Generative AI and LLMs
Useful for drafting, summarising or interrogating documents, but exposed to plausible fabrication, confidentiality issues and over-reliance.
The output is a generated proposition, not verified evidence.
Thomas Cole, The Savage State, 1836
I. UnderstandUnderstand AI without mystifying it.
Understand AI without mystifying it.
Start with what the system actually does.
AI systems, work organisation and occupational health.
Understand what a system produces, what it cannot guarantee and when its use remains acceptable. The core competence is not prompting; it is analysing real work.
In 30 seconds
Essential points
- AI is not one object: generative models, predictive systems and algorithmic management raise different questions.
- A language model generates plausible continuations; it does not inherently verify truth, lawfulness or clinical relevance.
- The central occupational-health question is how the system changes tasks, discretion, workload, relationships and responsibility.
EstablishedEmergingAnalysis
On this page
This is an editorial English edition. The French page remains the fuller reference version and contains the complete source list.
Scope
Three forms of AI that should not be conflated
The label “AI” commonly covers at least three distinct realities. Generative AI creates text, images or code from learned patterns. Predictive or decision-support systems estimate a score, classify a situation or recommend an action. Algorithmic management allocates tasks, sets priorities, measures performance or monitors behaviour. Their data, risks and governance requirements differ.
Predictive and decision-support systems
They may shape recruitment, occupational risk targeting, scheduling or access to opportunities. Bias, validation and human review become central.
The higher the consequence, the stronger the validation requirement.Algorithmic management
When a system allocates work, monitors activity or sets performance targets, it changes the organisation of work rather than merely assisting an individual.
Analyse autonomy, contestability and collective effects.Mechanism
How a language model produces an answer
A large language model represents text as tokens, estimates the probability of possible continuations and generates one token after another. Training, fine-tuning, system instructions, retrieval tools and user prompts all influence the output. Fluency can conceal uncertainty because the model is optimised to produce a coherent continuation, not to prove every statement.
Context is finite
The model only sees the information present in its current context and any connected tools. It does not possess unlimited access to the organisation’s knowledge.
Outputs are variable
Small changes in wording, sampling or model version can alter the answer. Repetition and stability testing matter.
Tools change the task
Web search, retrieval or code execution can improve performance, but they also add new failure modes, permissions and traceability needs.
Capabilities
What a model can do — and what it should not be credited with
Models can accelerate drafting, classification, extraction, translation and exploratory analysis. They do not thereby acquire professional judgement, legal accountability or knowledge of the local work situation. A good-looking answer may still be incomplete, outdated or contextually wrong.
Appropriate assistance
Reformulating a public document, producing a first draft, structuring notes or generating alternatives can be useful when a competent user reviews the result.
Sensitive use
Health information, employment decisions, individual recommendations or monitoring require stronger safeguards, validated data flows and clear human responsibility.
False anthropomorphism
Terms such as “understands”, “knows” or “decides” should not obscure the statistical and organisational mechanisms actually at work.
Occupational health
Translate technical properties into questions about work
For every project, ask who uses the system, which task is transformed, whether use is voluntary, what data enter the system, how much verification time is provided, who can challenge the output and who bears responsibility when it is wrong.
Task
Does the system assist, recommend, prescribe or automatically execute?
Data
Are personal, health, confidential or inferred data involved? Where are they stored and reused?
Organisation
Does the tool reduce discretion, accelerate pace, individualise performance or weaken cooperation?
Practice
Prompting is useful, but it is not governance
A well-designed prompt can clarify the expected format, role, sources and uncertainty. It cannot guarantee factual accuracy, confidentiality, non-discrimination or legal compliance. These depend on system design, data governance, validation, training, workload and organisational rules.
Specify the task
State the audience, purpose, allowed sources, required uncertainty and forbidden assumptions.
Require evidence
Ask for references and make verification a planned part of the workflow rather than an optional final check.
Keep a stop rule
Define situations in which the user must not rely on the model and must escalate to a qualified professional.
Key references
- European Commission, European approach to artificial intelligence.
- NIST, AI Risk Management Framework.
- Full references and detailed explanations are available in the French reference edition.