The AI Glossary for Marketers (Beginner Edition)

This glossary explains 50 artificial intelligence terms for marketing professionals who are just getting started. Every term follows the same pattern: a one-sentence definition, a picture drawn from your daily work, and what it means for you.

Where to start: the 5 essential terms

If you only have ten minutes, read these five entries first.

• Artificial intelligence: the umbrella word that covers everything else.
• Machine learning: the method that lets a tool learn from your data.
• Generative AI vs. predictive AI: two families of tools with different uses and different risks.
• Prompt: the instruction that determines the quality of what the AI gives back.
• Hallucination: the most common error, and the reason every piece of content must be reviewed by a person.

1. The basics: what do we mean by AI?

What is artificial intelligence (AI)?

In one sentence: artificial intelligence is the set of technologies that let a computer perform tasks once reserved for humans, such as understanding a text, recognizing an image or recommending a product.

The picture: imagine a statistician intern. They understand nothing about your business, but they have studied millions of examples and spot what looks like something they have already seen.

Why it matters to you: the same word covers the "you may also like" box on a shop website and the assistant that drafts your newsletter. They are very different tools. When a vendor sells you "AI", ask what it does exactly: the word alone guarantees neither quality nor novelty.

What is an algorithm?

In one sentence: an algorithm is a sequence of precise instructions that a computer follows to solve a problem, like a recipe applied to the letter.

The picture: a recipe gives the same dish every time, as long as the ingredients are the same. The computer invents nothing; it executes.

Why it matters to you: the rule "email customers who have been inactive for 90 days" is an algorithm with no AI at all. "The Instagram algorithm" simply means the system that decides which posts you see. Not every algorithm is AI: telling a fixed rule (predictable, checkable) from learning (powerful, less predictable) helps you judge a tool and talk accurately with your team.

What is machine learning?

In one sentence: machine learning is a method where the computer discovers rules by itself by analyzing many examples, instead of receiving them all written by a programmer.

The picture: it is a young salesperson who, after a thousand meetings, senses which prospects will sign, without always being able to explain why.

Why it matters to you: it has powered digital marketing for more than ten years: recommendations on streaming platforms, the score given to a prospect in a CRM, the detection of a suspicious payment. It only learns from your data: poor or outdated data gives poor results.

What is deep learning?

In one sentence: deep learning is a form of machine learning that stacks many layers of computation, called neural networks, to process complex data such as text, images or sound.

The picture: think of an assembly line. The first stations spot simple features; the last ones recognize a face or understand a sentence.

Why it matters to you: it is what explains the recent leap in generative AI tools, from producing a product visual out of a description to transcribing a customer call. The word "neuron" is a metaphor: it is not an artificial brain.

What are weak AI and strong AI?

In one sentence: weak AI is a system designed for one specific task, while strong AI would be a general intelligence comparable to a human's, which does not exist today.

The picture: weak AI is a very sharp specialist, like a spell checker. Strong AI would be a colleague able to do anything, which is still science fiction.

Why it matters to you: every tool you will use is weak AI. An assistant that writes ad hooks knows neither what your brand is nor why a customer buys: it imitates texts that look like hooks. This helps you put alarmist talk in perspective and be wary of vendors who suggest their tool "understands" your customers.

2. How it works: the technical words

What is an LLM (large language model)?

In one sentence: an LLM is a program trained on huge amounts of text that produces language by predicting, word after word, the most likely continuation.

The picture: it is your phone's autocomplete, but trained on an entire library.

Why it matters to you: ChatGPT, Claude, Gemini or Mistral rely on LLMs to draft an email, summarize a report or answer a customer. An LLM does not look for the truth, it produces plausible text. That explains its fluency and its mistakes (see "hallucination"). Always treat its output as a draft to review.

What is a hallucination?

In one sentence: a hallucination is false or invented information that a generative AI presents with the same confidence as an accurate fact.

The picture: it is an intern who, instead of admitting they do not know, makes up a convincing answer.

Why it matters to you: figures, quotes and proper names must be checked before publication. Your brand answers for what its chatbot says: in 2024, a Canadian tribunal held Air Canada responsible for a refund rule invented by its bot and ordered it to pay CAD 650.88 (CTV News). Not to be confused with bias, which reflects a distortion present in the data.

What is a prompt (and prompt engineering)?

In one sentence: a prompt is the written instruction you give a generative AI, and prompt engineering is the skill of wording it to get a good result.

The picture: it is a brief. Vague brief, vague result.

Why it matters to you: "Write a LinkedIn post" returns generic text. "You are the marketing manager of an organic cosmetics company: write a 600-character post announcing our new serum, warm tone, ending with a question" returns something usable. Give context, role, goal and expected format, as you would with a contractor.

What is a token?

In one sentence: a token is the small piece of text (a word, part of a word or a punctuation mark) that the AI reads and writes, and it serves as the unit for length and billing.

The picture: it is like the character count an agency bills for a translation: the AI counts in tokens.

Why it matters to you: analyzing a 100-page report uses far more tokens than summarizing three paragraphs, so it costs more in a pay-per-use tool. A word generally takes a little more than one token, and more in some languages. Knowing the unit helps you budget and avoid needlessly long prompts.

What is a foundation model?

In one sentence: a foundation model is a very large AI model, trained on massive and varied data, that is then adapted to many uses.

The picture: it is an engine that several carmakers fit into different cars.

Why it matters to you: many marketing tools that advertise "our AI" call a major provider's model behind the scenes rather than a homegrown technology. Ask the vendor which model it uses: your tool depends on that provider for price, availability and where your data is processed.

What is fine-tuning?

In one sentence: fine-tuning means continuing the training of an existing model on your own examples to specialize it for a task or a style.

The picture: it is training a generalist in your business by having them work for months on your files.

Why it matters to you: it is a heavy, often costly option. A brand that wants a model to write "like us" can fine-tune it on its best emails, but for a simple brand tone a good prompt with examples, or RAG, is often enough. Ask a vendor why fine-tuning would be necessary.

What is RAG (retrieval-augmented generation)?

In one sentence: RAG is a technique that lets the AI consult your documents at the moment it answers, instead of relying only on its training memory.

The picture: it is an open-book exam: the AI finds the right page before answering.

Why it matters to you: an internal assistant can answer your sales team from your up-to-date catalog, pricing and product sheets. It is the main way to reduce hallucinations and refresh answers, but it does not eliminate them: the AI can misread a document, so verify.

What is a multimodal model?

In one sentence: a multimodal model understands and produces several types of content (text, image, sound, video) in the same conversation.

The picture: it is a colleague you can show a photo, dictate an instruction to, and ask for text or a visual in return.

Why it matters to you: you can send the photo of a package and ask for three descriptions adapted to Instagram, your website and a marketplace. It cuts the number of tools you juggle, but quality varies from one format to another: test text, image and sound separately.

What is a context window?

In one sentence: the context window is the maximum amount of text an AI can keep "in mind" at once, your instructions and documents included.

The picture: it is the size of the AI's desk: the bigger it is, the more files stay open together; beyond that, sheets fall off.

Why it matters to you: if you ask for a summary of a 200-page market study and it exceeds the window, the AI may overlook part of it, or you have to split it. An incomplete answer or ignored instructions often come from there, and a large window does not guarantee equally careful reading of everything.

What is the difference between generative AI and predictive AI?

In one sentence: predictive AI anticipates an outcome from past data (who will buy, who will leave), while generative AI creates new content (text, image, sound).

The picture: the first is a weather forecaster, the second a writer.

Why it matters to you: the CRM that scores your prospects is predictive; the tool that writes the email sent to them is generative; many tools combine both. The checks differ: a prediction is verified by comparing with real results, generated content is reviewed piece by piece. Knowing which one you are dealing with avoids organizing the wrong checks.

What is inference?

In one sentence: inference is the moment when an already-trained model produces an answer to your request, as opposed to training, done upstream by the vendor.

The picture: training is the years of study; inference is each billed consultation.

Why it matters to you: every time your team asks a tool for a text, a visual or an analysis, an inference is triggered, with a computing cost. It explains pay-per-use pricing and quotas: the heavier the use, the higher the bill, a useful point in negotiation.

What is training data?

In one sentence: training data is the text, images and other content a model learned from, and it determines what the model "knows" and what it ignores.

The picture: these are the books the intern read before joining. If they only read old ones, they will not know your new competitors.

Why it matters to you: a model can miss a product launch that happened after its training, or reproduce stereotypes found in its sources. Training data is the source of bias and outdated information: ask vendors about its origin and freshness. Not to be confused with documents supplied through RAG.

3. AI in your daily marketing work

What is generative personalization?

In one sentence: generative personalization means using AI to create a unique message or visual for each person, instead of choosing among a few versions prepared for segments.

The picture: it is going from off-the-rack to made-to-measure: a salesperson who writes a different letter to each customer.

Why it matters to you: scale changes, and so do the risks: brand consistency and the use of personal data (see "GDPR and generative AI"). Review a sample before any mass send.

What is predictive analytics?

In one sentence: predictive analytics uses historical data to estimate the likelihood of future behavior: a purchase, a cancellation, a response to an offer.

The picture: it is the weather forecast. It announces a 70% chance of rain; it does not guarantee that it will rain.

Why it matters to you: it feeds targeting, scoring and recommendations, for instance to spot customers likely to leave and send them a win-back offer in time. The trap is treating a probability as a fact: regularly check that predictions come true.

What is synthetic content?

In one sentence: synthetic content is text, image, voice or video produced wholly or partly by an AI.

The picture: it is an industrial photomontage: the scene was never filmed, yet it looks real.

Why it matters to you: since August 2, 2026, the EU AI Act requires certain generated content to be labeled (see "AI Act"). Beyond the law, audiences can react badly: record internally how much human work went in.

What is a conversational chatbot (virtual assistant)?

In one sentence: a conversational chatbot is a program that talks with your customers in natural language and answers their questions without following a fixed script.

The picture: it is a shop assistant available around the clock who has read your whole catalog, but who may improvise when they do not know.

Why it matters to you: it relieves customer service, but it can make things up outside the intended scope (see "hallucination"). Provide a handover to a human and state clearly that it is an AI.

What is AI lead scoring?

In one sentence: AI lead scoring automatically gives each prospect a score estimating their likelihood of becoming a customer, based on their behavior and characteristics.

The picture: it is a seasoned salesperson sorting the mail and putting the hottest prospects' letters on top of the pile.

Why it matters to you: CRMs such as Salesforce or HubSpot offer predictive scoring that studies your past prospects who became customers in order to score new ones. A score is only as reliable as your data, and the history must be large enough. The model can also favor some profiles: check that high-scored prospects really convert.

What is GEO (Generative Engine Optimization)?

In one sentence: GEO is the set of practices aimed at getting your brand cited or recommended in the answers of AI assistants such as ChatGPT, Perplexity, Gemini or Google's AI overviews.

The picture: it is SEO for a new shelf. The goal is no longer to rank first in a list of links but to be quoted in the answer itself.

Why it matters to you: a buyer asks an assistant "which email tool for a small business?". If the sources the AI consults describe your product clearly, you have a chance of being named. The same idea also goes by AEO or LLMO (Ahrefs). Practices are young and nothing guarantees a citation: aim for clear, sourced and consistent content, alongside SEO.

What is synthetic data (market research)?

In one sentence: synthetic data is fictional but statistically plausible data generated by an AI to simulate consumer behavior.

The picture: it is a flight simulator. It lets you practice, but it does not replace the real sky.

Why it matters to you: it speeds up idea exploration at lower cost, but it also reproduces the biases of the model that generates it. Ipsos, which has partnered with Stanford University to test synthetic respondents (Ipsos), recommends pairing it with human rigor and real data before any investment decision.

What is a marketing copilot?

In one sentence: a marketing copilot is an AI assistant built directly into a tool you already use, such as your CRM, your email tool or your campaign manager.

The picture: it is an assistant sitting next to you in your usual software, proposing and preparing while you decide.

Why it matters to you: it is the easiest entry point, with no new tool to learn. HubSpot and Salesforce, for example, build such assistants into their suites. Its quality depends on the tool's data: ask what the copilot can access and where your data is processed.

4. Rules, ethics and risks

What is the AI Act (the EU regulation on AI)?

In one sentence: the AI Act is the European Union regulation that governs AI systems according to their level of risk, with transparency and human oversight requirements.

The picture: it is a highway code for AI: the more dangerous the situation, the stricter the rules.

Why it matters to you: since August 2, 2026, a chatbot must disclose that it is an AI unless this is obvious, and content generated or heavily modified by AI must be labeled (Blog du Modérateur). A short grace period runs until December 2, 2026 for the technical marking of systems already on the market (Banque des Territoires). Obligations for high-risk systems were postponed, not transparency, and staff AI training has been required since February 2025. Timelines keep moving, so check them regularly.

What is explainability?

In one sentence: explainability is the ability to understand and explain why an AI produced a given result, instead of treating it as a black box.

The picture: it is the difference between an adviser who justifies a recommendation and an oracle who answers "that's how it is".

Why it matters to you: some scoring tools display the factors behind a prospect's score, which lets a salesperson see why that prospect rises to the top. Explainability becomes essential when a decision affects a customer (an offer refused, an unfavorable score) and for your teams to trust the tool. The most powerful models are often the least explainable.

What is algorithmic bias?

In one sentence: algorithmic bias is a systematic distortion in an AI's results that reproduces the imbalances present in its training data.

The picture: it is a recruiter who has only seen résumés of one profile and concludes that only those people succeed.

Why it matters to you: a scoring model trained on your past customers can underrate profiles you never targeted, simply because they are absent from your history. You may exclude potential customers, with an ethical and regulatory stake whenever an AI influences an individual decision. Compare results by segment. Not to be confused with hallucination.

What is responsible AI?

In one sentence: responsible AI is an approach to designing and using AI that takes its ethical, social and environmental effects into account, not just its performance.

The picture: it is a quality or CSR program applied to AI.

Why it matters to you: a marketing team can adopt a short charter: what the AI may do alone, what is always reviewed, which data is banned from tools, how customers are informed. Customers, partners and regulators expect clear commitments. It is a voluntary approach that goes beyond legal compliance and brings together bias, explainability and data governance.

What is data sovereignty (sovereign AI)?

In one sentence: data sovereignty is an organization's ability to keep control over where its data is stored and processed, and over the law that applies to it.

The picture: it is the difference between keeping the keys to your own safe and renting a safe from a third party subject to other laws.

Why it matters to you: before entrusting customer data to a tool, ask where it is processed, by whom and under what law. Some vendors, such as Mistral AI, offer models processed in Europe (Siècle Digital). Do not mix this question up with the choice of a European vendor.

What does GDPR say about generative AI?

In one sentence: the GDPR applies to generative AI as soon as personal data is entered into the tool or produced by it, as with any other data processing.

The picture: pasting a customer file into an online tool is like handing it to a contractor: it needs a framework.

Why it matters to you: an employee who pastes a customer list into a public chatbot to draft follow-ups takes that data out of your control. Do not enter personal data into a non-approved tool and check each vendor's processing terms before any rollout. France's data protection authority (CNIL) advises starting from a precise need before deploying a tool (CNIL).

What is watermarking (marking and traceability)?

In one sentence: watermarking is the set of techniques that make it possible to flag and trace that a piece of content was generated or modified by an AI.

The picture: it is a banknote's watermark or a product's origin label.

Why it matters to you: it answers the AI Act's transparency obligations, but it has limits: a modification with a non-compatible tool can make verification fail (Google Cloud), and several systems still coexist. The most widespread standard, C2PA ("Content Credentials"), has been built into Adobe's tools since 2024. Keep your original files.

What is data governance?

In one sentence: data governance is the set of rules, roles and processes that govern how data is collected, quality-checked, accessed and used in an organization.

The picture: it is a library's house rules: who may borrow what, who shelves, who checks.

Why it matters to you: before plugging an AI copilot into your CRM, remove duplicates, define who accesses which records and keep proof of consent. AI amplifies the quality of your data, good or bad: poorly governed data produces mediocre or even risky results. It is a prerequisite much broader than AI, which AI has made urgent.

What is shadow AI?

In one sentence: shadow AI is the use of AI tools by employees outside any framework approved by the company, often with personal accounts.

The picture: it is an employee installing their own software on their computer without the IT department knowing.

Why it matters to you: data leaks and GDPR non-compliance are at stake. Engineers at Samsung entered confidential code into ChatGPT three times within weeks in 2023, after which the company banned such tools (Siècle Digital). According to a 2024 INRIA-Datacraft study, nearly 7 in 10 French employees use generative AI at work without telling their managers (Into The Minds). A pure ban pushes people to hide their use: better to provide approved tools and set which data is allowed.

Copyright and generative AI: what should you know?

In one sentence: copyright raises two questions here: the use of protected works to train models, and whether the content produced by AI is protected.

The picture: it is a very fast copyist: the copy may resemble the original, and it does not necessarily belong to you.

Why it matters to you: if an agency bills you for a transfer of rights on a visual that is 100% AI, its value is uncertain (Consultation.avocat.fr). In France, a work entirely generated by AI, without sufficient input from a human author, is not protected by copyright (National Assembly). The law is moving: the European Parliament is calling for prior authorization from rights holders for model training.

What is algorithmic transparency (and consent)?

In one sentence: it is the obligation to inform people when their data feeds an automated system, or when they are talking to an AI rather than a human.

The picture: it is the ingredients label on a food product: it informs you before you consume.

Why it matters to you: your website chatbot should open with "I'm a virtual assistant". This disclosure has been mandatory since August 2, 2026, unless the context makes it obvious. Discovering afterwards that you were talking to an AI harms trust: say so upfront, in plain words, and explain how to reach a human.

5. Organizing AI in your team

What is human-in-the-loop?

In one sentence: human-in-the-loop is a principle where a person steps in at a defined point of the process to validate, correct or stop the AI before the final action.

The picture: it is the proofreader at a print shop who signs off the proof before printing.

Why it matters to you: the more the AI acts alone, the more this check matters. Say an AI drafts your follow-up emails and a manager approves a sample before the mass send. It needs a real stopping point before the action, not a glance afterwards. Decide in advance who validates what.

What is orchestration (AI tools and workflows)?

In one sentence: orchestration means coordinating several AI tools or steps within one process, to automate a task from end to end.

The picture: it is a production line where each station hands over to the next.

Why it matters to you: a chain can analyze the week's sales, propose three campaign angles, write the emails, schedule them and measure the results. The time savings are significant, but an error at one step spreads to the next ones. Plan human checkpoints and start with a simple process.

What is AI literacy?

In one sentence: AI literacy is the level of understanding that lets a person or a team use AI wisely, know its limits and judge its results.

The picture: it is knowing how to read a map before setting off. You do not need to know how to make it.

Why it matters to you: a sophisticated tool without a trained team creates little value. The AI Act has required organizations to train their staff since February 2025 (Blog du Modérateur), and companies that pair AI with structured upskilling are twice as likely to get strong results (DataCamp).

What is an augmented marketer?

In one sentence: an augmented marketer is a professional who relies on AI to be more efficient while keeping control of judgment, strategy and creativity.

The picture: it is a pilot who sets and monitors the autopilot and takes the controls back when needed.

Why it matters to you: rather than wondering whether AI will replace you, list the tasks to delegate (first drafts, sorting, summaries) and those where your judgment makes the difference (strategy, brand tone, trade-offs). The AI can prepare three versions of a brief; you choose, correct and decide.

What is cross-functional governance?

In one sentence: cross-functional governance is a way of steering AI projects that involves marketing, data, IT and legal from the start, instead of letting each team move alone.

The picture: it is a building site where the architect, the electrician and the plumber coordinate. Otherwise you drill the walls after the painting.

Why it matters to you: before deploying a chatbot, marketing defines the need, the data team prepares the sources, IT handles security and legal reviews the notices. Many AI projects fail for organizational reasons (silos) rather than technical ones. Marketing has to co-build with other teams rather than command.

What is AI ROI (measurable value)?

In one sentence: AI ROI measures the value an AI project actually creates compared with its cost, in money, time or quality.

The picture: it is the balance sheet of a training course: what it cost, what it brought in.

Why it matters to you: in a 2026 survey of more than 500 US and UK executives, 21% report a significant positive return and 42% a moderate one (DataCamp). If AI cuts a newsletter from three hours to one, count the hours saved, the tool's cost and the review time. Set your indicators before launching, without forgetting speed or quality gains, which are harder to quantify.

6. What's coming: emerging trends

What is agentic AI?

In one sentence: agentic AI refers to systems able to chain several steps on their own to reach a goal, without asking for validation at each step.

The picture: it is moving from an assistant you ask for a text to a colleague you hand a mission to.

Why it matters to you: you ask "prepare and launch the back-to-school campaign" and the agent analyzes the audience, writes, schedules and adjusts. Salesforce and HubSpot already offer agents of this kind, with varying degrees of autonomy. The term is often oversold. In July 2026, France's data protection authority (CNIL) and the AI and Digital Council warned of a risk of losing control over personal data, and recommend limiting agents' access to data and applications and monitoring their actions (Le Monde Informatique).

What is an AI agent (and a multi-agent system)?

In one sentence: an AI agent is a program that acts autonomously to reach a goal, using tools such as email or the CRM; a multi-agent system has several specialized agents work together.

The picture: it is a colleague you give access to your mailbox and calendar. A multi-agent system is a small team: one searches, one writes, one proofreads.

Why it matters to you: unlike a chatbot, it acts: sending an email, changing a campaign. A prospecting agent, for example, finds companies matching your target, drafts a personalized email and prepares it in your CRM. The more autonomous it is, the more you should limit its permissions and keep a human for irreversible actions (see "human-in-the-loop").

What is an "AI-native" company or tool?

In one sentence: an AI-native tool is designed from the start around AI, rather than having an AI feature added afterwards.

The picture: it is a house whose plans include home automation, compared with an older house to which a connected thermostat was added.

Why it matters to you: the claim is increasingly used in sales pitches. Compare a content tool whose whole workflow (brief, writing, measurement) relies on AI with a classic email tool that just added a "write with AI" button. To judge it, ask what the AI really does in the tool, what data it uses and what would remain if you removed it.

What is an SLM (small language model)?

In one sentence: an SLM is a language model more compact than a large LLM, faster and cheaper, but less versatile.

The picture: it is a neighborhood specialist rather than a big hospital: fewer resources, but effective in its field.

Why it matters to you: a small model that sorts customer emails (complaint, quote request, after-sales) can be enough where a large model would be overkill. On repetitive, well-defined tasks, an SLM can cost less, answer faster and run on your own servers, which helps confidentiality (Cognizant). Ask a vendor why a large model would be indispensable.

What is the difference between open-source and proprietary models?

In one sentence: an open-source model makes its workings available to be used and adapted, while a proprietary model stays closed and is only used through its vendor's interface or API.

The picture: it is a published recipe you can cook at home, versus a dish ordered in a restaurant whose recipe is secret.

Why it matters to you: Mistral AI, for example, offers open models that can be run with data processing located in Europe (Mistral AI). The choice affects data sovereignty, cost and dependence on a single vendor. "Open source" does not always mean free, nor fully transparent: the degree of openness varies widely.

What is a world model?

In one sentence: a world model is an AI model that seeks to represent and simulate how the physical world works, where an LLM mainly predicts the next piece of text.

The picture: it is a digital mock-up of a store where you could move shelves to see how customers walk through.

Why it matters to you: Google DeepMind has presented Genie 3, which generates navigable 3D environments in real time, and the CNIL's innovation lab describes this field as a paradigm shift compared with language models (LINC-CNIL). Even the definition is debated and maturity for marketing is low: watch it for customer-experience simulation or prototyping, but do not budget for it yet.

What is an AI avatar (synthetic human)?

In one sentence: an AI avatar is a digital person generated by AI, who appears on video, speaks and embodies a brand or a contact without really existing.

The picture: it is a talking mascot, but so realistic it can be mistaken for a real presenter.

Why it matters to you: platforms such as Synthesia turn a script into a video with a virtual presenter, in more than 160 languages according to the vendor (Synthesia), mostly for training and corporate communication. The transparency question arises immediately: does the audience know it is not talking to a real person? Not to be confused with a deepfake, which reuses a real person's likeness without consent.

What is continual learning?

In one sentence: continual learning is an AI's ability to take in new information over time without forgetting what it already knew, something current models do poorly.

The picture: it is a colleague learning a new piece of software without ever forgetting the old one, whereas a naively updated AI often forgets part of the past.

Why it matters to you: a chatbot launched two years ago does not know your new offers unless you provide them. Researchers call "catastrophic forgetting" the loss of old knowledge during a new update (RTBF). An AI tool does not improve on its own with your use: plan to update and test it regularly. Not to be confused with RAG, which supplies recent information without changing the model.

Sources

The figures and rules in this glossary, and the research used to prepare it, come from pages found through web search in November 2026. Most pages are in French. Regulatory passages and figures should be rechecked at the source before publication.

Regulation and authorities

• Blog du Modérateur: AI Act, what changes on August 2, 2026
• Banque des Territoires: AI-generated content, transparency obligations
• CNIL: how to deploy generative AI
• CNIL and CIANum: note on agentic AI (reported by Le Monde Informatique)
• LINC-CNIL: world models
• French National Assembly: ministerial answer on AI-generated works
• Consultation.avocat.fr: agencies, advertisers and copyright on AI creations
• Google Cloud: Content Credentials and C2PA

Cases, studies and surveys

• CTV News: Air Canada and its chatbot
• Siècle Digital: Samsung and ChatGPT
• Into The Minds: shadow AI, INRIA-Datacraft study
• DataCamp: AI ROI in 2026, YouGov survey
• Ipsos and Stanford: synthetic data
• RTBF: catastrophic forgetting and continual learning

Tools and vendors (information supplied by the vendors)

• Mistral AI: regional inference and open models
• Siècle Digital: Microsoft and Mistral AI
• Ahrefs: GEO
• Synthesia: AI avatars
• Cognizant: small language models