Every answer and its explanation appears here once you have finished the path. Each one then links to the matching glossary entry, where the concept is set out in full with its worked example.
1. A firm uses a large language model to pre-analyse insurance contracts. The summaries read well, but the model sometimes introduces clauses that do not exist. What task is the model actually performing?
It predicts the most probable token given the preceding context
A large language model is trained on vast text corpora for a single, apparently modest task: predicting the next word, or more precisely the next token, given the preceding context. Complex capabilities emerge from that task, drafting, translation, summarising, apparent reasoning, code generation, and the transformer architecture lets it weigh the relative importance of different parts of a text. Understanding this mechanism is not a theoretical nicety, it explains the observed behaviour. The model does not search, does not verify and does not compare: it produces the most plausible continuation. A non-existent but perfectly idiomatic clause is therefore, from the model's standpoint, an excellent result, being exactly what a contract of that kind usually contains. Fluency of output says nothing about its truth, and it is precisely because the two are independent that human review cannot be removed.
Glossary entry · llm2. A compliance agent produces a report citing a regulatory article that does not exist, phrased with the same confidence as the accurate references. How should this behaviour be characterised?
A structural feature of how these models work, reducible but not eliminable
Algorithmic hallucination is the production of false or invented information, presented with a high apparent degree of certainty and with no signal distinguishing it from a correct answer. It follows directly from the probabilistic nature of large language models, which generate statistically plausible sequences rather than verified facts: it is therefore not an accidental defect but a structural feature, impossible to eliminate entirely and only reducible. Two consequences follow for anyone deploying these tools. The first is that waiting for a fixed version is a strategy without an end date, since the flaw lies in the mechanism and not its implementation. The second is graver: the absence of any distinguishing signal makes the error undetectable on reading, a fabricated citation looking exactly like an accurate one. That is what shifts the burden onto systematically checking references rather than reviewing style.
Glossary entry · hallucination-algorithmique3. An internal assistant answers staff by drawing on the company's document base rather than on its parameters alone. What attack surface does this architecture open?
A corrupted document slipped into the base can hijack the assistant, with no intrusion into the model at all
Retrieval-augmented generation pairs a large language model with document search: instead of answering from the knowledge frozen in its parameters alone, the model first queries a base of relevant documents and grounds its answer in the retrieved extracts. The benefit is real and twofold, anchoring answers in verifiable, current sources, which markedly reduces hallucination risk, and letting the system reason over private or recent information the model has never seen. But the architecture moves where trust sits: it is no longer the model that must be secured, it is the base. A corrupted document slipped into the corpus can carry a hidden instruction that hijacks the assistant, and the attack requires no intrusion into the model at all, only the right to add a file. For an insurer, that moves the underwriting question from the model vendor to write access controls on internal documentation.
Glossary entry · retrieval-augmented-generation4. A company specialises a foundation model on its own files to automate claims handling. What particular risk does that operation create?
The model can reproduce, in its answers, confidential data present in the training files
Fine-tuning specialises an already pre-trained model by giving it further training on targeted data particular to a task or domain. Rather than training from scratch, which would demand colossal resources, it builds on general knowledge already acquired and steers it toward a specific use, contract analysis, diagnosis, claims handling. This technique is what democratised artificial intelligence, putting it within reach of organisations that could never have built a model. The risk peculiar to the operation is that it writes the training data into the parameters themselves: badly controlled fine-tuning leads the model to reproduce, in its ordinary answers, confidential elements from the files used to train it. The leak then passes through no database and no intrusion, it passes through the model, which makes it hard to detect and impossible to claw back once the model is distributed.
Glossary entry · reglage-fin5. The term foundation model was popularised in 2021 by researchers at Stanford. What defines such a model?
A model built to serve as an adaptable base for a multitude of downstream applications
A foundation model is a very large model, trained on massive volumes of data, and built not for one task but to serve as an adaptable base for a multitude of downstream applications, from dialogue to code writing to image analysis. The Stanford researchers who popularised the term in 2021 stressed both its power and its risk, and both flow from the same property. The power lies in sharing: one pre-trained model can be reused and specialised cheaply, putting artificial intelligence within reach of organisations that could never have built one. The risk is exactly the same thing seen from the other side: when dozens of different tools rest on the same base, a reasoning flaw common to that base produces the same error in all of them at once. That is accumulation, in the sense insurance gives the word, and it is invisible in historical data because it did not exist before.
Glossary entry · modele-fondation6. A system declines a claim on reasoning that neither the insured nor the insurer can reconstruct. Where does that opacity come from?
From the nature of the model, whose decision emerges from billions of parameters with no interpretable representation
The algorithmic black box is a system, typically a deep neural network, whose inputs and outputs can be observed but whose internal path to a given decision remains largely inexplicable, including to its own designers. The distinction from commercial secrecy is essential and often missed: opening the code or publishing the weights would not make the decision explicable, because the opacity protects nothing, it results from how the decision forms, through the interaction of billions of parameters in very high-dimensional vector spaces, with no interpretable symbolic representation. The consequence for risk management is direct: a refusal that cannot be justified is still a refusal, but it becomes indefensible before a policyholder, a court and a supervisor. It is therefore not merely a technical problem, it is what makes certain decisions unusable whatever their accuracy rate.
Glossary entry · boite-noire-algorithmique7. A bank declines a loan on the recommendation of a model whose decision it cannot explain. Which central dilemma does this illustrate?
The most powerful models are frequently the least interpretable
Explainability is the ability to make a model's decisions or predictions comprehensible to a human being. It stands against the black box character of the most complex models, whose performance often comes with opacity such that even their designers struggle to explain a particular decision. The tension between performance and transparency is one of the central dilemmas of applied artificial intelligence, and it is real rather than rhetorical: the most powerful models are frequently the least interpretable, so gaining accuracy can cost defensibility. The declined loan shows the concrete stake, since three parties want different things from the same model: the customer wants a reason, the regulator wants a method, the institution wants performance. Choosing a slightly less accurate but explainable model is therefore not a technical retreat, it is a trade-off between error rate and the ability to answer for one's decisions.
Glossary entry · explicabilite8. A control requires a human to approve every automated decision. In practice the operator signs off hundreds of cases a day without examining them. What is left of the safeguard?
It exists on paper and controls nothing at all
The human in the loop principle keeps a human intervention, as supervision, approval or the ability to take back control, inside an automated decision process, and it is almost always presented as the ultimate safeguard against growing system autonomy. Several degrees are worth naming: human in the loop, approving each decision; human on the loop, watching a flow and intervening by exception; and human out of the loop, merely observing after the fact. The case described shows why the declared degree is not enough: an operator approving hundreds of cases a day without examining them is formally in the loop, while in fact intercepting nothing. The safeguard is then documented, auditable and empty. Which is why the useful question is not whether a human approves, but how much time they have per decision and how often they refuse one.
Glossary entry · humain-dans-la-boucle9. A fifty-employee company runs a fleet of sales agents and commits more legal acts per day than a five-thousand-person group. Which property of an autonomous agent produces that gap?
Tool access, which gives it an effect on the outside world
An autonomous agent differs from a language model by four technical properties, each carrying an insurance consequence, and they are worth keeping apart. Tool access gives it an effect on the outside world, turning an error of content into physical or financial damage: this is the property producing the case here, since the number of acts committed stops being bounded by the number of people. The execution loop, by which the agent observes the result of its action and continues, allows an initial error to propagate across a long sequence before any detection. Persistence, the ability to operate over time without being prompted, removes synchronisation with human attention. The underwriting consequence is direct and already measurable: a premium computed on headcount, the ordinary method in professional liability, entirely misses the risk of a company whose capacity to act is no longer proportional to its staff.
Glossary entry · agent-autonome