“AI” does not describe a single tool.
Generating text, assigning a score and allocating tasks are three different functions, with different risks.
Understand · Essential concepts
Understand artificial intelligence at work.
Distinguish between systems, understand how a language model produces an answer, and read performance results without making them say more than they do.
12-minute readSelected sourcesUpdated August 2026
What seems stable
The Savage State · 1834
The landscape still seems open, while the mountain gives the impression of a stable landmark. The work reminds us that an apparent equilibrium can quickly become fragile when the setting changes: understanding an AI system means identifying what must remain under control — the task, the data and the decisions — before novelty hides the weak points. View the artwork ↗
An LLM produces a plausible continuation.
It calculates probabilities from context; it does not independently verify whether its answer is true or compliant.
Real work is the right level of analysis.
The model matters, but so do the task, the data, the people, the decision rules and the organisation.
01 / Name the system
First, what kind of artificial intelligence are we talking about?
The term “AI” covers systems that do not produce the same kind of output. To understand a professional use, begin by identifying what the tool actually does.
Generative AI
It creates or transforms text, images, sound or code from an instruction. An LLM-based conversational assistant belongs to this family.
Predictive AI
It estimates a probability, assigns a score or classifies a case using past data: default risk, case priority or diagnostic support.
Algorithmic management
It allocates tasks, sets processing priorities, measures activity or influences a management decision. It may or may not include generative AI.
02 / Inside an LLM
How a language model produces an answer.
A large language model — or LLM — is a neural network trained to predict what comes next in a text. It does not retrieve a ready-made answer from a database: it composes one, unit by unit.
Text becomes tokens
The prompt and supplied documents are split into small units: words, word fragments or symbols.
Tokens become numbers
Each token is converted into a numerical representation that encodes patterns learned during training.
Attention weights the context
The Transformer examines relationships between tokens to determine which parts of the context matter most at that point.
The next token is selected
The model calculates a probability distribution. Generation settings affect the selection and its variability.
The answer is built step by step
The selected token is added to the context, then the calculation repeats until the answer is complete.
03 / Capabilities & limits
What language models can do — and what still needs checking.
Capabilities depend on the model and version, the language, supplied documents, available tools and how the task is framed.
| Task type | What the model can contribute | What cannot be assumed |
|---|---|---|
| 01Read and write | Summarise, rephrase, translate, extract information, compare documents or prepare a first draft. | Completeness, fidelity to sources, currency of information and suitability for the professional context. |
| 02Reason through a problem | Break down a question, propose hypotheses, apply a described procedure and explain the steps in a calculation. | Consistent correctness. A convincing explanation can accompany a wrong answer. |
| 03Write code and process data | Produce, explain or correct code; prepare a query or exploratory analysis. | Security, robustness and correct behaviour on untested cases. Execution and testing remain essential. |
| 04Use tools | With a suitable application, search the web, consult a document repository, call software or carry out several actions in sequence. | Source quality, authority to act and control of consequences. Access rights must be limited. |
| 05Process multiple media | Depending on the model: describe an image, transcribe audio, analyse a document or answer questions about a video. | Complete perception of detail, understanding of the situation and compliance with profession-specific rules. |
A demonstrated capability is not guaranteed reliability.
Frequent success on a category of tasks does not predict every answer. Test the errors that matter in your use case.
A model version is time-specific.
A provider may change the model, its settings, filters or tools. A useful evaluation records the version and date.
04 / Read the benchmarks
There is no universal ranking of the “best model”.
A benchmark measures performance on a set of tasks under a specific prompt, metric and conditions. Change any of them and the ranking may change.
Live links instead of figures that are already out of date.
The pages below are maintained by their authors. They provide the latest available results without freezing a ranking here that would quickly become outdated.
01General capabilitiesMaintained dashboardHELM Capabilities — Stanford CRFMKnowledge, reasoning, instruction following, dialogue and mathematics, with protocols and prompts available for inspection.View latest results ↗
02Human preferencesBlind comparisonsLMArenaUsers compare two answers without knowing the model. Useful for overall preference; insufficient on its own to judge accuracy.Open the leaderboard ↗
03Open modelsEvaluation directoryLeaderboards — Hugging FaceAn entry point to general and specialised evaluations, with results published on model cards.Explore evaluations ↗
04GeneralisationNovel problemsARC-AGI-2 — ARC PrizeVisual tasks designed to test interpretation, rule composition and adaptation to new problems.View the benchmark ↗
05HealthSpecialised evaluationMedHELM — Stanford CRFMA framework for medical tasks, showing why an overall average is not enough in a professional domain.Open MedHELM ↗
05 / Return to real work
A strong technical score does not, by itself, predict a good professional use.
A benchmark describes a model’s behaviour in a test. Deployment must also be assessed in the real task, with its constraints, responsibilities and effects on people.
Model evaluation
Does the system answer correctly?
- Accuracy on representative cases
- Frequent and serious errors
- Consistency from one answer to the next
- Cost, latency and resource needs
- Resistance to misleading inputs
Deployment evaluation
Is the work done better without creating new risks?
- Time actually saved, including review
- Final work quality and handling of exceptions
- Responsibilities and ways to challenge a result
- Workload, autonomy, skills and cooperation
- Effects observed before and after implementation
The principle to remember
We do not evaluate a model alone. We evaluate the work system formed by the tool, task, data, people who use it and the organisation’s rules.
06 / Before professional use
Five questions for moving from a demonstration to an informed decision.
These questions prevent the product or its reputation from becoming the starting point. They require a concrete description of the use and the means needed to control it.
01Which precise task do we want to change?
Describe the current situation, expected outcome, people concerned and points where an error would have serious consequences. “Saving time” is not a task.
02Which data will be sent, stored or reused?
Identify personal, confidential or legally protected data; check where it is processed, how long it is retained and how the provider intends to use it.
03Who will check the output, with what time and information?
Name the competent person, include review in the workload and specify the sources against which the answer can be checked.
04Who decides and who is accountable for the consequences?
Distinguish the tool’s suggestion, the human decision and the organisation’s responsibility. Provide a way to challenge, correct or suspend the use.
05Which effects will we measure after launch?
Before the pilot, choose a small set of useful indicators: quality, errors, review time, workload, autonomy, skills, mutual support and reported difficulties.
Further reading
Choose the resource that fits your question.
A short selection from public bodies and research projects that explain their methods. Level and language are shown.
What is a language model?
A short definition that establishes the vocabulary without going into technical detail.
CNIL ↗ChatGPT explained in 5 minutes
Seven short videos on how it works, its strengths and limits, prompting and tool use.
Inria ↗Introduction to Large Language Models
Tokens, context, parameters, attention, limits and model adaptation.
Google for Developers ↗Questions and answers on generative AI
Benefits, inaccurate outputs, personal data and initial deployment precautions.
CNIL ↗AI Risk Management Framework
Framework, resources and a profile devoted to generative AI risks.
NIST ↗Evaluate models in a professional context
Connect an evaluation protocol to professional knowledge and the errors that genuinely matter.
AI & Occupational Health ↗