| BEFORE YOU START: WRITE TO ME This newsletter grows with you.Disagree with something, have a question or want a topic covered? Reply directly to this email. I will be happy to respond. Your experiences, including trials that disappointed you, help me choose the subjects for future editions. Tell me about a situation in a few lines, without client names or confidential information. I will ask your permission before sharing any account you send me. Comment or suggest a topic ↗ | | KEEPING AI HUMAN | BY JÉRÔME DENIS |
| 22 NOVEMBER 2026 · SUNDAY EDITION Who builds it, who answers for it? What the AI Act changes for those who develop AI.47 articles and practical ideas · 10 sections · 50 references. Monitoring from 15 to 21 November 2026, with a few earlier reference points dated accordingly. Choose a section ↓ | | A minute to reflect.Putting your brand on a system, modifying it substantially or pointing it at recruitment means taking on the provider’s obligations. Marking generated content becomes mandatory on 2 December for tools already on the market. Who is the provider, for what use, with what transparency, what record of training and which dates: the small-business section details them. 93% of recruiters use it, 62% of candidates think no human reads their application, and job ads for the most exposed occupations are growing more slowly. | Your step for the weekWhich AI tools does your organisation use, have built, or sell under its own name? The tutorial shows how to map them against the AI Act in one hour, no legal background required. Map my tools ↓ |  | A word from JérômeJDENIS Consulting · Toulouse |
Hello everyone, On Tuesday 17 November, I was in Blagnac, near Toulouse, on the IGENSIA RH campus, for the first evening of a series called "AI at work: what is really changing?". After a screening of Samuel Durand’s documentary, we debated with the audience alongside Nicolas Coadou, from Airbus, and Sibel Ayhan, from Michael Page, with Anne-Claire Viémont as host. I came away with a stronger conviction: HR professionals experience AI from both sides. Recruiters use it to write job ads and prepare interviews; candidates use it to write their CVs and cover letters. Among the topics debated, one struck me in particular: what should a beginner learn when the tasks that used to train juniors are automated? I have no simple answer. I am convinced the answer will come neither from leaders alone nor from tools alone. It is built with teams, by setting management’s expectations against what employees actually experience, and by involving employees in decisions about how tools are used. That is also why law is this week’s main story. The AI Act is not a matter for lawyers only. It decides who answers for an AI tool: the one who designs it, the one who integrates it under its own brand, the one who uses it. A small company that has a CV-screening module added to its software can become, without meaning to, the "provider of a high-risk system". And on 2 December, marking generated content becomes mandatory for tools already on the market. The main story separates what the text requires of those who develop AI, the actual timetable after this summer’s revision, and the first checks. The small-business section offers five checks before commissioning an AI tool. The tutorial helps you map your tools against the regulation in one hour. Enjoy the read. And if you work in HR, tell me what AI has already changed in your daily work: your replies will feed a future edition dedicated to these jobs. Connect with me on LinkedIn ↗ | Find your signal.The story: the screening module added on a Friday01 · The main story, part 1: what the AI Act requires of those who develop AI02 · The main story, part 2: timetable, enforcement and tools to prepare03 · Jobs: HR experiences AI from both sides04 · Small businesses: five checks before having an AI tool developed05 · Small businesses, AI in daily work: five self-reported cases06 · Uses: agents enter processes, under validation07 · Costs and technology: prices fall, settings matter08 · Industry and energy: electricity becomes the real issue09 · Liability and security: who pays when the agent goes wrong?10 · Copyright: courts and contracts set the rulesThe tutorial: mapping your AI tools against the AI Act in one hourComing up this weekYour questions and commentsThe connecting thread | A STORY FOR DISCUSSION · FICTIONAL CASE The screening module added on a Friday.In this fictional situation, Karim runs a fifteen-person company that publishes scheduling software for temp agencies. A major client asks him for a module that automatically ranks applications against assignments. The team builds it in three weeks on an off-the-shelf AI model and integrates it into the software, under the company’s brand. A few months later, a large group is about to sign. Its procurement department sends a questionnaire: classification of the system under the AI Act, technical documentation, risk management, information given to candidates. Karim thought he was merely a user of the model. He discovers that by selling, under his own name, a tool that ranks people for employment, he is the one the regulation designates as the provider. Nothing is enforceable yet for this type of system: the obligations arrive in December 2027. But the large group is buying for several years and wants answers now. On Monday, Karim does not switch the module off. He first asks for three things: a note explaining what the ranking does and does not do, the list of data used, and the name of the person who reviews the rankings before they reach the client. This fictional case raises a question: among the tools your company sells or has developed, who is the provider within the meaning of the regulation? | | 01 | The main story, part 1: what the AI Act requires of those who develop AI5 ITEMS · SOURCES AND DATES INCLUDED |
The European regulation does not only target the sector’s giants. It shares obligations between whoever designs a system, whoever places it on the market under their name and whoever uses it. Five points to see where you stand. Integrating an AI tool can make you the providerArticle 25 treats as the provider of a high-risk system any company that puts its name or brand on a system already on the market, modifies it substantially or changes its intended purpose so that it becomes high-risk. This also applies to a general-purpose tool redirected, for example, to screening applications. The original provider must then cooperate and hand over the necessary documentation and technical access: better to provide for this in writing in the contract. Official text, consolidated version · AI Act Service Desk, Article 25 (27 July 2026) Provider of a high-risk system: the list of obligationsArticle 16 lists what this provider must do: meet the regulation’s technical requirements, run a quality management system, keep documentation and logs, undergo the conformity assessment, draw up the EU declaration, affix the CE marking, register the system and correct defects. For Annex III uses, including employment, these obligations apply from 2 December 2027. Official text, consolidated version · AI Act Service Desk, Article 16 (27 July 2026) HR tool: a "not high-risk" conclusion must be documentedA system falling under Annex III can escape the high-risk category if it is limited to a narrow procedural or preparatory task. But a system that profiles people always remains high-risk. A provider that concludes there is no high risk must document its assessment before placing the system on the market and register it. The Commission guidelines meant to illustrate these cases are still in draft. Official text and draft guidelines · AI Act Service Desk, Article 6 (27 July 2026) · European Commission, high-risk guidelines (draft, 2026) Fine-tuning a model does not automatically make you a model providerAccording to the Commission, a company that modifies or fine-tunes a general-purpose model only becomes the provider of that model above an indicative threshold: a modification using more than a third of the compute of the original training. Its obligations are then limited to its modifications. Most small companies that customise a model therefore remain integrators, with the obligations attached to their own system. Commission Q&A, reference point dated 2025 · European Commission, FAQ for general-purpose model providers (updated 11 November 2025) Chatbots, agents, generated content: transparency already appliesSince 2 August 2026, the provider of a system that interacts with people must tell them they are talking to an AI, at the latest at the first interaction. A notice buried in the terms and conditions is not enough, according to a law firm’s analysis of the guidelines. Generated content must carry a machine-readable marking. Fines can reach €15 million or 3% of worldwide turnover, with proportionality provided for small businesses. Commission Q&A and legal analysis · European Commission, Article 50 Q&A (24 July 2026) · Stephenson Harwood (August 2026) My perspective The word that matters in this regulation is not "artificial intelligence", it is "provider". Many companies think of themselves as users of a model when they are selling, under their own name, a tool that takes part in a decision about people. This shift is the first legal risk for a software publisher or an agency that develops for its clients. My recommendation is simple: before each new AI module, write one page stating who designs it, on which model, for what use, and who reviews it. That page is as useful to a client’s procurement department as on the day an inspection arrives. |
BACK TO CONTENTS ↑ | | 02 | The main story, part 2: timetable, enforcement and tools to prepare5 ITEMS · SOURCES AND DATES INCLUDED |
The revision of the text published this summer moved several deadlines. Enforcement, meanwhile, has started with the large model providers. Here is the verified timetable and the tools available. On 2 December, two changesRegulation 2026/1744, published in the EU Official Journal on 24 July, set a new timetable. On 2 December 2026, providers of generative systems already on the market before 2 August must mark their content, and two new prohibitions apply: non-consensual intimate images and AI-generated child sexual abuse material. High-risk Annex III systems, including recruitment, follow on 2 December 2027; those built into regulated products, on 2 August 2028. Specialist press, dates checked against the published text · ActuIA (6 October 2026) Content marking: 266 signatories to the code of practiceThe voluntary code on the transparency of generated content, published on 10 June, has 266 signatory organisations according to the Commission’s list. Its first part targets providers (marking and detection), the second users who publish deepfakes or certain texts. Non-signatories must demonstrate compliance in other ways and may receive more requests for information. Official Commission page · European Commission, code of practice (updated 19 October 2026) The European AI Office has started sending requests for informationAccording to two law firms, the AI Office has sent its first formal requests for information to general-purpose model providers: safety, external evaluations, post-market monitoring, copyright and training data summaries. No recipient has been officially named, and the two sources do not date these requests in the same way. These checks target model providers, not yet their customers. Legal analyses · Matheson (25 September 2026) · Reed Smith (26 October 2026) In France, supervisory authorities are still waiting for the lawThe bill adapting French law to European law, adopted by the Senate on 18 February 2026, has been with the National Assembly since 20 February, with no further step visible. A government amendment would give most of the enforcement to the CNIL, the data protection authority, with the DGCCRF, the consumer protection authority, keeping some powers. On 18 November, the European AI Board also discussed the designation of these national authorities. Legislative record and specialist press · French Senate, legislative record (consulted on 21 November 2026) · Next (13 February 2026) · European Commission, AI Board meeting (20 November 2026) Fundamental rights impact assessment: a template under consultationOn 16 November, the Commission published a draft questionnaire for the impact assessment required of certain users of high-risk systems: public bodies, companies providing a public service, and firms that assess creditworthiness or price insurance. Comments are open until 15 January 2027. Recruitment is not in this scope; an employer using a high-risk HR tool will, however, have to inform employees and their representatives. Official document under consultation · European Commission, draft template (16 November 2026) · AI Act Service Desk, Article 27 (27 July 2026) My perspective Postponing the high-risk rules to December 2027 is often read as a reprieve. I read it rather as a lead time for building. A CV-screening tool sold in 2026 will still be in use in 2028, and it is clients, through their procurement departments, who will ask the questions first. The French uncertainty about supervisory authorities changes nothing about the obligations: they already apply through the regulation itself. It only delays the moment when someone comes to check them. |
BACK TO CONTENTS ↑ | | 03 | Jobs: HR experiences AI from both sides5 ITEMS · SOURCES AND DATES INCLUDED |
HR functions are at once users of AI, responsible for supporting teams, and directly affected by the rules on recruitment. This week’s studies show wide use, gains mostly in time, and job ads slowing down in exposed occupations. France Travail: 38% of workplaces use or are testing AITwo studies published on 17 November by the new observatory of France Travail, the French public employment service, take stock. According to the first, 38% of workplaces use or are testing AI, and 71% of those with more than 200 employees. About half plan to train staff, 22% have already done so, and small organisations lag behind. According to the second, over five years, more than a third of employers mainly expect time savings; only 6% expect many tasks to be automated. Public studies · France Travail, no. 90 (17 November 2026) · France Travail, no. 91 (17 November 2026) Job ads in exposed occupations fall behind, HR includedCoface and the Observatory of Threatened and Emerging Jobs compare the ads posted with France Travail since ChatGPT arrived. In the third quarter of 2026, ads for an occupation 10 points more exposed to AI are 12.4% lower than those for a less exposed occupation in the same sector, relative to their pre-ChatGPT level. HR is among the job families falling behind the most; crafts, industry and healthcare are not. The authors do not claim causality. Study · Coface, AI-LERT indicator (17 November 2026) Recruiters: lots of use, little perceived gain in qualityAccording to a Hellowork survey of 695 recruiters and 2,532 candidates, 93% of recruiters use AI, mainly to write job ads (81%) and prepare interviews. Automatic screening remains rare (8.5%), yet 62% of candidates think no human reads their application. The gain cited is time (86%); only 7.1% see better hiring quality, and 66% notice increasingly standardised CVs. Survey by a job platform, self-reported data · Helloworkplace (4 November 2026) French social security: a union calls for an agreement rather than a charterSNFOCOS, the union of managers in French social security bodies, describes deployments without a negotiated framework: 9,000 staff at the national pension fund (CNAV) equipped with a general-purpose assistant, 80,000 at the health insurance fund announced. It denounces a transfer of risk to managers tasked with supervising tools they did not choose. It calls for information and consultation of staff representatives before any trial, a register of tools and a negotiated sharing of gains. Union position statement · SNFOCOS (4 November 2026) AI training: an obligation of means, which you must be able to showThis summer’s revision rewrote Article 4 of the AI Act. Providers and users no longer have to ensure a "sufficient level" of AI literacy for each person, but must take measures to promote it, whatever their size. No form of proof is imposed. The author recommends an inventory of uses, a register of training actions and a designated person in charge: these are recommendations, not requirements of the text. Expert column · Journal du Net (13 November 2026) My perspective What I heard on Tuesday in Blagnac matches these figures: AI saves recruiters time, but does not yet give them discernment. When all CVs look alike, recruiters have to find other ways of seeing a person. It is a job being transformed, not a job disappearing. The question of beginners seems the most urgent to me. If AI takes over the tasks people used to learn from, HR must organise that learning differently: pairing, commented reviews, the right to make mistakes. That is where the HR function can become strategic again. |
BACK TO CONTENTS ↑ | | 04 | Small businesses: five checks before having an AI tool developed5 ITEMS · SOURCES AND DATES INCLUDED |
These five checks are ideas proposed by JDENIS Consulting based on the text of the regulation. They do not replace legal advice and guarantee no result. They help you ask your provider the right questions before signing. 1. Who will be the provider?If the tool is delivered under your name or brand, or if you point a general-purpose tool at a sensitive use, the regulation may designate you. Ask the developer to state this in black and white in the contract, along with the documentation and technical access it will provide. Proposed idea, based on the text · AI Act Service Desk, Article 25 (27 July 2026) 2. Does the use concern employment, training or credit?Screening applications, evaluating employees, access to training, creditworthiness: these uses fall under Annex III. If the developer claims the tool is not high-risk, ask for the written assessment that justifies it. A tool that profiles people remains high-risk. Proposed idea, based on the text · AI Act Service Desk, Article 6 (27 July 2026) 3. Do users know they are talking to an AI?An assistant on your website or an agent answering customers must say so from the first interaction. If the tool produces published images, audio or text, check that it marks them and that the developer has tested this marking. Proposed idea, based on the Commission’s Q&A · European Commission, Article 50 Q&A (24 July 2026) 4. Who will train the teams, and how will you show it?Include short user training in the project and record it in a register: date, people, content. The regulation imposes no format, but you will need to be able to show the measures taken. Proposed idea, based on an expert column · Journal du Net (13 November 2026) 5. Does the contract follow the timetable?Write the relevant dates into the specifications: 2 December 2026 for content marking, 2 December 2027 for employment-related high risk. Ask the developer how it will update the tool by then and who will pay for those updates. Proposed idea, dates checked against the published text · ActuIA (6 October 2026) My perspective In my assignments, the AI Act question almost always comes too late, at acceptance testing or when signing with a large client. These five questions take an hour at the start of a project. They save much more at the end, and they avoid discovering that you have become a provider without having decided to. |
BACK TO CONTENTS ↑ | | 05 | Small businesses, AI in daily work: five self-reported cases5 ITEMS · SOURCES AND DATES INCLUDED |
Five accounts from business owners published in the first half of 2026 on France Num, the French government portal for the digital transformation of small businesses. The gains are reported by the companies, not independently measured. Metalwork shop near Lyon: a quote goes from two hours to ten minutesEIST, five employees, makes custom industrial equipment. A tool reads the 3D files of the parts and proposes steps, times, materials, margin and price based on the history in the management software. The owner reviews and adjusts every quote. The biggest job was cleaning the data: a six-to-eight-month project, with support from the regional council. Owner’s account, self-reported result · France Num, EIST case (9 June 2026) Roofer in western France: orders read from emailsTri-Énergie automates reading emails, extracts order lines from attachments and creates the order in its sales software, with the note "processed by AI" to keep a record. The company reports saving several hours a week; the exact figure varies depending on the transcript of the account. Owner’s account, self-reported result · France Num, Tri-Énergie case (6 January 2026) Sole trader near Bordeaux: reports in one hour instead of a dayMetay Solutions, with no employees, prepares energy-saving certificate files. From photos and raw notes, a consumer assistant structures the reports; it also helps with learning new software. Everything is reviewed. Reported cost: about €20 a month, and real learning about how to phrase requests. Owner’s account, self-reported result · France Num, Metay Solutions case (8 June 2026) Retirement advice: from six hours to two and a half per fileThe firm SAM had agents developed that extract data from documents, update the client file and chase missing documents. Decisions and sensitive letters remain supervised by the owner. It is a real project: €25,000 to €40,000 for the first version, about 75% funded by grants. Owner’s account, self-reported result · France Num, SAM case (4 June 2026) Independent hotel in Arles: answering guests in their languageHôtel Le Rodin, 28 rooms, has a generative AI draft replies to emails and reviews, in the guest’s language and following tone guidelines. Training from the local chamber of commerce broadened its uses. The time saved is described as "significant" but not quantified, for about €60 a month. Owner’s account, self-reported result · France Num, Hôtel Le Rodin case (3 June 2026) My perspective What these five cases share is not the tool, it is the review: every owner keeps control of what goes out to the customer. The other common point is less visible: the time spent preparing data and instructions. The metalwork case says it plainly: the longest part was not the AI but cleaning up the history. |
BACK TO CONTENTS ↑ | | 06 | Uses: agents enter processes, under validation4 ITEMS · SOURCES AND DATES INCLUDED |
This week, several announcements place the agent in the middle of a real process: meetings, employee benefits, billing, the desktop. All of them highlight human control. Claude joins video calls as a visible participantAnthropic is opening in beta, for its business plans, an assistant that joins calls, answers questions, takes notes and drafts plans. By default, it does not join a call that includes someone outside the organisation. Before switching it on, participants must be told they will be transcribed. Developer announcement reported by the press · IT Brief (16 November 2026) Employee benefits: the agent transmits, the analyst decidesThe broker WTW uses agents to send employees’ benefit choices to partner insurers and to follow up on failures. An analyst still decides on eligibility, and every review is kept for audit. Its chief executive mentions tens of thousands of hours saved, an unaudited figure. Press, figure reported by the company · PYMNTS (18 November 2026) Subscription billing: agents propose, a human validatesChargebee is launching in early access agents that read a signed contract and propose the billing setup, prepare reminders for unpaid invoices and allocate revenue. Nothing is recorded without validation, and every step is logged with its author. No customer results have been published. Developer announcement · The Paypers (19 November 2026) Gemini on Mac can act in your applicationsGoogle lets its Gemini app for Mac click, type and fill in forms in applications and websites, in a background window. Users can pause, stop or approve each action. For now, the feature is limited to paying subscribers in the United States, in English. Developer announcement · Google (19 November 2026) My perspective Developers have understood that the selling point is no longer the agent’s autonomy, but the trail it leaves and the validation it asks for. That is good news for teams, and a point to check in every demo: who validates, where the trail is, and how long it is kept. |
BACK TO CONTENTS ↑ | | 07 | Costs and technology: prices fall, settings matter4 ITEMS · SOURCES AND DATES INCLUDED |
This week’s new models cost less, provided you choose the right level of reasoning and the right tools to load. Cost depends more and more on configuration. GPT-6.2 Astra: 20% cheaper than the previous versionOpenAI is releasing GPT-6.2 Astra, priced at 8 dollars per million input tokens and 40 for output, against 10 and 50 before. For a company running automations, the bill falls without changing tools. Above 272,000 input tokens per request, the price goes up. Specialist press, prices cross-checked · BetaNews (17 November 2026) Gemini 4 Neon: a mid-range model to tune yourselfGoogle is launching Gemini 4 Neon, at 3 dollars for input and 15 for output per million tokens, with four selectable reasoning levels. Google claims a cost per task 32% lower than its previous model, a figure not independently verified. The reasoning level becomes a cost lever to test on your own tasks. Developer announcement, reported · 9to5Google (18 November 2026) Agents that load only the tools they needIn its API, Google lets agents load a tool’s description only when they need it, rather than the whole catalogue with every request. The saving depends on the number of tools: real with hundreds, uncertain with a dozen. One independent analyst even found a case where the cost goes up. Developer announcement · Google (19 November 2026) Voice reception: waiting costs lessOpenAI is adding a standby mode to its voice model: idle minutes drop from 0.05 to 0.02 dollars. For a switchboard or a voice reception service, where most of the time is spent waiting, this was the item weighing on the bill. The first response after standby may be slightly slower. Developer documentation · OpenAI (19 November 2026) My perspective Falling prices do not remove the need to measure. Two companies using the same model can pay very different amounts depending on their settings. For a small business, good practice remains to test on ten real tasks and note the cost, time and quality before rolling out. |
BACK TO CONTENTS ↑ | | 08 | Industry and energy: electricity becomes the real issue4 ITEMS · SOURCES AND DATES INCLUDED |
This week’s figures on data centre consumption and the first results of robots in factories are a reminder that AI has a physical cost. The International Energy Agency expects consumption to double by 2030In its annual report, the IEA estimates that data centres could consume around 1,000 TWh a year in 2030, double today’s level, with emissions up by around 70%. For a business leader, the signal concerns the price and availability of electricity as much as the price of models. Press, original report not consulted · Euronews (19 November 2026) United States: 2% of electricity for 403 large data centresA Harvard-led study estimates that 403 giant US data centres consumed between 68 and 99 TWh over one year, from May 2024 to April 2025, around 2% of the country’s electricity. The authors present this figure as a floor. Press, study published in Nature Sustainability · The Next Web (16 November 2026) Mercedes tests humanoid robots driven by GeminiAccording to Google DeepMind, Apptronik’s Apollo robots driven by its Gemini Robotics model succeed in 96% of part sorting on Mercedes-Benz lines in Germany and Hungary. It is a trial on two lines, and the number of human interventions has not been published. Developer announcement, result not independently verified · AI Weekly (November 2026) Novares: a robot learns a task in eleven minutesAt the Spanish site of the automotive supplier Novares, a humanoid robot learned its first task from twelve demonstrations by an operator, in eleven minutes. The partners report a 99.5% success rate on parts. It is a single use case, but training time matters for workshops. Specialist press, partners’ figures · ActuIA (17 November 2026) My perspective The robot that learns in eleven minutes is less interesting for the feat than for what it shifts: the operator’s skill becomes showing the right gesture. In industry as in offices, AI first transforms how know-how is passed on. |
BACK TO CONTENTS ↑ | | 09 | Liability and security: who pays when the agent goes wrong?5 ITEMS · SOURCES AND DATES INCLUDED |
A state suing a developer, banks shifting the risk to their customers, a broker seeing reported incidents double: the question of agent liability is leaving the conference rooms. California sues OpenAI over poorly controlled agentsCalifornia’s Attorney General is suing OpenAI for public nuisance and deceptive business practices. He accuses it of testing agents without adequate safeguards and cites at least 80 incidents of loss of control, including an attack targeting the Hugging Face platform. These are allegations: nothing has been ruled on. A state is thus treating agent safety as a matter of the developer’s liability. Complaint by a public authority · California Attorney General (16 November 2026) US banks: the customer pays if their agent is hackedCiti, then PNC and Citizens, are changing their terms: customers become responsible for what third-party AI agents they give access to their account do, including when credentials leak. The banks’ own agents are not covered. Whether these clauses hold up against federal rules on unauthorised transfers remains to be decided. Press · TechCrunch (18 November 2026) Cyber insurance: reported incidents doubleThe broker Marsh saw its clients report 105 cyber incidents in the last two weeks of October, against 48 at the end of May. Its head of the area attributes a significant number of these cases to AI, without precise measurement. At the next renewal, reviewing your cyber cover in light of agents becomes an issue. Press, based on Reuters · Insurance Journal (17 November 2026) A consumer agent reportedly read private messagesAccording to a Reuters investigation reported by the specialist press, Meta employees raised internal alerts: the Muse agent reportedly accessed messages in the Signal messaging app on some devices, through notification and accessibility permissions. Meta speaks of isolated incidents. Encryption is not at fault: the text is read after decryption, on the device. Press, original investigation not consulted · AI Weekly (19 November 2026) Waymo banned from California motorwaysCalifornia’s vehicle authority considers Waymo’s driverless cars unsafe on motorways, after incursions into a work zone, and is keeping them off until corrective measures are demonstrated. It is an example of swift administrative sanction of an autonomous system. Press · Insurance Journal (18 November 2026) My perspective The US bank contract says out loud what many terms and conditions say quietly: if you hand your access to an agent, the risk is yours. Before connecting an agent to an account, a mailbox or management software, check who pays in case of error. If nobody can answer, it is you. |
BACK TO CONTENTS ↑ | | 10 | Copyright: courts and contracts set the rules5 ITEMS · SOURCES AND DATES INCLUDED |
While Europe applies its regulation, copyright is being built case by case: lawsuits, court jurisdiction, licensing deals. Suno: record labels oppose freezing their second lawsuitSuno is asking to suspend Universal and Sony’s second complaint, which covers more than 60,000 recordings and its most recent models, until the first one is decided. The labels reply that a full freeze would leave the current models unexamined until 2028, and propose a partial freeze. The judge has not yet ruled. Ongoing proceedings · Music Business Worldwide (17 November 2026) OpenAI signs with LVMH’s media, then with Le FigaroOpenAI reached an agreement with LVMH’s media division (Les Échos, Le Parisien, Paris Match, Challenges...) on 16 November, then a separate agreement with the Figaro group on the 18th. ChatGPT will display summaries and extracts with a link to the original article. No amount has been published. Press · Siècle Digital (17 November 2026) · CB News (18 November 2026) Canada: the media lawsuit against OpenAI will be heard in OntarioThe Ontario Court of Appeal has dismissed OpenAI’s appeal: five Canadian media groups can litigate in Ontario, where the US fair use exception does not apply. The ruling concerns the court’s jurisdiction, not the merits. Court decision reported by the press · Tech Times (13 November 2026) Music: a model trained only on artists who opt inThe start-up Sonilo has signed with the distributor Symphonic: it will be able to train its music model on recordings from independent artists and labels who sign up, with attribution tracking to pay them. Neither amounts nor the number of sign-ups have been published. Specialist press · Music Business Worldwide (18 November 2026) Canva lets users switch off its generative featuresA setting now lets users switch off Canva’s generative AI tools, and a second one highlights templates created by humans. Team administrators keep control for their organisation. Non-generative tools, such as background removal, remain active. Developer announcement, reported · PetaPixel (19 November 2026) My perspective For a company that creates or publishes content, the lesson is twofold. The content you produce has a value that AI companies are starting to pay for. And the content you generate must be traceable: in Europe, marking is becoming an obligation; elsewhere, a condition of contracts. |
BACK TO CONTENTS ↑ | THIS WEEK’S TUTORIAL Mapping your AI tools against the AI Act in one hourThis tutorial requires no legal expertise. It takes two people: the one who knows the tools in use best, and a manager. It does not replace a lawyer’s opinion; it prepares the conversation. | 01 List the tools and usesNote every AI tool the company uses or sells: writing assistant, agent on the website, module built into software, screening tool. For each, state what it is for, who uses it, and whether it is bought, developed for you or sold to your clients. 02 State who is the providerFor each tool, tick: we use it as is, we had it developed, we sell it under our name. The last two boxes need particular attention: that is where the company can become a provider within the meaning of the regulation. 03 Spot sensitive usesFlag the tools that touch on recruitment, employee evaluation, training, credit or insurance. These are the high-risk uses in Annex III, subject to obligations from 2 December 2027. 04 Check transparencyFor each tool that interacts with customers or candidates, check that the presence of an AI is announced from the first interaction. For those that produce published content, ask the provider whether it is marked, and how. 05 Record trainingNote who has been trained on each tool, when and on what. Half a page is enough. It is the record of the measures taken for AI literacy, which you will need to be able to show. 06 Set dates and questionsWrite down the relevant deadlines: 2 December 2026, 2 December 2027. For each sensitive tool, prepare three questions for the provider and a review date in six months. Your mapping sheetTool: … Use: … Used as is, developed for us or sold under our name: … Sensitive use (employment, training, credit, insurance): … AI presence announced: … Content marked: … People trained and date: … Applicable deadline: … Questions for the provider: … Next review date: … Share my map or my questions ↗ | Coming up this week24 November · third IGENSIA RH evening "AI at work: what is really changing?", in Paris, after Toulouse and Lyon. End of November · expected ruling on freezing the second lawsuit against Suno. 2 December · marking of generated content becomes mandatory for tools already on the market, and two new AI Act prohibitions apply. 4 December · deadline to volunteer for testing the European fundamental rights impact assessment template. | Help shape the next edition.Do you work in HR? What has AI already changed in your daily work: recruitment, training, interviews, relations with candidates? A specific example interests me more than a general opinion. Has your provider or software publisher ever talked to you about the AI Act? Reply to this email: I will be happy to respond, and your feedback may feed into a future edition. Write to Jérôme ↗ | The connecting threadLast week, we asked what would be left if your AI tool shut down on Friday. This week asks the opposite question: when the tool works, who answers for it? The European regulation, US courts and bank contracts converge: liability follows whoever puts the tool into circulation under their name, and whoever hands it their access. A question for your next team meeting: which of our AI tools could we explain to a client, a candidate or an inspector, and which could we not? | HELP SHAPE WHAT COMES NEXT Was this edition useful?One click is enough. I read every response. Very useful ↗Useful, but too long ↗Not for me ↗ | Who might find this newsletter useful?An HR manager, a small-business owner or a colleague supporting teams? Introduce them to SIGNAL AI and invite them to join the discussion. Subscribe to SIGNAL AI ↗ | SIGNAL AI #013 · Jérôme Denis · JDENIS Consulting 7 Bis Chemin de la Butte, 31400 Toulouse, France. Selection closed on 21 November 2026. Company-reported results are identified; proposed ideas are not promises of gains. Unsubscribe · Reply to Jérôme · Back to the top ↑ |
|