Artificial intelligence lawyers (AI)

The Article 5 prohibitions and the Article 4 AI literacy obligation have applied since 2 February 2025. Obligations on general-purpose AI models (GPAI, Articles 51 to 55), governance and penalties have applied since 2 August 2025. Obligations on Annex III high-risk systems and on transparency (Article 50) are due to apply on 2 August 2026, subject to the final adoption of the 'digital omnibus' proposed by the Commission on 19 November 2025, which could conditionally postpone this deadline until 2 December 2027 at the latest.

Until the omnibus is formally adopted, the 2 August 2026 deadline remains legally binding: serious operators do not bet on the postponement, they prepare ahead.

The regulation classifies AI systems into four levels that determine the intensity of compliance required: unacceptable risk (prohibited since 2 February 2025), high risk (strict obligations on risk management, data quality, human oversight and CE marking), limited risk (transparency obligations and AI content marking), and minimal risk (no specific obligations).

If you wrap the OpenAI or Mistral API to offer a service to your clients, you are a deployer towards the GPAI provider, and a provider of your own AI system built around it. If you fine-tune an open-source model, you remain in principle a deployer, unless your modifications cross the indicative threshold set by the Commission on 18 July 2025: use of more than one third of the original model's compute or one third of the 10²³ FLOP threshold.

In practice, that threshold is crossed only by a tiny minority of fine-tuners. If your use case falls under Annex III (HR, scoring, biometrics, justice, education), you become a high-risk system provider and inherit all Article 16 obligations, regardless of the GPAI provider you rely on.

The European Commission published, through the AI Office, on 10 July 2025, the Code of Practice for General-Purpose AI Models, formally endorsed by the Commission and the AI Board on 1 August 2025. This voluntary code is the preferred channel to demonstrate compliance with Articles 53 and 55. It contains three chapters: transparency (applicable to all GPAI providers), copyright (policy of respecting Article 4 of Directive 2019/790 opt-outs and mitigating the risk of infringing outputs) and safety and security (reserved for systemic-risk models exceeding the 10²⁵ FLOP threshold). The AI Office also published, on 24 July 2025, a standardised template for the summary of training content, mandatory for all GPAI providers, to be updated at least every six months and published in an accessible manner.

Providers of high-risk systems must, under Article 16: implement a risk management system across the lifecycle, ensure quality and representativeness of training data, maintain technical documentation in line with Annex IV, automatically log system operation, give transparent information to deployers, ensure effective human oversight, and pass a conformity assessment before placing on the market.

Deployers, under Article 26, must use the system in line with its instructions, entrust oversight to trained persons with real authority, monitor operation and report incidents, retain logs for at least six months, inform affected persons and, in the cases set out in Article 27, conduct a fundamental rights impact assessment prior to deployment.

Each Member State must establish at least one regulatory sandbox operational by 2 August 2026. The CNIL has already run several thematic sandboxes (public services, silver economy). The interest is legal as much as operational: a provider following the validated plan benefits from administrative immunity (no AI Act fine) and the exit report facilitates subsequent conformity assessment.

Penalties are among the highest in EU law. For Article 5 prohibited practices: up to €35 million or 7% of worldwide annual turnover. For other breaches: up to €15 million or 3%. For inaccurate information given to authorities: up to €7.5 million or 1%. GPAI model providers are subject, under Article 101, to specific fines of up to 3% of global turnover or €15 million.

Under French law, settled case law holds that only a natural person can be recognised as an author (Article L.111-1 of the Intellectual Property Code, as interpreted in particular by Cass. ass. plén., 7 March 1986, 'Pachot', and by the CJEU, Infopaq, C-5/08, 16 July 2009). The work must also bear the imprint of the author's personality.

A creation purely generated by AI, without creative human intervention, is therefore not protectable. This aligns with US courts: the D.C. Circuit Court of Appeals, in Thaler v. Perlmutter on 18 March 2025, unanimously confirmed that the Copyright Act requires human authorship, and the US Copyright Office, in its January 2025 report, took the same line.

AI-assisted works can however be protected if human intervention is creatively substantial: iterative and nuanced prompt selection, creative curation of outputs, recompositions, retouching, editorial arrangement.

Proof comes from rigorous documentation of the creative process (preserving prompts, iterations, choices made). This is precisely what we set up for the creative studios and agencies we advise.

Platform licences differ and their terms shift with US litigation.

Midjourney provides that the subscriber 'owns all Assets' created subject to payment and compliance, and grants other users a licence on public images (Stealth mode in the Pro plan restricts visibility). For OpenAI / DALL-E, the user owns the outputs subject to compliance with the terms, but those terms do not guarantee copyrightability. Suno authorises commercial use of songs generated under Pro or Premier plans, with no royalty take, but explicitly states that music entirely generated by AI is not eligible for copyright protection.

ElevenLabs varies by plan; cloning a real voice requires explicit consent and a separate commercial licence. For agencies delivering to clients, we systematically include in service contracts transparency clauses (mention of AI use), limited eviction warranties, framed assignment of rights covering arrangement and human modifications, and explicit allocation of third-party claim risk.

Law no. 2024-449 of 21 May 2024, known as 'SREN' (securing and regulating digital space), significantly strengthened the repression of deepfakes in French law. It amended Article 226-8 of the Penal Code to punish bringing to the public's or a third party's knowledge visual or audio content generated by algorithmic processing representing a person's image or words without their consent, where it is not obvious that the content is algorithmically generated or expressly stated to be so. The penalty is one year imprisonment and €15,000 fine, raised to two years and €45,000 when distributed via an online public communication service.

Article 21 of the same law created Article 226-8-1 of the Penal Code, specifically punishing non-consensual sexual deepfakes: two years imprisonment and €60,000 fine, raised to three years and €75,000 online.

These offences combine with identity theft (Art. 226-4-1), invasion of privacy (Art. 226-1), cyberbullying (Art. 222-33-2-2), sexual blackmail (Art. 312-10) and, on the civil side, image rights (Article 9 of the Civil Code) and neighbouring vocal rights.

The AI Act adds, from 2 August 2026, a transparency obligation requiring deployers to visibly label deepfakes (Article 50(4)) and providers to technically mark AI-generated content (watermarking, C2PA metadata, cryptographic signatures). The first draft Code of Practice on Transparency was published on 17 December 2025 and proposes a common 'AI' pictogram across Europe.

Synthetic reproduction of a performer's voice without authorisation infringes their neighbouring rights, guaranteed by Articles L.212-1 et seq. of the Intellectual Property Code. The performer holds an inalienable moral right to the respect of their name, status and performance, and economic rights to authorise or prohibit the fixation, reproduction and communication of their performance.

Collective management runs through ADAMI for principal performers and SPEDIDAM for accompanying performers. The voice is also protected under Article 9 of the Civil Code, extended by case law to vocal attributes.

The dispute between Scarlett Johansson and OpenAI over the GPT-4o 'Sky' voice in May 2024 brought these issues to the spotlight: OpenAI suspended the voice without going to trial, after an independent acoustic analysis concluded an exceptional similarity with the actress's voice.

In France, no specific case law has yet been issued on synthetic voices, but the field is mapped by neighbouring rights, the Civil Code and now Articles 226-8 and 226-8-1 of the Penal Code. The positions of ADAMI, SACEM and SCAM converge on the requirement of explicit prior consent and equitable remuneration.

Using protected works to train AI models raises the text and data mining (TDM) exception.

Directive (EU) 2019/790, transposed into French law in Article L.122-5-3 of the Intellectual Property Code, provides two exceptions: one for scientific research (Article 3 of the directive, non-overridable), and a general one (Article 4 of the directive), overridable subject to rights holders' opt-out, in particular by machine-readable means for online content.

On 12 March 2025, the Syndicat national de l'édition, the Société des gens de lettres and the Syndicat national des auteurs et compositeurs sued Meta Platforms before the Paris judicial court for training Llama, in particular, on the Books3 dataset containing about 200,000 protected books.

The first French case against a major generative AI player, based on counterfeiting and parasitism.

Abroad, Bartz v. Anthropic (N.D. Cal., 23 June 2025) found training on lawfully acquired books to be 'quintessentially transformative' but refused fair use for building a library from pirated works, leading to a $1.5 billion settlement granted preliminary approval on 25 September 2025. Kadrey v. Meta (N.D. Cal., 25 June 2025) also retained fair use with more nuanced reasoning.

NYT v. OpenAI (S.D.N.Y., order of 4 April 2025) is ongoing and led to OpenAI producing 20 million ChatGPT logs. In the UK, Getty Images v. Stability AI (EWHC, 4 November 2025) dismissed secondary infringement and found only limited trademark infringement.

In Germany, OLG Hamburg, in its 10 December 2025 decision, clarified that an opt-out expressed in natural language but not technically machine-readable is not valid. For publishers, creators and platforms, we deploy opt-out strategies cumulating robots.txt, the TDMRep protocol, C2PA or IPTC PLUS metadata and clear mentions in the terms.

Algorithms as such are not patentable in Europe, but the software implementing them can be protected by copyright (Article L.112-2, 13°, of the Intellectual Property Code). Computer-implemented inventions may also be patented under certain conditions (technical effect, technical character of the solution). We define the protection strategy best suited to your architecture.

Databases benefit from dual protection under European law: copyright on their original structure (Article L.112-3 of the Intellectual Property Code) and the sui generis right of database producers for substantial investment (Articles L.341-1 et seq. of the same code, transposing Directive 96/9/EC). Your training datasets are strategic assets whose protection must be anticipated from collection.

Law no. 2018-670 of 30 July 2018, transposing Directive (EU) 2016/943, provides specific protection for trade secrets (Articles L.151-1 et seq. of the Commercial Code). For your models, hyperparameters, training pipelines and technical know-how, implementing reasonable protection measures (NDAs, access restrictions, traceability of consultations, logging, information classification) is essential to benefit from this regime, which does not apply to information whose confidentiality has not been actively preserved.

The European Data Protection Board (EDPB) issued, on 17 December 2024, its Opinion 28/2024 on data protection aspects in AI models, at the request of the Irish authority.

The opinion lays down several key points: an AI model can only be deemed anonymous if the probability of extracting or re-identifying personal data, by reasonable means, is insignificant (many large language models do not meet this criterion, due to the risk of memorisation and regurgitation); legitimate interest can serve as a legal basis for development and training, subject to the three-step test set out in EDPB Guidelines 1/2024; unlawful processing in the development phase may, on a case-by-case basis, contaminate the deployment phase.

The CNIL published, on 19 June 2025, its final recommendations on legitimate interest as a legal basis for the development and training of AI models, following a public consultation. Three cumulative conditions must be met: a lawful, real and precise interest; necessity of processing (minimisation test); a favourable balancing.

Expected additional safeguards include pseudonymisation, access restriction, exclusion of sensitive data, transparent information of individuals, a discretionary and prior right to object, a right to erasure, and technical anti-memorisation and anti-regurgitation measures.

The CNIL also published, on 19 June 2025, a web scraping practical sheet, and rolled out in December 2025 a traceability tool for open-source models. The PANAME project, launched in February 2026, is to deliver a GDPR audit framework for models.

For executives facing uncontrolled deployment of generative AI by their teams ('shadow AI'), the internal AI policy is no longer best practice: it is now the only enforceable tool.

Concretely, the policy must define authorised tools and their version (the Enterprise version and the standard version do not offer the same contractual guarantees), list categories of prohibited data (identifiable client data, ongoing contracts, health and HR data, undisclosed financial data, proprietary source code, industrial secrets), require systematic human supervision of outputs, organise a procedure for declaring new tools, and provide for disciplinary sanctions.

To make the policy enforceable against an employee, it must be annexed to the internal regulations under the procedure of Articles L.1321-4, R.1321-2 and L.2312-8 of the Labour Code (consultation of the social and economic committee, filing with the labour inspectorate and the registry of the local labour court). Without this formal integration, the policy remains a mere recommendation, with no disciplinary basis.

Since 2 February 2025, Article 4 of the AI Act requires providers and deployers to take measures to ensure a sufficient level of AI literacy of their staff and any persons acting on their behalf. No certification is imposed, but a training register is strongly recommended.

In the event of a breach of another AI Act obligation, the absence of training will likely be retained as an aggravating factor, by parallel with GDPR case law. INFLUXIO designs and delivers enforceable AI training programmes, tailored to your sector and the maturity of your teams: executive sessions for boards and management committees (AI Act framework, governance, directors' liability), role-based modules for exposed functions (HR, marketing, legal, R&D, IT, customer support), technical workshops for data and product teams (provider/deployer qualification, risk classification, Annex IV technical documentation, logging, human oversight), and dedicated programmes for creators and agencies (generative AI, copyright, Midjourney, Suno and ElevenLabs licences, deepfakes and the SREN law). Each programme is delivered with its enforceable AI policy, time-stamped training register, sector-specific case studies and skills assessment, forming a file directly enforceable in any review by the AI Office, the CNIL or a market authority.

Any structuring AI deployment processing data at scale, sensitive data or involving systematic evaluation triggers a data protection impact assessment (Article 35 GDPR). Transfers outside the EU to AI providers (often US-based) must be governed by standard contractual clauses, where applicable relying on the EU-US Data Privacy Framework.

Processing contracts must contain all Article 28 GDPR safeguards, supplemented by AI-specific stipulations (prohibition on using data to train models, unless explicit opt-in; log retention; cooperation with authorities).

For a startup founder marketing a SaaS built on a GPAI provider, the terms must contain: definitions aligned with the AI Act, a clear qualification of provider and deployer roles, allocation of compliance (who bears the risk management system, technical documentation, six-month minimum logging, post-market surveillance), an explicit hallucination disclaimer with a human supervision obligation for the deployer, IP framing on outputs and inputs, a DPA aligned with Article 28 GDPR, prohibition on reuse of user data for training (unless opt-in), a serious incident notification obligation combining Articles 26(5) AI Act and 33 GDPR, liability caps, and a termination clause for regulatory reclassification.

Ownership of rights over developments, deliverables, training data and intermediate models must be defined precisely, distinguishing what is assigned, what is licensed and what remains the contractor's property (generic models, pre-existing technical building blocks, know-how). We secure your contracts with providers, freelancers and internal teams.

For agencies and creative studios, client contracts must now include seven specific clauses: transparency on AI use, limited eviction warranty (the agency cannot fully guarantee the absence of third-party claims on training data), assignment of rights covering arrangement and human modifications (without warranty of copyrightability of purely AI elements), capped liability and exclusion of indirect damages, GDPR compliance for data injected by the client, audit and log retention of at least six months (consistent with Article 26 of the AI Act), reversibility and secure destruction at the end of the relationship.

Liability for AI-caused damage is rapidly evolving. The proposed AI Liability Directive (Proposal 2022/0303) was withdrawn by the Commission in 2025, due to lack of agreement between co-legislators.

We already integrate limitation, sharing and subrogation clauses adapted to this emerging framework, working with specialised brokers to identify the relevant cyber and professional civil liability covers.