Kohen Avocats, artificial intelligence law in Paris
AI lawyer in Paris: AI Act, AI at work and generated content
Classifying an AI system and the timetable of the EU regulation as amended in 2026, the obligations of providers and deployers, introducing AI tools in the workplace and consulting the works council (CSE), copyright in generated content and training data, deepfakes, contracts and liability: as AI lawyers in Paris, the firm advises businesses that design or use AI, and defends people who bear its effects.
The firm can be reached 24/7. Regulation (EU) 2026/1744 of 8 July 2026 moved the deadlines: the rules for the high-risk systems listed in Annex III apply from 2 December 2027, and two new prohibitions, covering realistic sexual images produced without consent and child sexual abuse material, apply from 2 December 2026.
4,9/5
Average Google review rating270 reviews
I received high-quality support from this criminal lawyer. His professionalism, his complete command of the case and his constant involvement at every stage of the proceedings deserve to be highlighted. Communication was clear, structured and responsive, with precise advice suited to my situation. I recommend this lawyer to anyone looking for a rigorous and effective criminal defence.
Mr Kohen handled my criminal case. I am very satisfied with the work he did and I recommend him to you without any hesitation.
Excellent criminal defence lawyer, he got me out of trouble in a protection order case brought abusively by my ex
Very satisfied with the solution he found for my defence in court. All the best
Hassan Kohen is a young, seasoned and highly competent lawyer. He is an excellent listener, with a great deal of empathy and a strong human sensitivity. Available and responsive, he has a perfect command of his profession.
I had a very good experience with Hassan Kohen. He took the time to listen to me carefully and to understand my situation properly. He is a lawyer who gives sound advice, kind and available, and who knows how to steer his clients in the right direction. I recommend him without hesitation to anyone looking for an attentive and competent professional.
Mr Kohen advised and supported us well. I thank him for everything he did.
Paris Bar
Emergencies 24/7
Preliminary case review
Send your documents to an artificial intelligence lawyer. Receive an analysis.
Hassan Kohen
Avocat at the Paris Bar, artificial intelligence law: compliance with the EU regulation, AI at work, intellectual property in generated content, contracts and litigation.
Official profile in the national directory of French lawyers (CNB) · 4.9/5 on Google (270 reviews)
Page updated on 1 November 2026.
Quick answer
Regulation (EU) 2024/1689 on artificial intelligence (the AI Act) applies, according to their role, to providers that develop an AI system and place it on the market under their own name, and to deployers that use it under their own authority in a professional context (Article 3). Its prohibitions and the AI literacy obligation have applied since 2 February 2025, the rules on general-purpose AI models since 2 August 2025, and the rest of the text since 2 August 2026, transparency included (Article 113). Regulation (EU) 2026/1744 of 8 July 2026 postponed the requirements for high-risk systems to 2 December 2027 for those in Annex III and to 2 August 2028 for those in Annex I. French law adds further rules: consultation of the works council (CSE), in companies with at least fifty employees, before a new technology is introduced (Article L. 2312-8 of the French Labour Code, Code du travail), copyright reserved for original human creation, offences targeting content generated without consent (Articles 226-8 and 226-8-1 of the French Criminal Code, Code pénal), and civil and contractual liability. Kohen Avocats, in Paris 16th, acts for businesses and for the individuals concerned alike. Contact the firm on +33 6 89 11 34 45.
An AI tool to deploy, a works council to consult, generated content to challenge?
Understanding your situation
EU AI Act and artificial intelligence law in Paris, explained
The AI Act defines an AI system as “a machine-based system that is designed to operate with varying levels of autonomy” (Article 3 of Regulation (EU) 2024/1689). What applies and from when, who is a provider or a deployer, what a business must do before deploying a tool, who owns generated content, and who is liable for damage.
Key dates
The EU AI Act and its timetable, as amended in July 2026
The dates that matter, taken from the published text and from the regulation that amended it.
2 February 2025
Chapters I and II apply: definitions, the AI literacy obligation and prohibited practices, such as social scoring, exploiting vulnerabilities or inferring emotions in the workplace and in education institutions, except for medical or safety reasons (Articles 4, 5 and 113 of Regulation (EU) 2024/1689).
2 August 2025
The rules on general-purpose AI models (Chapter V), governance (Chapter VII), penalties (Chapter XII) and notifying authorities (Chapter III, Section 4) apply, with the exception of Article 101 (Article 113, point (b)).
2 August 2026
General date of application of the regulation (Article 113, first paragraph): it applies in particular to the transparency obligations of Article 50, which cover conversational systems, synthetic content and deep fakes.
2 December 2026
The two prohibitions added by Regulation (EU) 2026/1744 apply: systems that generate realistic images, videos or sounds depicting the intimate parts of an identifiable person, or such a person engaged in sexually explicit activities, without their consent, and systems that generate child sexual abuse material (Article 5(1), points (ba) and (bb)). The same date applies to the marking of outputs from generative systems placed on the market before 2 August 2026 (Article 111(4)).
2 December 2027
Chapter III, Sections 1 to 3, applies to the high-risk systems in Annex III: biometrics, critical infrastructure, education, employment and workforce management, access to essential services, law enforcement, migration, justice (Article 113, point (c)(i), as worded by Regulation (EU) 2026/1744).
2 August 2028
The same rules apply to the high-risk systems of Article 6(1), that is, safety components of products covered by the harmonisation legislation listed in Annex I (Article 113, point (c)(ii)). High-risk systems intended for use by public authorities that are already in service must comply by 2 August 2030 at the latest (Article 111(2)).
The postponement is no longer a hypothesis. Regulation (EU) 2026/1744 of 8 July 2026, published in the Official Journal of the European Union on 24 July 2026, entered into force on the third day following its publication (Article 4 of that regulation). It replaces point (c) of Article 113: Chapter III, Sections 1, 2 and 3, applies “from 2 December 2027 as regards AI systems classified as high-risk under Article 6(2) and Annex III” (our translation). The other deadlines are unchanged: the prohibitions and AI literacy have applied since February 2025, and transparency since August 2026.
The same regulation rewrote the AI literacy obligation. Providers and deployers “take measures to promote the development of AI literacy” (our translation) among their staff, and “This obligation does not require providers or deployers to ensure a specific level of AI literacy of an individual” (our translation) (new Article 4, Regulation (EU) 2026/1744, Article 1, point (5)). The obligation remains: training, usage instructions and an internal AI policy are still the means of documenting it.
The regulation is directly applicable, but the Member States lay down the rules on penalties (Article 99) and designate their national competent authorities (Article 70). On generated content, see our article on AI Act obligations for AI-generated content.
Classification
Provider or deployer, and which risk level: four questions in order
The obligations depend on the company's role and on how the system is used, not on the tool itself.
Is the tool an AI system? The regulation covers a machine-based system, with a degree of autonomy, that infers from the input it receives how to generate predictions, content, recommendations or decisions (Article 3, point (1)). A general-purpose AI model, capable of performing a wide range of distinct tasks, is also subject to its own regime (Article 3, point (63), and Article 53).
What is the company's role? The provider develops the system or has it developed and places it on the market or puts it into service under its own name or trademark; the deployer uses it under its own authority, except in the course of a personal non-professional activity (Article 3, points (3) and (4)). A deployer or distributor becomes a provider if it markets a high-risk system under its own trademark, makes a substantial modification to it, or changes its intended purpose so that it becomes high-risk (Article 25).
Is the use prohibited? Article 5 bans, among other things, subliminal or manipulative techniques, the exploitation of vulnerabilities, social scoring, emotion inference at work and in education and, from 2 December 2026, systems that generate realistic sexual images of an identifiable person without their consent. This last prohibition is qualified: placing on the market is prohibited only if producing such content is the system's intended purpose, or a reasonably foreseeable result that adequate technical safeguards do not prevent, and use is prohibited only if the deployer uses the system for that purpose (Article 5(1a), inserted by Regulation (EU) 2026/1744).
Is the system high-risk? It is if it constitutes or is part of a product subject to third-party conformity assessment under Annex I, or if it falls within an Annex III area, such as recruitment and the screening of applications (Article 6). An Annex III system escapes this classification if it does not pose a significant risk to health, safety or fundamental rights, but it remains high-risk whenever it performs profiling of natural persons (Article 6(3)).
The deployer of a high-risk system uses it in accordance with the instructions for use, assigns human oversight to competent persons, ensures that the input data under its control is relevant and monitors the operation of the system (Article 26). An employer that deploys such a system in the workplace has a specific obligation: “deployers who are employers shall inform workers’ representatives and the affected workers that they will be subject to the use of the high-risk AI system” (Article 26(7)). These obligations will apply to Annex III systems from 2 December 2027.
Most everyday uses, such as a writing assistant, a summarisation tool or a chatbot, are not high-risk. They are nonetheless subject to the transparency obligations that have applied since 2 August 2026: the provider designs a system intended to interact with people so that they “are informed that they are interacting with an AI system”, unless this is obvious, and the outputs of generative systems must be “marked in a machine-readable format and detectable as artificially generated or manipulated” (Article 50 of Regulation (EU) 2024/1689).
The provider of a general-purpose AI model keeps technical documentation, informs the providers that integrate it and puts in place “a policy to comply with Union law on copyright and related rights” (Article 53). A business that integrates such a model into its product checks these points in the contract. See our page on IT contracts, and the one dedicated to deeptech.
At work
Deploying AI in the workplace, or challenging its introduction
Consulting staff representatives, the duty to ensure safety, informing the people concerned: what the employer checks before deploying a tool.
You are the employer
- List the AI tools in use, including those employees use on their own initiative, and classify each one under the regulation.
- In a company with at least fifty employees, inform and consult the works council (CSE) before introducing a new technology that affects working conditions.
- Assess the tool's effects on workload and work organisation with the employees concerned, in real conditions, before any reorganisation.
- Adopt an AI usage policy, train teams, and set rules for confidential and personal data.
You are the works council or an employee
- Request that the consultation be opened and that written information be provided on the tool, its functions and its effects.
- Apply to the urgent applications judge (juge des référés) to have the deployment suspended until the consultation is closed, or a reorganisation whose risks have not been assessed.
- Require prior information on recruitment methods and data collection systems.
In companies with at least fifty employees, the works council is informed and consulted on “the introduction of new technologies, and any significant change altering health and safety conditions or working conditions” (our translation) (Article L. 2312-8 of the French Labour Code). Urgent applications judges apply this provision to AI tools. The Paris Court of Appeal upheld the suspension, subject to a daily penalty (astreinte), of a writing assistant and of ChatGPT until the consultation was closed, noting that the in-house tool “is not a mere variant of word-processing software” (our translation) (CA Paris, pôle 1, ch. 2, 21 May 2026, No. 25/13234); the Nanterre urgent applications judge suspended two skills-management programs, even though the previous software already used AI (TJ Nanterre, réf., 29 January 2026, No. 25/02856). The consultation procedure and the works council's urgent application are explained on our page dedicated to digital employment law.
A completed consultation does not shield the employer. In Créteil, the information-consultation procedure on the elimination of eight sub-editor posts had ended with the works council's opinion of 15 July 2026, and the works council did not challenge it. The urgent applications judge ruled on the employer's duty to ensure safety (Articles L. 4121-1 and L. 4121-2 of the French Labour Code): the share of tasks taken over by generative AI had been assessed “without any participation of the employees concerned, or any testing in real conditions” (our translation), and implementing the project “on this approximate basis, not supported by real-life testing” (our translation) appeared to represent “a serious and specific risk to the physical and mental health of several categories of employees” (our translation). Nor did the employer show any measures to prevent AI-related risks recorded in its occupational risk assessment document (document unique). The reorganisation and the collective redundancy were suspended until tests in real conditions had been carried out (TJ Créteil, réf., 15 September 2026, No. 26/00991).
Other provisions govern AI applied to people. In companies with at least fifty employees, the works council is informed, before they are used, of recruitment-assistance methods or techniques and of automated personnel-management processing (Article L. 2312-38 of the French Labour Code); job applicants are expressly informed of the methods used in their regard, which must be relevant to the purpose pursued (Article L. 1221-8), and “no information concerning an employee personally may be collected by a device that has not been brought to their attention beforehand” (our translation) (Article L. 1222-4). See also our article on AI imposed without consulting the works council.
The points the firm examines in every case:
Content
Generated content, training data, deepfakes
What is protected, what may be used to train a model, what is punishable.
Copyright protects works of the mind, whatever their kind, form of expression, merit or purpose, and arises from the mere fact of creation (Articles L. 111-1 and L. 112-1 of the French Intellectual Property Code, Code de la propriété intellectuelle). It requires an original work that bears the stamp of its author's personality.
Authorship belongs, unless proven otherwise, to the person under whose name the work is disclosed (Article L. 113-1); in a dispute where an architect's ownership was contested on the ground that the work had been disclosed under the name of his company, the Paris judicial court (tribunal judiciaire) inferred that “only a natural person can be an author” (our translation) (TJ Paris, 3e ch. 2e sect., 29 May 2026, No. 24/09710).
Any reproduction or performance made without the author's consent is unlawful, including by adaptation or transformation (Article L. 122-4), and infringement (contrefaçon) is a criminal offence (Article L. 335-2).
Text and data mining is permitted on lawfully accessible works, whatever its purpose, “unless the author has objected to it in an appropriate manner, in particular by machine-readable means for content made available to the public online” (our translation) (Article L. 122-5-3, III).
The maker of a database who can show a substantial investment may prohibit the extraction and reuse of a substantial part of its contents (Articles L. 341-1 and L. 342-1).
Disseminating without consent visual or audio content generated by algorithmic processing and depicting a person's image or words is punishable, if it is not obvious that such content is involved or if this is not expressly stated (Article 226-8 of the French Criminal Code). Sexual content is punished more severely, and the provision that covers it contains neither of these two exceptions (Article 226-8-1).
None of the decisions cited concerns content generated by an AI. The firm infers that such content is protected only if a person has expressed identifiable creative choices in it. The Paris judicial court, following the Court of Justice of the European Union (4 December 2025, C-580/23), recalls that “not only choices dictated by various constraints, in particular technical ones, that bound the author when creating the object are not freely made and creative, but also those which, although freely made, do not bear the stamp of the author's personality by giving that object a unique appearance” (our translation). It refused protection to photomontages composed from a stock image library, which “do not bear the stamp of the personality of an author giving them a unique appearance” (our translation), and dismissed the copyright infringement claim (TJ Paris, 3e ch. 2e sect., 27 March 2026, No. 24/08115). The rights holder must also identify for itself what constitutes the originality it claims.
Training a model on third-party content is assessed in light of the text and data mining exception and the reservation of rights. The provider of a general-purpose AI model must identify and comply with reservations of rights expressed by rights holders (Article 53(1)(c) of Regulation (EU) 2024/1689). Scraping a database, by contrast, is assessed in light of the maker's investment: according to the French Court of Cassation (Cour de cassation), “where a substantial part of the contents of a database is extracted and reused, the court must ascertain whether this use creates a risk to the recovery of the database maker's investments” (our translation). It dismissed the appeal of the aggregator that was reproducing the listings of a car sales website (Cass. 1re civ., 15 October 2025, No. 23-23.167).
Deepfakes fall under both criminal law and the regulation: deployers of a system that generates a deep fake “shall disclose that the content has been artificially generated or manipulated” (Article 50(4)). See our page on copyright, and our articles on deepfakes published without a label and on removing a photograph from a training dataset.
Litigation
Contracts, liability and automated decisions
Who is liable for an incorrect output, for damage or for a decision taken by an algorithm.
Before signing
A party who knows information that is decisive for the other party's consent must disclose it, where the other party is legitimately unaware of it or relies on its co-contractor; this duty does not extend to the estimate of the value of the service, and the parties may neither limit nor exclude it (Article 1112-1 of the French Civil Code, Code civil). An AI tool contract sets out each party's role under the regulation, the use of the company's data, confidentiality, ownership of outputs and warranties.
Non-performance
The debtor is ordered to pay damages for non-performance of the obligation or for delay, unless it can show force majeure (Article 1231-1). An obligation of result or of means, a service level or a limitation of liability is negotiated before any dispute arises.
Damage caused to a third party
Any act whatever of a person that causes damage to another obliges the person by whose fault it occurred to make reparation (Article 1240), including through negligence or carelessness (Article 1241). A user who publishes a false or harmful output without checking it incurs personal liability.
Defective product
The producer is liable for damage caused by a defect in its product, whether or not it is bound by a contract with the victim (Article 1245), for personal injury and, above an amount set by decree, for damage to property other than the product itself (Article 1245-1). A product is “any movable property” (our translation) (Article 1245-2): the text does not say whether software or an AI system is one. A product is defective “when it does not provide the safety that one is legitimately entitled to expect” (our translation) (Article 1245-3), subject to the grounds for exemption, including the development risk defence (Article 1245-10).
Automated decision
The data subject has “the right not to be subject to a decision based solely on automated processing, including profiling”, which produces legal effects or similarly significantly affects them, subject to exceptions that come with the right to obtain human intervention (Article 22 of the GDPR).
Public administration
An individual decision taken on the basis of algorithmic processing includes an explicit statement informing the person concerned, who may obtain the rules defining the processing and the main characteristics of its implementation (Article L. 311-3-1 of the French Code of Relations between the Public and the Administration, Code des relations entre le public et l'administration), including the degree to which the processing contributed to the decision and its parameters (Article R. 311-3-1-2).
The expected safety is assessed in concrete terms. The French Court of Cassation overturned a court of appeal judgment that had ruled out product liability “without examining whether, in light of the circumstances and in particular their presentation and the use that could reasonably be expected of them, the products alleged to be defective provided the safety that one was legitimately entitled to expect” (our translation) (Cass. 1re civ., 9 December 2020, No. 19-17.724). The judgment concerned chemical products that had spoiled wine, and recalls that the regime compensates, besides personal injury, only damage “exceeding 500 euros, resulting from damage to property other than the defective product itself” (our translation). Its application to software or an AI system depends on whether it qualifies as a product, which neither the Civil Code nor the decisions cited settle. Directive (EU) 2024/2853 of 23 October 2024 settles the question for the future: a product “includes electricity, digital manufacturing files, raw materials and software” (Article 4, point (1)), for products placed on the market or put into service after 9 December 2026, the deadline for its transposition (Articles 2 and 22). As at 31 October 2026, Article 1245-2 of the Civil Code is still worded as in 2016. If the regime applies, the presentation of the tool, its marketing claims and its instructions will weigh in the assessment of the expected safety.
An algorithm does not dispense with human assessment. For the Parcoursup university admissions procedure, the French Council of State (Conseil d'État) notes that “the decision taken on each application cannot be based exclusively on an algorithm and requires, on the contrary, an assessment of the merits of the applications by the selection committee, then by the head of the institution” (our translation). This solution rests on Article L. 612-3 of the French Education Code (code de l'éducation), which is specific to Parcoursup, interpreted in light of the decision of the Constitutional Council of 3 April 2020; it is not a general rule of the Code of Relations between the Public and the Administration. The student union that sought the algorithmic processes and source codes in fact lost on that point, as these provisions are interpreted “as not requiring the publication or disclosure to third parties of the algorithmic processing itself and of the corresponding source codes” (our translation) (CE, 4e et 1re ch. réunies, 15 July 2020, No. 433297). In employment relationships, the French Court of Cassation inferred from an app with a geolocation system that tracked the courier's position in real time, combined with a power to impose sanctions, “the existence of a power to direct and to monitor the performance of the service, characterising a relationship of subordination” (our translation) (Cass. soc., 28 November 2018, No. 17-20.079).
IT disputes follow their own rules of evidence and expert assessment: see our pages on IT litigation, the GDPR and legaltech, and our article on a defamatory accusation produced by ChatGPT.
FAQ
Frequently asked questions
Fourteen detailed answers, based on the provisions in force and on verified decisions.
What does an AI lawyer do?
An AI lawyer translates a project or a dispute into the applicable rules, then into action. For a business that designs or integrates a tool, the lawyer classifies the system under Regulation (EU) 2024/1689: AI system or ordinary software, provider or deployer, prohibited practice, high risk or a mere transparency obligation (Articles 3, 5, 6 and 50). The lawyer then sets out the project's own timetable, as it results from Regulation (EU) 2026/1744, and the list of documents to be produced: usage policy, AI literacy measures, information for the people concerned, documentation required from the provider.
The lawyer also deals with the French law that applies alongside the regulation. In companies with at least fifty employees, introducing an AI tool that amounts to a new technology affecting working conditions requires consultation of the works council (CSE) (Article L. 2312-8 of the French Labour Code), failing which the urgent applications judge may suspend its use, as the Paris Court of Appeal confirmed in May 2026 (CA Paris, pôle 1, ch. 2, 21 May 2026, No. 25/13234). Generated content raises questions of copyright, database rights and criminal law (Articles 226-8 and 226-8-1 of the French Criminal Code).
The firm drafts and negotiates AI tool contracts, assists with the consultation of staff representatives, advises on the use of content and data, and litigates both for businesses and for individuals harmed by an AI system: employees, job applicants, authors, people depicted in a deepfake.
When does the EU AI Act apply?
In stages. The original text sets general application at 2 August 2026, with two earlier dates: Chapters I and II, which contain the definitions, AI literacy and the prohibited practices, since 2 February 2025, and the rules on general-purpose AI models, governance and penalties since 2 August 2025 (Article 113 of Regulation (EU) 2024/1689).
Regulation (EU) 2026/1744 of 8 July 2026, published on 24 July 2026, amended this timetable. Chapter III, Sections 1 to 3, which sets out the requirements and obligations for high-risk systems, applies “from 2 December 2027 as regards AI systems classified as high-risk under Article 6(2) and Annex III” (our translation), and from 2 August 2028 for those under Article 6(1) and Annex I. Two new prohibitions, covering the generation without consent of realistic sexual images of an identifiable person and the generation of child sexual abuse material, apply from 2 December 2026, within the limits set by the new Article 5(1a).
The same regulation gives providers of generative systems placed on the market before 2 August 2026 until 2 December 2026 to mark their outputs (Article 111(4)), and keeps at 2 August 2030 the deadline for high-risk systems intended for use by public authorities that are already in service (Article 111(2)). The transparency obligations of Article 50, for their part, have applied since 2 August 2026. The firm determines the relevant date for each system.
Our company uses ChatGPT or Copilot: are we a “deployer” under the AI Act?
In most cases, yes. A deployer is the natural or legal person “using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity” (Article 3, point (4), of Regulation (EU) 2024/1689). A company that makes a chatbot available to its teams falls within this definition, even if the tool is supplied by a third party.
The obligations then depend on the use. For drafting or summarising, the main ones are the AI literacy obligation, which requires measures to promote AI literacy among staff (Article 4 as worded by Regulation (EU) 2026/1744), and the transparency obligations: the deployer of a system that generates a deep fake discloses it, and a deployer that publishes generated text to inform the public on matters of public interest discloses that it was generated by AI, unless there has been human review and editorial responsibility (Article 50(4)). If the tool is used to screen applications or assess employees, it falls under Annex III, and the obligations of the deployer of a high-risk system will apply from 2 December 2027 (Article 26).
The company may also become a provider, if it markets a high-risk system under its own trademark or changes its intended purpose (Article 25). Employment law applies in addition: in companies with at least fifty employees, making ChatGPT available was held to be the introduction of a new technology subject to consultation of the works council (CA Paris, pôle 1, ch. 2, 21 May 2026, No. 25/13234). See our article on Copilot, ChatGPT, the AI policy and works council consultation.
How do we know whether our AI system is “high-risk”?
By applying Article 6 of Regulation (EU) 2024/1689. A system is high-risk in two cases. The first covers a system intended to be used as a safety component of a product covered by the harmonisation legislation listed in Annex I, or which is itself such a product, where that product is subject to a third-party conformity assessment. The second covers the systems listed in Annex III: biometrics, critical infrastructure, education, employment and workforce management, access to essential services, law enforcement, migration, administration of justice (Article 6 and Annex III).
An Annex III system is not high-risk where it “does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons”, for example if it performs a narrow procedural task or improves the result of a previously completed human activity. This derogation falls away whenever the system performs profiling of natural persons, and a provider that relies on it documents its assessment before placing the system on the market (Article 6(3) and (4)). Regulation (EU) 2026/1744 specified that systems used solely for user assistance, performance optimisation or quality control, with no link to safety, are not safety components.
Classification governs everything else: risk management system, data quality, documentation, human oversight, conformity assessment. It is based on the system's actual intended purpose and must be reviewed whenever its use changes.
Can an employer cut jobs based on an assessment of the tasks taken over by AI?
It can, but the assessment must be serious, because it engages the employer's duty to ensure safety. The employer “takes the measures necessary to ensure the safety and protect the physical and mental health of workers” (our translation), and ensures that these measures are adapted to take account of changing circumstances (Article L. 4121-1 of the French Labour Code); it assesses the risks that cannot be avoided and takes account of the state of technical progress (Article L. 4121-2). A reorganisation that transfers tasks to an AI tool changes the workload of those who remain: that is the risk that has to be measured.
The Créteil urgent applications judge demonstrated this in September 2026. The works council consultation on the elimination of eight sub-editor posts was closed, but the assessment that the tool took over about 70% of their tasks had been carried out “without any participation of the employees concerned, or any testing in real conditions” (our translation), and the employer did not show the measures to prevent AI-related risks that it had itself identified. Finding a manifestly unlawful disturbance in light of the duty to ensure safety, the judge suspended the reorganisation and the collective redundancy until tests in real conditions and a new assessment of the other employees' workload had been carried out, subject to a daily penalty of €8,000, while refusing the interim payment requested by the works council (TJ Créteil, réf., 15 September 2026, No. 26/00991).
In companies with at least fifty employees, the project is also subject to works council consultation, as a measure likely to affect the size or structure of the workforce and as the introduction of new technologies (Article L. 2312-8). The firm helps the employer document the assessment, the trials and the prevention measures, or assists the works council and employees who challenge it. The consultation itself is explained on our page dedicated to digital employment law.
Can a job applicant or an employee be assessed by AI?
Yes, under strict conditions. Under French employment law, “a job applicant is expressly informed, before they are implemented, of the recruitment-assistance methods and techniques used in their regard” (our translation), and these methods must be relevant to the purpose pursued (Article L. 1221-8 of the French Labour Code). No information concerning the applicant may be collected by a device that has not been brought to their attention beforehand (Article L. 1221-9), and the same rule protects employees (Article L. 1222-4). In companies with at least fifty employees, the works council is informed, before they are used, of recruitment-assistance methods or techniques (Article L. 2312-38).
The EU regulation adds further rules. Systems intended for recruitment, screening applications, evaluating candidates or making decisions on employment relationships are listed in Annex III and will be subject to the high-risk requirements from 2 December 2027; the employer, as deployer, will have to inform workers' representatives and the affected workers before the system is put into service (Article 26(7)). Emotion inference in the workplace, for its part, has been prohibited since 2 February 2025, except for medical or safety reasons (Article 5).
Finally, a decision based solely on automated processing that produces legal effects or similarly significantly affects the person is in principle excluded, with a right to human intervention where it is permitted (Article 22 of the GDPR). See our article on job applications rejected by AI.
Is an AI-generated image or text protected by copyright?
Not in itself. Copyright protects a work of the mind by the mere fact of its creation (Article L. 111-1 of the French Intellectual Property Code), provided that it is original, and authorship requires a person: “Authorship belongs, unless proven otherwise, to the person or persons under whose name the work is disclosed” (our translation) (Article L. 113-1). Ruling on an architect's ownership of an interior design, the Paris judicial court inferred that “only a natural person can be an author” (our translation) (TJ Paris, 3e ch. 2e sect., 29 May 2026, No. 24/09710). None of these decisions concerns content generated by an AI: what follows is the firm's analysis.
The question therefore becomes one of human input. The same court, following the Court of Justice of the European Union, requires creative choices, freely made, that reflect the author's personality, and recalls that “it is for the person claiming copyright protection, who is able to characterise the creative choices, freely made, reflecting their personality, to identify what characterises that originality” (our translation). It refused protection to photomontages composed from a stock image library (TJ Paris, 3e ch. 2e sect., 27 March 2026, No. 24/08115). An output obtained through a simple prompt is hard to defend; a work whose composition, selection, editing or arrangement reflects personal choices may be protected for that contribution.
The firm helps preserve evidence of the creative process: successive versions, prompts, edits, composition choices. It also checks the tool's terms of use, which govern ownership and commercial use of the outputs. See our article on proving authorship of a novel written with AI.
Can an AI be trained on works or data published online?
Under conditions. The French Intellectual Property Code permits copies made for text and data mining on works that have been lawfully accessed, “unless the author has objected to it in an appropriate manner, in particular by machine-readable means for content made available to the public online” (our translation), the copies being destroyed at the end of the mining (Article L. 122-5-3, III and Article L. 122-5, 10°). The provider of a general-purpose AI model must also put in place a copyright compliance policy and identify the reservations of rights expressed (Article 53(1)(c) of Regulation (EU) 2024/1689).
Databases have their own protection. A maker who can show a substantial investment may prohibit the extraction or reuse of a substantial part of the contents of its database (Article L. 342-1). The French Court of Cassation requires the court to ascertain whether the use creates a risk to the recovery of the maker's investments, and dismissed the appeal of a website that was reproducing a competitor's listings on a massive scale, the loss of traffic and the fall in turnover having been established (Cass. 1re civ., 15 October 2025, No. 23-23.167).
Personal data contained in training corpora remains subject to the GDPR. The firm audits the provenance of datasets, reservations of rights and licences, or acts for rights holders whose content has been used. See our article on database scraping by a competitor.
What are the legal risks of sharing a deepfake in France?
A criminal conviction, and a civil claim. The French Criminal Code treats as equivalent to a montage the act of bringing to the knowledge of the public or of a third party “visual or audio content generated by algorithmic processing and depicting a person's image or words, without their consent” (our translation), if it is not obvious that the content was generated algorithmically or if this is not expressly stated. The penalty is one year's imprisonment and a €15,000 fine, increased to two years and €45,000 where an online public communication service is used (Article 226-8 of the French Criminal Code).
Sexual content is punishable by two years' imprisonment and a €60,000 fine, and by three years and €75,000 where it is published online (Article 226-8-1); this provision contains neither the obviousness exception nor the statement exception, so flagging that the content is generated does not remove the offence. Impersonating someone in order to disturb their peace or harm their honour is also an offence (Article 226-4-1). The victim may also bring a civil claim based on the right to respect for private life (Article 9 of the French Civil Code).
The EU regulation supplements this framework: since 2 August 2026, the deployer of a system that generates a deep fake discloses that the content has been artificially generated or manipulated (Article 50(4)), and from 2 December 2026, systems that generate realistic sexual images of an identifiable person without their consent are prohibited (Article 5(1), point (ba), inserted by Regulation (EU) 2026/1744). This prohibition does not cover a general-purpose generator equipped with adequate technical safeguards, nor its user who does not use it “for the purpose of generating or manipulating such material” (our translation) (Article 5(1a)). The firm has the content formally recorded, obtains its removal and files a criminal complaint. See our article on intimate deepfakes.
What transparency obligations apply to a chatbot or to AI-generated content?
Those of Article 50 of Regulation (EU) 2024/1689, applicable since 2 August 2026. The provider of a system intended to interact directly with people designs it so that they “are informed that they are interacting with an AI system”, unless this is obvious from the point of view of a reasonably well-informed, observant and circumspect person (Article 50(1)). The provider of a generative system ensures that its outputs are marked in a machine-readable format and detectable as artificially generated or manipulated (Article 50(2)); for systems placed on the market before 2 August 2026, the deadline runs until 2 December 2026 (Article 111(4), inserted by Regulation (EU) 2026/1744).
The deployer has its own obligations: informing people exposed to an emotion recognition or biometric categorisation system (paragraph 3), and disclosing deep fakes and generated text published to inform the public on matters of public interest, unless there has been human review and editorial responsibility (paragraph 4). Regulation (EU) 2026/1744 tasked the Commission with encouraging codes of practice on marking and labelling, and allows it to adopt an implementing act if they prove inadequate.
The firm drafts the disclosures, checks the marking solutions offered by the provider and keeps a record of human review where the exception is relied on. See our article on proving human review of AI-generated text.
What are the penalties for breaching the EU AI Act?
Heavy administrative fines, under rules laid down by each Member State within the framework of the regulation. Non-compliance with the prohibitions in Article 5 is subject to a fine of up to €35 million or, for a company, 7% of its total worldwide annual turnover for the preceding financial year, whichever is higher (Article 99(3) of Regulation (EU) 2024/1689).
Breaches of the obligations of providers, authorised representatives, importers, distributors and deployers, and of the transparency obligations of Article 50, carry fines of up to €15 million or 3% of worldwide turnover, and supplying incorrect, incomplete or misleading information to the authorities up to €7.5 million or 1% (Article 99(4) and (5)). For SMEs and start-ups, the lower of the two amounts applies (paragraph 6). Regulation (EU) 2026/1744 extended this lower-ceiling rule to small mid-cap companies for the fines under paragraphs 4 and 5, and added certain Article 25 obligations to the breaches subject to penalties (amended Article 99).
Penalties must be effective, proportionate and dissuasive, and take into account the interests of SMEs and their economic viability (Article 99(1)). Other sanctions coexist: GDPR fines where personal data is involved, and civil and criminal judgments under French law. The firm assists businesses during an inspection and in challenging a penalty.
AI and the GDPR: what rules apply to personal data?
The GDPR applies whenever an AI system processes data relating to an identified or identifiable person, both in training and in use. It requires compliance with the principles of lawfulness, purpose limitation and data minimisation (Article 5 of the GDPR), a legal basis (Article 6), informing people at the time of collection (Article 13) and a data protection impact assessment where the processing, “in particular using new technologies”, is likely to result in a high risk to people's rights and freedoms (Article 35). Above all, it protects against full automation: the person has “the right not to be subject to a decision based solely on automated processing, including profiling”, which produces legal effects or similarly significantly affects them (Article 22 of the GDPR). Where the decision is permitted by the contract or by explicit consent, the person retains the right to obtain human intervention, to express their point of view and to contest the decision.
Regulation (EU) 2026/1744 inserted an Article 4a which allows, exceptionally and to the extent strictly necessary, the processing of special categories of data to detect and correct bias, subject to cumulative safeguards: impossibility of achieving this with other data, pseudonymisation, restricted access, deletion once the bias has been corrected, and an entry in the record of processing activities.
The public sector follows its own rules. An individual decision taken on the basis of algorithmic processing “includes an explicit statement informing the person concerned” (our translation) (Article L. 311-3-1 of the French Code of Relations between the Public and the Administration). For Parcoursup, the Council of State held, interpreting Article L. 612-3 of the Education Code, which is specific to that procedure, in light of the decision of the Constitutional Council of 3 April 2020, that the decision cannot be based exclusively on an algorithm, while refusing disclosure of the source codes to third parties (CE, 4e et 1re ch. réunies, 15 July 2020, No. 433297). See our page on the GDPR.
Who is liable when AI causes damage?
A person, never the system. Depending on the situation, the victim brings a claim against the user, the integrator or the producer. Anyone who causes damage to another through their fault, negligence or carelessness must make reparation (Articles 1240 and 1241 of the French Civil Code): a business that publishes a false, defamatory or infringing output without checking it is liable for its own fault. Between contracting parties, non-performance of an obligation, for example of availability, security or compliance, gives rise to damages unless there is force majeure (Article 1231-1).
Product liability applies without a contract: the producer is liable for damage caused by a defect in its product (Article 1245), whether personal injury or, above an amount set by decree, damage to property other than the product itself (Article 1245-1). The defect is assessed in concrete terms, taking into account in particular the presentation of the product (Article 1245-3 and Cass. 1re civ., 9 December 2020, No. 19-17.724), and the producer may escape liability by proving that the state of knowledge did not allow the defect to be detected (Article 1245-10). It remains to be seen whether software or an AI system is a product: Article 1245-2 refers to “any movable property” (our translation) without saying so, and the decisions cited do not settle the point. Directive (EU) 2024/2853 expressly includes software for products placed on the market or put into service after 9 December 2026, the deadline for its transposition; as at 31 October 2026, Article 1245-2 is still worded as in force since 1 October 2016.
The contract allocates the rest: warranties, limitation of liability, insurance, recourse between provider and integrator. The firm negotiates these before deployment, and builds the evidence after the damage: logs, versions, prompts, formal records. See also our article on case law invented by AI.
What are an AI lawyer's fees in Paris?
Except in an emergency, in cases of force majeure or where the lawyer acts under full legal aid, fees are set out in a written agreement, in accordance with Article 10 of the Law of 31 December 1971: before you commit, you know what the engagement covers and what it costs. The same article provides that fees take into account, according to custom, the difficulty of the case, the costs incurred and the work carried out.
The firm favours a fixed fee for each stage: classifying a system under the regulation, AI usage policy and AI literacy plan, information note and support for the works council consultation, drafting or auditing an AI tool contract, analysing rights in generated content or training data, formal notice, urgent application, criminal complaint, proceedings on the merits. Any new engagement is covered by an amendment to the agreement, with no unexpected invoicing.
For a first opinion on documents, the preliminary case review at €80 incl. VAT, through the form on this page, results in a written analysis within 24 working hours. First contact on +33 6 89 11 34 45.
References
Official texts and related pages
The sources for this page, checked on 1 November 2026, and the firm's related areas of practice.
Official texts
- Regulation (EU) 2024/1689 of 13 June 2024 on artificial intelligence: definitions (Article 3), AI literacy (Article 4), prohibited practices (Article 5), high risk (Article 6 and Annex III), value chain and deployers (Articles 25 and 26), transparency (Article 50), general-purpose models (Article 53), penalties (Article 99), application (Articles 111 and 113)
- Regulation (EU) 2026/1744 of 8 July 2026, digital omnibus on AI: new Article 4, Article 4a, new prohibitions in Article 5, amended Articles 99, 111 and 113
- Regulation (EU) 2016/679, Articles 5, 6, 13, 22 and 35: principles, lawfulness, information, automated individual decision-making, impact assessment
- Directive (EU) 2024/2853 of 23 October 2024: liability for defective products, software included (Articles 2, 4 and 22)
- Articles L. 2312-8, L. 2312-38, L. 1221-8, L. 1221-9 and L. 1222-4, L. 4121-1 and L. 4121-2 of the French Labour Code: works council consultation, recruitment, prior information, duty to ensure safety
- Articles L. 111-1, L. 112-1, L. 113-1, L. 122-4 and L. 335-2 of the French Intellectual Property Code: work, author, unlawful reproduction, infringement
- Articles L. 122-5, L. 122-5-3, L. 341-1 and L. 342-1 of the French Intellectual Property Code: text and data mining, databases
- Articles 226-1, 226-4-1, 226-8 and 226-8-1 of the French Criminal Code: privacy, identity theft, content generated without consent
- Articles 9, 1112-1, 1231-1, 1240 and 1241 of the French Civil Code: privacy, duty to inform, contractual and tort liability
- Articles 1245, 1245-1, 1245-2, 1245-3 and 1245-10 of the French Civil Code: liability for defective products
- Articles L. 311-3-1, R. 311-3-1-1 and R. 311-3-1-2 of the French Code of Relations between the Public and the Administration: administrative decisions based on algorithmic processing
- Article 10 of Law No. 71-1130 of 31 December 1971: fee agreement
Decisions cited
- CA Paris, pôle 1, ch. 2, 21 May 2026, No. 25/13234: ChatGPT and a writing assistant, suspension for lack of works council consultation
- TJ Nanterre, réf., 29 January 2026, No. 25/02856: AI-based HR software, consultation and suspension
- TJ Créteil, réf., 15 September 2026, No. 26/00991: reorganisation based on generative AI suspended under the duty to ensure safety, after the consultation
- TJ Paris, 3e ch. 2e sect., 29 May 2026, No. 24/09710: an architect's ownership, only a natural person can be an author
- TJ Paris, 3e ch. 2e sect., 27 March 2026, No. 24/08115: originality refused to photomontages made from a stock image library
- Cass. 1re civ., 15 October 2025, No. 23-23.167: database scraping and recovery of the investment
- Cass. 1re civ., 9 December 2020, No. 19-17.724: concrete assessment of the safety expected of a product, recoverable damage
- Cass. soc., 28 November 2018, No. 17-20.079: algorithmic geolocation, power to impose sanctions, subordination
- CE, 4e et 1re ch. réunies, 15 July 2020, No. 433297: Parcoursup, decision not exclusively algorithmic, source codes not disclosable
An AI tool to deploy, a works council to consult, generated content to challenge?
The firm can be reached 24/7. Send us the description of the tool, the provider's contract, the works council notice of meeting or the content in question: Hassan Kohen, avocat at the Paris Bar, examines the classification under the EU AI Act, the applicable deadlines and the rights at stake, then tells you the appropriate course of action: compliance, AI usage policy and consultation, contract negotiation, formal notice, urgent application, criminal complaint or proceedings on the merits.