust along with time. And often, sorting out this mess of confusion, can add more time to your workflow – you can waste an awful lot of time, ‘saving time’ with AI.
AI doesn’t come with airbags as standard
An LLM is like a hypercar: it ‘looks the business’; it is making a large amount of noise and everyone takes notice; everybody wants to drive one; it can go from 0 to 60 mph in less than 2 seconds and will get you where you want to go so fast it will make your eyes spin; people in lesser vehicles will simply not be able to keep up; but if you don’t know what you are doing with that much power, you will put it in a tree, and the AI doesn’t come with airbags as standard.
Whilst using LLMs is obviously quicker in most instances, the human-in-the-loop still has a great deal of work to do in curating and cleansing data to feed the beast and checking its output for accuracy. These are now non-negotiable elements in our efficient, AI-enabled workflows.
You can’t be angry with the car that crashes when you drive it like you stole it.
LLM – Limited Liability Machine
One thing certainly remains true with the advent of the AI hypercar – walking is no longer an option when your competitors are using 1,000 horsepower. Now it’s time to learn what your car can really do and get on an advanced driving course.
And let’s not forget what happens to your precious data or ideas when you input them into a free version of an LLM. You may as well graffiti your secrets on the main entrance of King’s Cross Station.
So whatever you do, don’t switch off the best LLM in the known universe, the one with 86 billion neurons, that never reaches capacity, the one that evolution built for you – your human brain.
In the end, it’s your name and your company’s brand on your work, and that is a responsibility no AI will take on for you.
AI is just a very powerful tool; don’t let it make a tool out of you.
46-48 East Smithfield, London, E1W 1AW