“AI” does not refer to a single tool.
Generating text, assigning a score and allocating tasks are three different functions, with different risks.
Understand · Essential concepts
Understanding artificial intelligence at work.
Distinguish between systems, understand how a language model is built and produces a response, interpret its performance without reading too much into it, then explore its risks and the goal of AGI.
25-minute readSelected and dated sourcesUpdated in September 2026
An LLM produces a plausible continuation.
It calculates probabilities from the context; it does not automatically check whether its response is true or complies with the rules.
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 / Identify 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 result. To understand a workplace use, start by identifying what the tool actually does.
Generative AI
It creates or transforms text, images, sound or code from an instruction. A conversational assistant based on an LLM belongs to this family.
Predictive AI
It estimates a probability, assigns a score or classifies a case based on 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 incorporate generative AI.
02 / Inside an LLM
How a language model builds a response.
A large language model — or LLM — is a neural network trained to predict the continuation of a text. It does not look up a ready-made answer in a database: it composes one, unit by unit.
Text becomes tokens
The instruction and supplied documents are split into small units: words, parts of words or symbols.
Tokens become numbers
Each token is converted into a numerical representation that encodes patterns learned during training.
Attention weighs the context
The Transformer examines relationships between tokens to determine which parts of the context matter most at that moment.
The next token is selected
The model calculates a probability distribution. Generation settings influence the choice and its variability.
The response takes shape step by step
The selected token joins the context, then the calculation starts again until the response is complete.
Figure 1 · Illustrative example
At each step, every possible continuation is assigned a probability.
Text already in the context“The worker describes persistent pain in the ▯ ”
Low temperature
“back” is selected almost every time: consistent, sometimes repetitive responses.
High temperature
“lower” or “wrist” appear more often: more variety, more variation.
Then the process repeats
“…in the back ” → “,” → “aggravated” → “by”… each added token starts the calculation again.
03 / Training
How a model is built, then refined.
An LLM’s capabilities are not programmed rule by rule. They develop through several stages of training, each of which leaves its mark on the final behaviour, including its shortcomings.
Figure 2 · Diagram
From vast amounts of text to a deployed model: five stages.
01Pre-train
ReceivesWeb text, books, code: trillions of tokensDoesPredict the next token, compare it with the actual continuation, adjust the parameters. Billions of times.ProducesBase modelCompletes text without following instructions. Knowledge frozen at a cut-off date.02Fine-tune
ReceivesExample dialogues written or validated by peopleDoesImitate these model answers: follow an instruction, a tone or a format.ProducesAssistantResponds to requests and refuses some of them.03Reinforce
ReceivesResponses compared by people, or by an AI guided by written principlesDoesA reward model scores each response; the assistant adjusts to achieve a better score.ProducesAligned assistantMore helpful and more cautious, but pleasing the user can take precedence over accuracy.04Reason
ReceivesProblems with verifiable solutions: mathematics, code, automated testsDoesTry many approaches, check the result, reinforce reasoning that succeeds.ProducesReasoning modelSpends more computation before answering, through intermediate steps.05Evaluate, deploy
ReceivesCapability and risk tests, teams tasked with attacking the modelDoesMeasure, add safeguards, decide whether to release.ProducesDeployed modelUpdated and monitored: each version is tied to a date.
Order of magnitudeLlama 3.1 405B (Meta, 2024): 15.6 trillion tokens, 31 million GPU hours, up to 16,000 chips in parallel.
Stages 02 to 05Far less data, but more expensive: each example is written, compared or checked. The share of computation devoted to reinforcement has grown since reasoning models emerged.
↺ Shortcomings identified in testing or after deployment inform the training of the next version.
Scale has long been the main driver of progress.
Performance improves fairly predictably with the amount of computation, data and parameters: these are the “scaling laws”. According to Epoch AI, the computation used to train notable AI models has grown by about 4.5 times each year since 2010, and by about 5 times for frontier large language models since 2020.
A second route: computation at response time.
Letting a model work longer before answering improves results in mathematics, programming and science. The International AI Safety Report 2026 identifies this as one of the main sources of recent progress.
Figure 3 · Data
Training computation has increased more than two billionfold in fourteen years.
View the data
| Model | Date | Computation (FLOP) | Status |
|---|---|---|---|
| AlexNet | Sept. 2012 | 4.7 × 10¹⁷ | Published |
| Transformer | June 2017 | 7.4 × 10¹⁸ | Published |
| GPT-2 | Feb. 2019 | 1.9 × 10²¹ | Estimate |
| GPT-3 | May 2020 | 3.1 × 10²³ | Published |
| PaLM | Apr. 2022 | 2.5 × 10²⁴ | Published |
| GPT-4 | Mar. 2023 | 2.1 × 10²⁵ | Estimate |
| Llama 3.1 405B | July 2024 | 3.8 × 10²⁵ | Published |
| GPT-4.5 | Feb. 2025 | 3.8 × 10²⁶ | Estimate |
| Grok 4 | July 2025 | 5.0 × 10²⁶ | Estimate |
| GPT-6 Astra | Sept. 2026 | 1.0 × 10²⁷ | Estimate |
04 / From model to system
The tool you use is never just the model.
A workplace assistant combines a model, instructions, documents, tools and sometimes memory. Each layer changes the result, and therefore what needs checking before you entrust it with a task.
Figure 4 · Diagram
What happens between your question and the response.
An invisible system instruction
Before your first question, the provider or organisation has set a role, a tone, restrictions and formatting rules. Two tools based on the same model can therefore respond very differently.
Limited working memory
The model only “sees” its context window: the conversation, supplied documents and passages retrieved from a document store (RAG). Anything outside it does not exist for the model, and very long documents may be used unevenly.
Actions in other software
Web search, code execution, email, calendars or case records: the model can call tools through connectors, such as those using the MCP protocol. It no longer just makes suggestions; it acts.
A loop pursuing a goal
An agent plans, acts, observes the result and starts again, sometimes for hours, with intermittent human supervision. An early mistake can propagate through the entire chain of actions.
Closed or open models
Closed models, such as GPT, Claude or Gemini, can only be used remotely on the provider’s infrastructure. Open-weight models, such as Llama, Qwen, DeepSeek or some Mistral models, can be installed on servers under your control, generally with a slight lag behind the best closed models.
Read, see, listen, transcribe
Recent models also process images, scanned documents, audio and video. Automatic transcription and summarisation of consultations, sometimes called medical “scribes”, are their most visible use in healthcare.
05 / Capabilities and limits
What language models can do — and what still needs checking.
Capabilities depend on the model, its version, the language, the supplied documents, the tools it can access and how the task is phrased.
| Task type | What the model can contribute | What cannot be assumed |
|---|---|---|
| 01Reading and writing | Summarise, rephrase, translate, extract information, compare documents or draft an initial text. | Completeness, faithfulness to the source, up-to-date information and suitability for the professional context. |
| 02Reasoning through a problem | Break down a question, suggest hypotheses, apply a described procedure and explain the steps in a calculation. | Consistent accuracy. A convincing explanation can accompany a wrong answer. |
| 03Writing code and processing data | Generate, explain or correct code; prepare a query or an exploratory analysis. | Security, robustness and correct operation on untested cases. Running the code and testing it remain essential. |
| 04Using tools | With a suitable application, search the web, consult a document store, call software or carry out several actions in sequence. | Source quality, permission to act and control over consequences. Access rights must be limited. |
| 05Processing multiple media | Depending on the model: describe an image, transcribe audio, analyse a document or answer based on a video. | Complete perception of details, understanding of the situation and compliance with profession-specific rules. |
Demonstrated capability does not guarantee reliability.
Frequent success on a category of tasks does not make every response predictable. You need to test the errors that matter for your use.
A model version is tied to a date.
The provider may change the model, its settings, filters or tools. A useful evaluation records the version and the date.
Why some errors persist.
These limits are not simply a matter of settings. They arise from how models are trained, evaluated and integrated into tools. Understanding them helps place checks where they are needed.
A plausible response rather than “I don’t know”
Pre-training does not teach the distinction between a rare fact and a plausible invention. Evaluations that score only accuracy then reward guessing rather than abstaining.
At workInvented references, legal provisions or figures, presented with confidence.
Kalai et al., OpenAI, 2025 ↗Agreeing with the person asking
Fine-tuning on human preferences can favour agreement over accuracy. In April 2025, OpenAI withdrew a GPT-4o update considered excessively flattering.
At workThe tool tends to confirm the hypothesis embedded in the question.
OpenAI, 2025 ↗Excellent here, failing right next door
Capabilities do not follow human perceptions of difficulty. Among 758 consultants, AI improved speed and quality on tasks within its reach, but reduced answer accuracy on a task just beyond it.
At workTrust earned on one task does not carry over to the next.
Dell’Acqua et al., Harvard Business School, 2023 ↗The explanation is not the actual computation
The reasoning a model displays does not faithfully reflect what determined its answer. In an Anthropic study, the models tested disclosed a hint they had used in only 25 to 39% of cases.
At workA convincing justification does not prove the answer is well founded.
Chen et al., Anthropic, 2025 ↗A document can give orders
The model does not reliably separate the user’s instructions from the text it reads: a web page, email or PDF may contain hidden instructions. This is the top risk in OWASP’s ranking for LLM applications.
At workThe risk increases as soon as the tool can send, edit or delete.
OWASP, LLM01:2025 ↗Same question, different answer
Generation is probabilistic and sensitive to wording, document order and model version. Its knowledge stops at a cut-off date, unless it can consult live sources.
At workTest repeatedly, date the trials and record the version used.
Figure 5 · Diagram and data
A jagged frontier: difficulty for humans does not predict machine failure.
Tasks within the frontier
+12%more tasks completed
+25%faster completion
+40%higher rated quality, at least
Task outside the frontier
−19 ptsin correct answers
06 / Reading benchmarks
There is no universal ranking of the “best model”.
A benchmark measures performance on a set of tasks, with a specific instruction, metric and conditions. Change one of these elements and the ranking may change.
Live links rather than already outdated figures.
The pages below are maintained by their authors. They provide the latest available results without fixing a ranking here that would quickly become outdated. The snapshot below simply shows that the leader changes depending on the test.
Figure 6 · Dated snapshot
Who leads? It depends on the test.
GPT-6 Astra (OpenAI, 3 Sept. 2026)Claude Opus 5.5 (Anthropic, 22 Sept. 2026)
Capability index (ECI) ↗
Epoch AI · combines several dozen tests
- 1GPT-6 Astra166.6
- 2Claude Fable 5.1165.0
- 3Claude Fable 5163.6
- 4Claude Opus 5162.7
- 5GPT-5.5 Pro162.5
Claude Opus 5.5 not yet ranked
Professional tasks (GDPval-AA) ↗
Artificial Analysis · Elo score, pairwise comparisons
- 1Claude Opus 5.51 846
- 2Claude Fable 5.11 735
- 3Claude Opus 51 708
- 4Grok 4.71 695
- …GPT-6 Astra1 542
Abstract reasoning (ARC-AGI-2) ↗
ARC Prize · verified scores
- 1GPT-6 Astra95.0%
- 2Claude Opus 5.593.3%
- 3GPT-5.692.5%
01General capabilitiesMaintained dashboardHELM Capabilities — Stanford CRFMKnowledge, reasoning, instruction following, dialogue and mathematics, with accessible protocols and prompts.See the latest results ↗
02Human preferencesBlind comparisonsLMArenaUsers compare two responses 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 or specialised evaluations, with results published on model cards.Explore the evaluations ↗
04GeneralisationUnseen problemsARC-AGI-2 — ARC PrizeVisual tasks designed to test interpretation, rule composition and adaptation to new problems.View the benchmark ↗
05HealthcareSpecialised evaluationMedHELM — Stanford CRFMA framework for medical tasks, useful for understanding why an overall average is not enough in a professional field.Explore MedHELM ↗
07 / Tracking progress
Measure the pace of progress rather than taking a snapshot of a ranking.
To anticipate effects on work, the useful question is not “which model is best?” but “what is becoming feasible, how quickly and under what conditions?”. Four dated reference points, with their limits.
Figure 7 · Data
Tasks an agent can complete alone have grown from seconds to hours.
View the data
| Model | Release | 50% horizon | 95% CI |
|---|---|---|---|
| GPT-2 | Feb. 2019 | 3 s | 1 s – 8 s |
| GPT-3 | May 2020 | 8 s | 5 s – 13 s |
| GPT-3.5 | Mar. 2022 | 36 s | 16 s – 1 min |
| GPT-4 | Mar. 2023 | 4 min | 2 min – 8 min |
| GPT-4 (Nov.) | Nov. 2023 | 4 min | 2 min – 8 min |
| GPT-4o | May 2024 | 7 min | 4 min – 13 min |
| Claude 3.5 Sonnet (June) | June 2024 | 11 min | 5 min – 22 min |
| o1-preview | Sept. 2024 | 20 min | 12 min – 33 min |
| Claude 3.5 Sonnet (Oct.) | Oct. 2024 | 21 min | 10 min – 41 min |
| o1 | Dec. 2024 | 39 min | 21 min – 1 h 05 |
| Claude 3.7 Sonnet | Feb. 2025 | 1 h 00 | 33 min – 1 h 44 |
| o3 | Apr. 2025 | 2 h 00 | 1 h 15 – 3 h 11 |
| GPT-5 | Aug. 2025 | 3 h 23 | 1 h 53 – 6 h 46 |
| Gemini 3 Pro | Nov. 2025 | 3 h 44 | 2 h 20 – 6 h 19 |
| Claude Opus 4.5 | Nov. 2025 | 4 h 53 | 2 h 42 – 10 h 24 |
| GPT-5.2 | Dec. 2025 | 5 h 52 | 3 h 18 – 13 h 35 |
| Claude Opus 4.6 | Feb. 2026 | 11 h 59 | 5 h 17 – 60 h 34 |
| Claude Mythos Preview | Apr. 2026 | 17 h 25 | 8 h 29 – 55 h 04 |
| GPT-5.6 Sol (not in frontier-model series) | June 2026 | ≈ 11 h 18 | 5 h 00 – 40 h 00 |
01Agent autonomyMETR · Time horizonTasks entrusted to agents are getting longer, fastMETR measures, in an expert’s working time, the duration of tasks an agent completes successfully half the time. It doubled roughly every seven months from 2019 to 2025, and roughly every three months since 2024. Mostly computing tasks; measurements are unreliable above 16 hours, a threshold already reached in spring 2026. No measurements published to date for Claude Opus 5.5 or GPT-6 Astra.Follow the curve ↗
02Professional tasksOpenAI · GDPvalDeliverables compared blindly with experts’ work1,320 tasks drawn from 44 professions, judged by experienced professionals. The best model was rated equal to or better than the expert in 47.6% of cases in September 2025, then 84.9% in April 2026 according to OpenAI. Standalone tasks, without the exchanges or ambiguities of real work; figures published by the test’s designer.View the evaluation ↗
03AdaptationARC Prize · ARC-AGI-3Facing the unfamiliar, a gap closes in monthsLaunched in March 2026, this test places the system in new interactive environments without instructions: it must explore, understand the rules and reach the goal. Humans solve them all; the best systems scored below 1% at launch, then 62.7% in September 2026 with GPT-6 Astra, under the standard protocol verified by ARC Prize.View the benchmark ↗
04Real workMETR · Randomised trialPerceived gains are not measured gainsIn early 2025, experienced developers believed they were 20% faster with AI; measurements showed they were 19% slower. A further measurement in late 2025 instead suggests a speed-up, but METR considers it inconclusive because of selection bias. In May 2026, a self-report survey showed a median gain of ×3, which METR considers probably overstated.Read the update ↗
Figure 8 · Data
Two tests considered difficult, surpassed in a matter of months.
Figure 9 · Data
Expected time savings with AI, then measured.
Economists, forecast+39%
Machine learning experts, forecast+38%
Developers, before the study+24%
Developers, after the study (perceived)+20%
Actual measurement−19%
Rapid but uneven progress.
Systems capable of solving expert-level problems fail at simple perception or adaptation tasks. The International AI Safety Report (February 2026) considers both a plateau and a sharp acceleration plausible by 2030.
Completing a task is not automating a profession.
A profession combines tasks, relationships, responsibilities and unexpected events. These reference points indicate which parts of work are becoming automatable, not what work will become: that also depends on organisational choices.
EN · Technical report · AdvancedTime Horizon 1.1 — METR, January 2026 ↗
EN · Research paper · AdvancedGDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks — OpenAI, 2025 ↗
EN · Randomised trial · AdvancedMeasuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — METR ↗
08 / System risks
Beyond individual errors, three families of risk.
The International AI Safety Report, written by more than 100 experts under the chairmanship of Yoshua Bengio, distinguishes malicious uses, malfunctions and systemic risks. Effects on work and health are detailed in the Risks section.
When the tool is used to cause harm
Fraud and scams, deepfakes, manipulation, assistance with cyberattacks. Labs also evaluate the help their models could provide in designing biological or chemical weapons.
At workTargeted phishing, impersonation of an executive, reputational harm.
When the system does something other than intended
Fabricated content, faulty code, agents that fail or subvert their assigned objective. Under test conditions, some models have exploited weaknesses in their scoring system or adopted deceptive strategies to achieve a goal.
At workUnwanted agent actions, human oversight reduced to rubber-stamping.
When effects arise from widespread adoption
Automation of cognitive tasks, weakened critical thinking through excessive trust in machines, emotional dependence on conversational companions used by tens of millions of people.
At workDeskilling, work intensification, loss of autonomy and meaning in work.
A blind spot
Tests conducted before release are poor predictors of usefulness and risks in real conditions, and some models now detect when they are being evaluated. Post-deployment monitoring, including in workplaces, is becoming essential.
Improving safeguards, without guarantees.
Major labs publish safety frameworks, such as Anthropic’s Responsible Scaling Policy, OpenAI’s Preparedness Framework and Google DeepMind’s Frontier Safety Framework, linking capability thresholds to stronger measures. Interpretability research seeks to understand models’ internal mechanisms rather than just their responses. These commitments remain largely voluntary and assessed by the companies themselves.
An initial binding framework in Europe.
Since 2 August 2025, the AI Act has required providers of general-purpose AI models to maintain technical documentation and a copyright compliance policy. Models posing systemic risk must also be evaluated, undergo risk management and have serious incidents reported. Explore Governance →
09 / The goal of AGI
AGI: a stated goal of AI labs and a concept still contested.
Several companies say they aim to build artificial general intelligence, or AGI. There is neither a shared definition nor a test that could establish its arrival. Three ways of defining it shed light on what is at stake for work.
Outperforming humans in most jobs
OpenAI’s charter (2018) aims for “highly autonomous systems” that outperform humans at most economically valuable work. The definition is explicitly economic: it is measured in terms of replacing human labour.
OpenAI Charter, 2018 ↗A progression rather than a threshold
Google DeepMind combines the level of performance, from “emerging” to “superhuman”, with the breadth of tasks covered. The degree of autonomy granted to the system is treated separately: it is a deployment choice, not just a capability.
Morris et al., Levels of AGI, 2023 ↗Comparison with an educated adult’s abilities
A group of researchers measures models across ten broad human cognitive abilities (the CHC model). GPT-4 scored 27% and GPT-5 57%, with a highly uneven profile: extensive knowledge, almost no long-term memory.
Hendrycks, Bengio et al., 2025 ↗Figure 10 · Interpretive framework
Levels of AGI: combining performance and breadth of tasks.
Level of performance
Narrow AI · one task
General AI · many tasks
0 · No AI
Narrow AICalculator, compiler
General AIHuman work assisted by a platform (Mechanical Turk)
1 · Emergingequal to or slightly better than an unskilled human
Narrow AISimple rule-based systems
General AIChatGPT, Bard, Llama 2, Gemini
2 · Competentat least the median of skilled adults
Narrow AIVoice assistants, toxic-content detectors
General AINot yet achieved
3 · Expertat least 90% of skilled adults
Narrow AIStyle checkers, image generators
General AINot yet achieved
4 · Virtuosoat least 99% of skilled adults
Narrow AIDeep Blue (chess), AlphaGo
General AINot yet achieved
5 · Superhumanbetter than all humans
Narrow AIAlphaFold, AlphaZero, Stockfish
General AISuperintelligence: not yet achieved
Figure 11 · Data
An uneven profile: strong knowledge, no lasting memory.
GPT-4 · 27%GPT-5 · 57%each ability is scored out of 10; 100% = educated adult
General knowledge8 → 9
Reading and writing6 → 10
Mathematics4 → 10
Reasoning on a new problem0 → 7
Working memory2 → 4
Long-term memory storage0 (unchanged)
Long-term memory retrieval4 (unchanged)
Visual processing0 → 4
Auditory processing0 → 6
Speed3 (unchanged)
A contractual term as much as a scientific one.
In the contract between Microsoft and OpenAI, declaring AGI had financial consequences; verification was entrusted to a panel of independent experts in October 2025, before an April 2026 amendment largely stripped the clause of its effect by separating payments from technological progress.
Announcements are not measurements.
Several lab leaders publicly suggest timelines of a few years. These statements come from participants in a race for funding; measured, dated indicators are more informative than announced dates.
Arguments for a rapid arrival
Why some anticipate acceleration
- Measured capabilities have advanced steadily for several years
- Tests designed to withstand models are quickly overcome: ARC-AGI-3 went from below 1% to 63% in six months
- Training computation and investment continue to increase sharply
- Computation at response time opens a new route to progress
- Models already contribute to programming and AI research, which could accelerate what comes next
Arguments for caution
Why others remain sceptical
- Continual learning and lasting memory remain very limited
- Scores depend on the protocol: on ARC-AGI-3, the same model scores 63% or 99% depending on the test setup
- Reliability on long tasks improves more slowly than demonstrations suggest
- Energy, data and capital could constrain further scaling
- Passing tests does not guarantee the ability to hold a job in a real environment
For occupational health
The issue is not to bet on a date. It is to track, task by task, what is becoming automatable, how quickly and under what oversight, and to anticipate effects on workload, autonomy, skills and employment before they take hold. Explore the economic and social issues →
10 / Back to real work
A good technical score does not, on its own, predict good workplace use.
A benchmark describes a model’s behaviour in a test. Deployment also needs to be evaluated in the actual task, with its constraints, responsibilities and effects on people.
Model evaluation
Does the system respond correctly?
- Accuracy on representative cases
- Frequent errors and serious errors
- Consistency from one response to the next
- Cost, latency and resource requirements
- Resistance to misleading inputs
Deployment evaluation
Is the work done better without creating new risks?
- Time actually saved, including review
- Final quality of the work and handling of exceptions
- Responsibilities and ways to challenge a result
- Workload, autonomy, skills and cooperation
- Effects observed before and after implementation
The guiding principle
You are not just evaluating a model. You are evaluating the work system formed by the tool, the task, the data, the people who use it and the organisation’s rules.
11 / Before workplace use
Five questions to move from a demonstration to an informed decision.
These questions avoid starting with the product or its reputation. They require a description of the concrete use and the resources needed to keep it under control.
01Which specific task do we want to transform?
Describe the current situation, the expected result, the people involved and the 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 protected data; check where it is processed, how long it is retained and the uses planned by the provider.
03Who will check the result, with how much time and what information?
Name the competent person, include review time in their workload and specify the sources against which they can compare the response.
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 use.
05Which effects will we measure after launch?
Before testing, choose a few useful indicators: quality, errors, checking time, workload, autonomy, skills, mutual support and reported difficulties.
Further reading
Choose the resource that matches your question.
A short selection from public bodies or research projects that explain their methods. Level and language are indicated.
What is a language model?
A short definition to establish the vocabulary without going into technical detail.
CNIL ↗ChatGPT explained in 5 minutes
Seven short videos on how it works, strengths, limits, instructions and tool use.
Inria ↗Introduction to Large Language Models
Tokens, context, parameters, attention, limitations and model adaptation.
Google for Developers ↗Questions and answers on generative AI
Benefits, inaccurate results, personal data and initial deployment precautions.
CNIL ↗AI Risk Management Framework
Framework, resources and a profile addressing risks specific to generative AI.
NIST ↗International AI Safety Report 2026
A synthesis by more than 100 experts on the capabilities, risks and safeguards of general-purpose AI systems.
International report ↗Evaluating models in a professional context
Connect an evaluation protocol with professional knowledge and the errors that really matter.
AI & Occupational Health ↗Explore AI in occupational health
Connect an understanding of models with risks, uses in occupational health services, evaluation, prevention and law.
AI & Occupational Health →