Blog
Articles about how these systems work, how they are evaluated, what obligations they bring and where they fail. No unsourced figures and no promises nobody can check.
20 articles
A language model does not look things up and does not \u201cknow\u201d facts. It predicts the next piece of text. That difference explains almost everything that surprises you about it — including the mistakes.
Read onA \u201challucination\u201d is not a fault that a new version fixes. It is a consequence of how the model works. It can be reduced, though — and, more importantly, detected.
Read onA single percentage does not tell you whether the system works. It tells you how often it is right on average — which is precisely the figure that hides the problems that matter.
Read onA model learns the regularities in its data, including the ones you would rather it did not. The problem is not that the system \u201chas prejudices\u201d, but that it faithfully reproduces something already there.
Read on\u201cAn explainable system\u201d can mean five different things to five different people. Before asking for explainability, it is worth settling who you are explaining to, and why.
Read onThe regulation classifies systems by risk, not by technology. The first practical question is not \u201cwhich model are we using\u201d, but \u201cwhat does the system do, and to whom\u201d.
Read onData protection did not become irrelevant because AI arrived. If anything, a few old principles suddenly became much harder to honour.
Read onHaving someone press \u201capprove\u201d is not oversight. It is a signature on a decision made by something else — responsibility transferred without the control that should come with it.
Read onIf a model reads text from outside your organisation, that text is code arriving at the model. It is a class of vulnerability with no direct equivalent in classical software.
Read onIn public administration, an error does not cost you a customer. It costs someone a right. That changes the benefit-versus-risk calculation entirely.
Read onRetrieval plus generation is the most common architecture for making a model answer from your documents. The hard part is not the model — it is the retrieval.
Read onThree approaches that are often confused and solve different problems. The wrong choice means a month of work for something that took a day — or the other way round.
Read onMost AI projects fail before they start, at the moment the wrong problem is chosen.
Read onThe cost of model calls is usually the small part. Budgets break in places nobody puts in the initial estimate.
Read onGenerated data solves real problems of privacy and volume. It also, sometimes, solves the problem of getting good evaluation results without having a good system.
Read on\u201cOpen source\u201d means very different things in this field. For an organisation, the useful question is not ideological but operational.
Read onThe subject is real, and the figures in circulation are often impossible to verify. What can be claimed with grounding, and what cannot.
Read onPublic debate talks about jobs. What concretely changes inside an organisation are tasks — and the consequences are less obvious.
Read onThe question \u201chow do we stop students using AI\u201d has taken over the conversation. It is probably the least interesting question available.
Read onA short set of questions that separates, in under a minute, a real result from a press release.
Read onEditorial note: Articles on this blog explain mechanisms and methods. They do not describe the association's projects and contain no figures or studies we cannot point to. The field moves quickly — the publication date matters.
