The ICIA Référentiel

#LLM#ChatGPT#generative AI

How does a language model like ChatGPT work?

A language model predicts the most likely next word based on billions of text examples. It does not consult a knowledge base: it generates a plausible response from statistical reg…

What is generative AI and how is it different from traditional AI?

Generative AI produces new content — text, images, sound, code — while traditional AI classifies, predicts or recommends from existing data. Both rely on machine learning, but gene…

#generative AI

#hallucinations#verification

Why does AI get things wrong: understanding hallucinations

A hallucination is a false answer delivered with confidence. It stems from how language models work — they generate what is plausible, not what is verified. No current model is imm…

What is an algorithm and where do you encounter one daily?

An algorithm is a sequence of instructions that transforms data into a result. You encounter dozens daily: news feed, GPS, ticket prices, applicant sorting, university admissions. …

#algorithm

How does AI learn: data, training, models

An AI system learns by adjusting billions of parameters to minimize errors on training data. The quality and representativeness of this data directly determine the quality and bias…

#training#data

What is a prompt and how to formulate requests effectively

A prompt is the instruction you give to a generative AI. Its quality determines the quality of the response. Effective principles are documented and stable: provide context, specif…

#prompt

#understanding#anthropomorphism

Does AI really understand what it says?

The question divides the scientific community itself. What is established: current models manipulate statistical representations of language without experience of the world, intent…

#machine learning#deep learning

What are the differences between AI, machine learning and deep learning?

AI is the general field, machine learning is the family of methods where machines learn from data, deep learning is the sub-family using deep neural networks, behind the spectacula…

What is an AI agent and how is it different from a chatbot?

A chatbot answers, an agent acts: it chains actions, uses tools, navigates, executes multi-step tasks with a goal. This growing autonomy shifts questions of responsibility and cont…

#AI agents

#bias#discrimination#AI Act

AI bias: where does it come from and how to spot it

AI bias primarily comes from training data, which reflects real-world imbalances, and from design choices. It is documented in hiring, lending, facial recognition. The AI Act impos…

How much does AI consume: energy, water, environmental footprint

Training and using large models consumes significant and growing amounts of electricity and cooling water, documented notably by the International Energy Agency. Orders of magnitud…

#environment#energy

#players#sovereignty#open source

Who builds AI: landscape of players

Large model development is concentrated among a few American and Chinese players, with European players including France's Mistral AI. Alongside proprietary models, an active open-…

Open-source AI: what it is and why it matters

An open-source model publishes its weights and allows reuse, enabling auditing, research and independence from major vendors. The AI Act provides lighter obligations for some open …

#open source#AI Act

#disinformation#deepfake#critical thinking

How to verify information in the age of AI

Verification relies on stable reflexes: trace back to the primary source, cross-check independent sources, check the date, be wary of content that triggers strong emotion. Syntheti…

How to use AI to learn more effectively

AI is effective for learning when it serves as a tutor that questions, reformulates and adapts — not as a dispenser of answers. Uses that work, backed by cognitive science research…

#learning#tutor

#personal data#GDPR#privacy

How to use an AI assistant without exposing your personal data

Basic rule: never submit to a consumer AI any data you would not publish. Check whether your conversations are used for training and disable that option when available, never enter…

#verification#hallucinations

Can you trust AI answers? A verification method

No by default, yes under conditions. Trust should be calibrated by stakes: brainstorming and reformatting tolerate error, health, law and money do not. Four-step method: assess the…

How to use AI for writing without losing your voice

AI writes fast and smooths everything out. To keep your voice: write the first draft or outline yourself, use AI for review and contradiction rather than production, systematically…

#writing#transparency

#deepfake#detection#AI Act

How to spot AI-generated content

No method is 100% reliable, and automated detectors produce documented false positives. Useful clues: factual inconsistencies, visually aberrant details, lack of primary sources, r…

#administration#public services

How to use AI for administrative procedures

AI can help understand a letter, prepare a file, rephrase a request. It does not replace official sources: any administrative information must be verified on service-public.fr or w…

Parents & teachersEveryday use

AI and health: what AI can and cannot do for you

A consumer AI can explain a medical term or prepare questions for a consultation. It does not provide a reliable diagnosis and never replaces a healthcare professional. AI-based me…

#health

How to protect children in an AI world

Three fronts: synthetic content, AI companions designed to maximize engagement, and the data children expose. Effective levers are regular dialogue, service settings, knowledge of …

#minors#protection

How seniors can embrace AI

No age barrier exists for conversational AI: talking to an AI is precisely the most accessible interface ever created. High-value uses for seniors: understanding a document, gettin…

#seniors#inclusion

Which free AI tools to start with

All major general-purpose assistants offer a free tier sufficient for discovery and practice. This article provides a neutral choice framework: intended use, language, privacy, lim…

#tools#free

How to talk to an AI: basics of effective dialogue

A conversational AI does not guess your context: it only knows what you tell it. Four effective reflexes: provide context, make a precise request, react to the response instead of …

#prompt#dialogue

#professions#responsibility

Can AI replace a teacher, a doctor, a lawyer?

For standard information tasks, AI already competes. For diagnosis involving liability, personalized advice and human relationships, no — and the law sets boundaries: practicing me…

What does the AI Act, the European regulation on artificial intelligence, say?

The AI Act, EU Regulation 2024/1689, is the world's first comprehensive legal framework on AI. In force since August 1, 2024, it classifies AI systems by risk level: prohibited pra…

#AI Act

Workers & professionalsRights & regulation

Which AI practices are banned in Europe since February 2025?

Since February 2, 2025, the AI Act notably bans: generalized social scoring, manipulation exploiting vulnerabilities, emotion recognition at work and in education, biometric catego…

#AI Act#bans

What is the AI literacy obligation under Article 4 of the AI Act?

Since February 2, 2025, Article 4 of the AI Act requires providers and deployers of AI systems to ensure sufficient AI literacy among their staff. This obligation concerns any orga…

#AI Act#AI literacy#training

GDPR and AI: what rights do you have over your data with AI systems?

The GDPR, EU Regulation 2016/679, fully applies to AI whenever personal data is processed, both in training and in use. Your rights: access, rectification, erasure, objection, rest…

#GDPR#CNIL#personal data

#Article 22#GDPR#automated decision

Can you challenge a decision made by an algorithm?

Yes, in specific cases. Article 22 of the GDPR generally prohibits fully automated decisions producing significant legal effects, with exceptions, and guarantees a right to human i…

Who is responsible when AI causes harm?

AI has no legal personality: responsibility always traces back to people — provider, deployer or user — depending on the case. The framework combines general liability law, the rev…

#liability

#copyright#intellectual property

AI and copyright: who owns what AI creates?

Under French and European law, copyright-protectable works require original human creation: content generated solely by AI is not, in principle, protected by copyright. Upstream, u…

Are your conversations with AI private?

Not by default. Depending on the service and settings, your conversations may be stored, reviewed for moderation and used for training. The GDPR gives you rights of access and eras…

#privacy#personal data

#facial recognition#biometrics#CNIL

Facial recognition: what the law allows and prohibits in France

Real-time facial recognition in public spaces for law enforcement purposes is banned by the AI Act except for strict exceptions subject to authorization. In France, the May 19, 202…

What is the CNIL doing about AI?

The CNIL is the French data protection authority and a central player in AI regulation: recommendations on GDPR compliance for AI systems, a dedicated AI service created in 2023, i…

#CNIL

Workers & professionalsRights & regulation

Deepfakes: what does French law say?

The SREN law of May 21, 2024 adapted the penal code: publishing AI-generated visual or audio content depicting a person without their consent is punishable if it is not manifestly …

#deepfake#penal code

AI at work: what your employer can and cannot do

An employer may deploy AI tools, but within a framework: prior information of employees for any system collecting their data (Article L1222-4 of the Labour Code), consultation of t…

#Labour Code#works council#surveillance

How to exercise your GDPR rights with an AI service

The procedure is the same as for any digital service: identify the data controller in the privacy policy, submit your request (access, erasure, objection), get a response within on…

#GDPR#CNIL#remedies

The AI Act timeline: which obligations by which dates

The AI Act applies in phases: February 2, 2025 for bans and the AI literacy obligation, August 2, 2025 for rules on general-purpose AI models and governance, August 2, 2026 for mos…

#AI Act#timeline

What is a high-risk AI system?

The AI Act classifies as high-risk those AI systems used in domains where errors harm fundamental rights: hiring, education, credit, justice, essential public services, biometrics,…

#AI Act#high risk

Will AI destroy your job? What studies actually say

Studies converge: AI massively transforms tasks, much more than it eliminates entire jobs in the short term. The OECD, IMF and France Stratégie estimate that a large share of jobs …

#employment#studies

General publicWork & skills

Which AI skills will be in demand in the coming years?

Three levels emerge from public frameworks like the European DigComp: AI literacy for using tools in one's own profession (accessible to all); hybrid skills — oversight, verificati…

#skills#DigComp

How to train in AI for free

Quality free resources exist: public platforms, open university courses, CNIL and Pix resources for digital skills, and the MentivisOS ICIA platform, free for life, which builds a …

#training#free

#Labour Code#AI Act#training

Is your employer required to train you in AI?

Two foundations combine: the general obligation to adapt employees to their position (Article L6321-1 of the Labour Code) and, since February 2025, the AI literacy obligation of Ar…

How to use your personal training account (CPF) for AI training

The personal training account, topped up with €500 per year for a full-time employee up to €5,000, funds training leading to nationally registered qualifications. A co-payment of a…

#CPF#funding

How to include AI in your CV and job applications

Mentioning AI on a CV is only valuable if concrete: tools actually used, use cases in your profession, results. The article also covers the other side: recruiters use automated scr…

#recruitment#CV

Career change into AI: where to start

AI careers are not limited to coding: data, quality, compliance, training, project management and expert use also hire. The realistic starting point: map your transferable skills, …

#career change

Workers & professionalsRights & regulation

AI and management: what changes for supervisors

The manager becomes the first level of AI governance: define authorized uses within the team, ensure human verification of outputs, upskill everyone without widening gaps, and carr…

#management

Algorithmic surveillance of employees: framework and limits

Algorithmic monitoring of employees is strictly regulated: proportionality and prior information required by the Labour Code and GDPR, works council consultation, and since Februar…

#surveillance#CNIL#works council

#works council#social dialogue

What employee representatives need to know about AI

The works council must be consulted on the introduction of new technologies and any significant changes to working conditions (Article L2312-8 of the Labour Code). Faced with an AI…

Freelancers and independents: using AI without legal risk

Three vigilance points for independents: confidentiality of client data submitted to AI tools (may breach contractual commitments), ownership and originality of AI-generated delive…

#freelance#contracts

Jobs created or transformed by AI

AI creates direct jobs (engineering, data, compliance) and mostly transforms existing jobs by shifting time toward supervision, relationship and expertise. The article draws on DAR…

#jobs#prospective

#national education#usage framework

Can AI be used in schools: the framework in France

The Ministry of Education has published a framework for AI use in education that authorizes supervised pedagogical uses and sets limits: protection of student data, no autonomous u…

#pedagogy#personalization

How AI is transforming learning

Research in education sciences identifies real potential — personalization, immediate feedback, on-demand tutoring — and documented risks — cognitive offloading, loss of retrieval …

AI and homework: helping children without cheating

The dividing line is simple to state: AI can explain, help review, provide additional examples; it should not produce the homework. The article provides concrete uses by age, signs…

#homework#families

How teachers can use AI

Immediate value uses for teachers: differentiating materials, generating exercises and variations, preparing lesson sequences, providing formative feedback. Conditions: never submi…

#teachers#practices

#critical thinking#media literacy

Critical thinking in the age of AI: what needs to be taught

Critical thinking toward AI is taught as a method, not a posture: understanding how content is generated, learning to trace back to sources, practicing error detection on real case…

#students#academic integrity

Parcoursup and admissions algorithms: understanding to avoid being subject to them

Parcoursup does not assign candidates via a central algorithm: the platform transmits applications and rankings are established by individual programs, which may use decision-suppo…

#Parcoursup#admissions

General publicEducation

Students: what you are allowed to do with AI

There is no single national rule: each institution sets its own policy, from prohibition to supervised use with declaration. The stable framework: plagiarism and exam fraud remain …

How good is AI-generated educational content?

AI-generated educational content is worth what its verification is worth: generation produces structure and clarity, but factual errors and approximations are documented across all…

#content#quality

General publicEducation

Pix and digital skills: where does AI fit in?

Pix is the public online service for assessing and certifying digital skills, aligned with the European DigComp framework, which now integrates AI-related competencies. The article…

#Pix#DigComp

General publicEducation

Learning AI with AI: method and limits

Using AI to learn AI works on one condition: that the system makes you practice, question and correct rather than deliver content. This is the principle of adaptive learning, docum…

#adaptive learning

#democracy#DSA#disinformation

AI and democracy: elections, manipulation, information

Documented risks: synthetic content during election periods, mass targeting, pollution of the information space. Existing protections: the EU Digital Services Act requires large pl…

#public services#administration

AI in French public services

The French administration uses algorithmic processing governed by the code of relations between the public and the administration: mandatory explicit mention, right to obtain the p…

#inclusion#digital illiteracy

Digital divide and AI: who risks being left behind

INSEE documents that a significant share of the population still struggles with basic digital tasks, with digital illiteracy particularly affecting older and less educated individu…

Local authorities & stakeholdersSociety & environment

Why AI acculturation is a civic issue

Understanding AI now conditions the concrete exercise of rights: challenging an algorithmic decision, protecting one’s data, evaluating information, training, working. The article …

#citizenship#AI literacy

How a local authority can foster AI acculturation in its territory

A territorial AI acculturation approach follows four proven steps: diagnosis of local needs and digital maturity, awareness-raising for elected officials and staff, prioritized act…

#territories#local authorities

AI and ecological transition: real help or false promise?

Both statements are true and must be held together: AI has its own environmental footprint, growing and documented, and it contributes to measurable optimizations — energy grids, a…

#environment#climate

Local authorities & stakeholdersEveryday use

Women and AI: usage gaps and representation issues

Available data shows gender gaps in generative AI tool usage, in training pathways and in technical careers, with direct consequences on system design and biases. The article prese…

#gender#inclusion

AI and nonprofits: where to start with no budget

A nonprofit can save significant time with free AI tools on three low-risk uses: writing and adapting content, preparing grant applications, summarizing documents. Prerequisites: n…

#nonprofits#associations

Digital sovereignty: why Europe wants its own AI

European AI dependency is factual: models, chips and cloud infrastructure are mostly non-European. The public response combines regulation (AI Act), investment, European programs a…

#sovereignty#Europe