A text predictor. That’s it.
Let’s talk about what AI actually is.
An LLM (Large Language Model, the engine behind ChatGPT, Claude, Gemini and the rest) is not intelligent. It’s a statistical model that predicts the most probable next word in a sequence. Then the next one. Then the next one.
It doesn’t understand what it writes. It doesn’t reason. It produces text that looks like reasoning because it was trained on billions of texts that contained human reasoning. The difference is fundamental… and it’s systematically glossed over in the current discourse.
The term has existed since 2021. Emily Bender, Timnit Gebru and colleagues called them “stochastic parrots” in a now-landmark paper: On the Dangers of Stochastic Parrots. Over 8,000 citations since. Systems that reproduce statistical patterns without understanding what they’re saying. The name stuck.
When an LLM gives you a convincing answer, it’s not because it understood your problem. It’s because the structure of your question matched a pattern it’s seen millions of times before. Change one contextual parameter… and the answer can become absurd.
Hallucinations are not a bug
People talk about AI “hallucinations” as a flaw to be fixed. A temporary issue. The next version will sort it out.
No.
Hallucinations are a direct consequence of how the model works. A text predictor has no way of knowing whether what it generates is true. It knows it’s probable. Probable and true are not the same thing.
And the numbers are not trivial. According to a study published in Nature in 2025, hallucination rates range from 15% to 52% depending on model and context. In the legal domain, a recent analysis found rates of 69% to 88% on high-stakes queries. Even the best current models hover between 8% and 17% under favourable conditions.
The worst part is that nothing in the form tips you off. The text is fluent, structured, confident. An LLM that cites a non-existent source isn’t “making a mistake”. It’s doing exactly what it was designed to do: produce a statistically coherent sequence of words. Whether that sequence describes reality or not… that’s not part of the equation.
The cost nobody wants to see
AI is sold as an efficiency gain. People talk less about what it costs.
Not the subscription. Energy. Water. Infrastructure.
Training GPT-3 consumed 1,287 MWh of electricity. The equivalent of 120 American households for a year. 552 tonnes of CO2. GPT-4? Roughly $100 million in training costs. And that’s just the training, not the daily use by hundreds of millions of people.
On the usage side, 2.5 billion queries per day for ChatGPT alone. A volume that doubled in six months. And for what? Rephrasing a three-line email. Summarising a paragraph you could have read. Generating a joke for a WhatsApp group. Asking for a weekly meal plan. A good portion of these queries doesn’t justify the computing power involved.
Then there’s water. A standard data centre uses over a million litres of water per day for cooling. The largest ones, close to 19 million litres per day. The equivalent of a town of 50,000 people. In 2024, 78% of the water used by Google for its data centres was drinking water.
The results show up in carbon reports. Google’s emissions: +48% in five years. Microsoft: +29% between 2020 and 2024. Same cause in both cases: data centre expansion for AI.
This isn’t an argument against using AI. It’s an argument against using it for everything.
Meanwhile, in the office
While senior management wonders whether they should “do something with AI”… their teams are already using it.
ChatGPT for writing emails. Copilot for code. Online tools to summarise internal documents. No policy. No framework. Nobody knowing what data is passing through these services.
The numbers speak for themselves. According to a 2026 Salesforce survey, 67% of employees use AI tools at work. Only 18% of organisations have an AI security policy. 38% of employees share confidential data with AI platforms without authorisation. And 60% believe the time savings justify the risks.
An employee pasting a confidential document into ChatGPT for a summary isn’t doing anything malicious. They’re doing what everyone does: looking for a shortcut with the tools available. The problem is that nobody has set a framework.
It’s called “shadow AI”: employees using AI tools without the company knowing or having authorised it. It’s the equivalent of “shadow IT” from the 2010s, when teams installed their own software without going through IT. Except this time, it’s faster, more discreet, and the data leaking out is often more sensitive.
Data quality. Always.
AI doesn’t create quality. It amplifies what already exists.
Clean, structured, well-documented data? AI can do remarkable things. Messy data, poorly named, no metadata? AI will automate the chaos. Faster.
The industry prefers not to dwell on this point. “AI will solve everything” sells better than “start by cleaning up your data”. But no layer of AI, however sophisticated, compensates for shaky foundations.
You see it very concretely in DAM. Vendors integrate visual recognition models like YOLO (You Only Look Once), designed to detect and classify objects in an image. Nothing to do with an LLM. But the principle is the same: AI sold as a turnkey solution.
Except these models are usually trained on generic datasets. They recognise a “dog” or a “car”. Not your product categories, your business typologies, your internal conventions. For that, you’d need a custom dataset. Manually annotating thousands of images costs between $0.03 and $1 per image, not counting design and quality control. Nobody mentions that during the demo.
And even with a well-trained model… tagging at 90% accuracy across 100,000 assets means 10,000 errors. The model doesn’t raise its hand when it gets it wrong. To find the 10,000 errors, you have to check all 100,000. 90% accuracy doesn’t mean 90% of the work is done. Far from it.
What AI actually does well
This isn’t an anti-AI indictment. It’s a recalibration.
AI is useful when applied in the right place, with the right expectations:
- Repetitive, predictable tasks. Classifying documents, extracting structured information, pre-filling forms. Anything that follows an identifiable pattern.
- Research and exploration. Searching through a corpus, spotting trends, suggesting leads. AI as a research assistant. Not an oracle.
- Production acceleration. Generating a first draft, proposing variations, translating. Provided a human validates the output.
The common thread: in each of these cases, AI is a supervised tool. Not an autonomous decision-maker.
The real issue is scoping
The question isn’t “should we use AI?” Obviously yes, in certain cases.
The question is: on what, how, with which data, and who checks?
Without scoping, AI becomes a problem generator disguised as a solution. With serious scoping, it can save considerable time on specific tasks.
The difference between the two isn’t the technology. It’s the work you do upfront.