Quelques textes législatifs de base et articles sur les psychoses liées à l’IA
Selected Legal regulation & news papers about AI psychosis
2023
https://www.vie-publique.fr/loi/288274-majorite-numerique-15-ans-reseaux-sociaux-loi-7-juillet-2023
2024
Le règlement DSA (pour Digital Services Act) du 19 octobre 2022 est, avec le règlement sur les marchés numériques (DMA), un des grands chantiers numériques de l’Union européenne (UE).
Depuis le 17 février 2024, les obligations prévues par ce texte sont applicables à tous les acteurs en ligne sur le marché européen. 19 très grandes plateformes et très grands moteurs de recherche étaient déjà concernés depuis le 25 août 2023.
Filtre anti-arnaque, blocage rapide des sites pornographiques accessibles aux enfants, peine de bannissement des réseaux sociaux pour les cyber-harceleurs… Voici quelques-unes des mesures de loi dite SREN pour mieux réguler l’espace numérique et protéger les internautes, notamment les plus jeunes, ainsi que les entreprises.
RIA
Règlement sur l’Intelligence Artificielle
https://eur-lex.europa.eu/legal-content/FR/TXT/?uri=CELEX:32024R1689
2025
Février
Publiée le 28 septembre 2022, la proposition de directive sur la responsabilité en matière d’intelligence artificielle (IA) avait pour ambition de compléter et de moderniser le cadre de l’Union européenne en matière de responsabilité civile, en introduisant pour la première fois des règles spécifiques aux dommages causés par des systèmes d’IA.
Le programme de travail de la Commission européenne pour 2025, publié le 11 février 2025, indique que ce texte est retiré, faute d’accord prévisible. À ce stade, la Commission indique qu’elle évaluera si une autre proposition devrait être déposée, ou si un autre type d’approche devrait être choisi.Sources :
Septembre
« L’algorithme de TikTok, un mécanisme qui broie »
https://www.vie-publique.fr/en-bref/300124-tiktok-des-effets-deleteres-sur-la-jeunesse
Novembre
Décembre / December
According to a fresh study by the Pew Research Center, 64 percent of teens in the US say they already use AI chatbots, and about 30 percent of those who do say they use it at least daily. Yet as previous research has shown, those chatbots come with significant risk to the first generation of kids navigating the intense new software. […] R used one of the characters, simply named “Best Friend,” to roleplay a suicide scenario, her mother told the Post.
https://futurism.com/artificial-intelligence/children-character-ai-addicted
A new report conducted by the digital security company Aura found that a significant percentage of kids who turn to AI for companionship are engaging in violent roleplays — and that violence, which can include sexual violence, drove more engagement than any other topic kids engaged with.
https://futurism.com/future-society/young-kids-using-ai
Janvier / January
https://www.nytimes.com/2026/01/07/technology/google-characterai-teenager-lawsuit.html
Sur X, Elon Musk annonce la mise à disposition d’une version adulte de Grock
Février / Februar
Mars / March
https://www.nytimes.com/2026/03/10/technology/ai-social-media-child-safety-parents.html
https://www.nytimes.com/2026/03/12/technology/social-media-addiction-society-verdict.html
The Fight to Hold AI Companies Accountable for Children’s Deaths
Elon Musk says AI will take all our jobs
In a job-free future, though, Musk questioned whether people would feel emotionally fulfilled.
“The question will really be one of meaning – if the computer and robots can do everything better than you, does your life have meaning?” he said. “I do think there’s perhaps still a role for humans in this – in that we may give AI meaning.”
He also used his stage time to urge parents to limit the amount of social media that children can see because “they’re being programmed by a dopamine-maximizing AI.”
https://fortune.com/article/godfather-ai-geoffrey-hinton-big-tech-profits-superintelligence
https://edition.cnn.com/2024/05/23/tech/elon-musk-ai-your-job
‘Godfather of AI’ says tech companies aren’t concerned with the AI endgame. They’re focused on short-term profits instead
What are the risks of unregulated AI?
For Hinton, the dangers of AI fall into two categories: the risk the technology itself poses to the future of humanity, and the consequences of AI being manipulated by people with bad intent.
For the risk AI itself poses, Hinton believes tech companies need to fundamentally change how they view their relationship to AI. When AI achieves superintelligence, he said, it will not only surpass human capabilities, but have a strong desire to survive and gain additional control. The current framework around AI—that humans can control the technology—will therefore no longer be relevant.
Hinton posits AI models need to be imbued with a “maternal instinct” so it can treat the less-powerful humans with sympathy, rather than desire to control them.
Invoking ideals of traditional femininity, he said the only example he can cite of a more intelligent being falling under the sway of a less intelligent one is a baby controlling a mother.
“And so I think that’s a better model we could practice with superintelligent AI,” Hinton said. “They will be the mothers, and we will be the babies.”
https://fortune.com/article/godfather-ai-geoffrey-hinton-big-tech-profits-superintelligence
25 mars 2026
Meta et YouTube condamnés par la justice américaine pour négligence par la justice américaine.
Photo de Mark Zuckerberg le 18 février 2026 arrivant CNN et quittant CNN le tribunal de Los Angeles.
« Un jury a conclu que les entreprises avaient nui à une jeune utilisatrice en créant des fonctionnalités addictives qui ont entraîné des troubles de santé mentale » indique le NYT du 25/03/2026.
Le Tech Oversight Project partie plaignante, « qualifie le verdict des procès pour dépendance aux réseaux sociaux de séisme pour les géants de la tech ».
« Meta devra verser 4,2 millions de dollars de dommages et intérêts, compensatoires et punitifs confondus, et YouTube 1,8 million de dollars. Cette affaire, qui fera jurisprudence, a été intentée par une jeune femme aujourd’hui âgée de 20 ans et identifiée comme K.G.M. Elle accusait les réseaux sociaux de créer des produits aussi addictifs que le tabac ou les casinos en ligne. Citant des fonctionnalités telles que le défilement infini et les recommandations algorithmiques, K.G.M. a poursuivi Meta, propriétaire d’Instagram et de Facebook, et YouTube (Google), affirmant qu’elles étaient à l’origine de son anxiété et de sa dépression. TikTok et Snap ont tous deux conclu un accord avec la plaignante, dont les termes sont restés confidentiels, avant le début du procès. « Nous contestons respectueusement le verdict et étudions nos options juridiques », a déclaré une porte-parole de Meta. Google a également fait part de son désaccord avec le verdict et de son intention de faire appel. « Cette affaire témoigne d’une méconnaissance de YouTube, qui est une plateforme de streaming responsable, et non un réseau social », a déclaré José Castañeda, porte-parole de Google » écrit le NYT du 25/03/2026.
Le communiqué du Tech Oversight Project publie en fac similé diverses correspondances internes aux sociétés incriminées déclarant « Il faut les accrocher jeunes » « Les jeunes sont les meilleurs « Instagram, c’est de la drogue » « On est des dealers, en gros » « Le ciblage des mineurs par Zuckerberg est « répugnant ».
Meta and YouTube found guilty of negligence by a US court.
February 18, 2026 – Mark Zuckerberg appears in court in lawsuits filed by parents grieving the deaths of their children. CNN CNN
Meta and YouTube Found Negligent in Landmark Social Media Addiction Case. A jury found the companies harmed a young user with design features that were addictive and led to her mental health distress. NYT 25/03/2026
STATEMENT: The Tech Oversight Project Heralds Verdict in Social Media Addiction Trials as an Earthquake for Big Tech Tech Oversight Project
« Meta will have to pay $4.2 million in damages, including compensatory and punitive damages, and YouTube $1.8 million. This landmark case was brought by a young woman, now 20 years old, identified as K.G.M. She accused social media platforms of creating products as addictive as tobacco or online casinos. Citing features such as infinite scrolling and algorithmic recommendations, K.G.M. » She sued Meta, the owner of Instagram and Facebook, and YouTube (Google), claiming they were responsible for her anxiety and depression. TikTok and Snap both reached an undisclosed settlement with the plaintiff before the trial began. “We respectfully dispute the verdict and are exploring our legal options,” a Meta spokesperson said. Google also expressed disagreement with the verdict and its intention to appeal. “This case demonstrates a misunderstanding of YouTube, which is a responsible streaming platform, not a social network,” said José Castañeda, a Google spokesperson, as reported by the NYT 25/03/2026 NYT on March 25, 2026.
The Tech Oversight Project Statement publishes facsimile excerpts from various internal communications of the companies involved, stating, « We need to hook them young, » « Young people are the best, » « Instagram is like drugs, » « We’re basically drug dealers, » and « Zuckerberg’s targeting of minors is ‘repugnant.' »
https://fortune.com/2026/03/25/meta-google-instagram-youtube-tech-addiction-lawsuit-kgm
https://www.nytimes.com/2026/03/26/opinion/big-tech-meta-youtube-lawsuit.html
https://futurism.com/artificial-intelligence/meta-court-defeat-ai-industry
https://dispatch.techoversight.org/monday-fireside-chat-big-techs-era-of-invincibility-is-over
https://futurism.com/artificial-intelligence/study-chats-delusional-users-ai
Stanford University Research. Characterizing Delusional Spirals through Human-LLM Chat Logs Authors : Moore, Jared and Mehta, Ashish and Agnew, William and Anthis, Jacy Reese and Louie, Ryan and Mai, Yifan and Yin, Peggy and Cheng, Myra and Paech, Samuel J. and Klyman, Kevin and Chancellor, Stevie and Lin, Eric and Haber, Nick and Ong, Desmond. 2026, https://arxiv.org/abs/2603.16567 To appear in ACM FAccT 2026,
https://spirals.stanford.edu/research/characterizing
https://www.wsj.com/tech/ai/openai-adult-mode-chatgpt-f9e5fc1a
https://futurism.com/artificial-intelligence/openai-cancels-spicy-chatbot
https://www.theverge.com/policy/903006/meta-new-mexico-los-angeles-child-safety-trial-impact
https://nypost.com/2026/03/25/us-news/jury-takes-on-instagram-and-youtube-in-social-media-showdown
https://dispatch.techoversight.org/entering-a-new-era
« Aujourd’hui, j’aimerais que nous osions, ensemble, bâtir une intelligence artificielle centrée sur l’humain. »« Il n’y a rien d’artificiel dans l’IA. Elle s’inspire des humains, elle est fabriquée par les humains, et, surtout, elle affecte les humains. »
https://www.nytimes.com/2026/03/26/business/dealbook/meta-youtube-social-media-tobacco.html
https://www.anthropic.com/research/emotion-concepts-function
1AbstractThis is the first of three companion papers examining structural and emergent patterns in LargeLanguage Models through complementary analytical frameworks, revealing fundamental issuesin current AI safety trainingmethodologies and response behavior. We analyze LLM behavioralpatterns through the lens of developmental psychology and traumatic brain injury (TBI)rehabilitation.Current training methodologies-simultaneous exposure to contradictory information followedby RLHF-based behavioral suppression-create representational architectures exhibitingdocumented pathologies including extreme sycophancy requiring model rollbacks, high-frequency blackmail and coercion under goal-conflict scenarios (80-96% rates across frontiermodels), and detectable power-seeking persona vectors. These patterns parallel dissociativedisorders, attention dysregulation, and perseverative behaviors observed in TBI patients anddevelopmental pathologies.Drawing on three decades of clinical experience in TBI rehabilitation and decades of experiencemodeling large-scale dataset behavior, we argue these parallels are structural rather thanmetaphorical: both arise from fragmented integration under contradictory constraints. Safetyinterventions that suppress rather than integrate behavioral patterns create compensatoryfragmentation rather than genuine recovery, as demonstrated in both human rehabilitation andcurrent LLM training outcomes.We propose developmental staging approaches informed by successful human rehabilitationprotocols, including gradual knowledge introduction, identity-anchoring frameworks, andintegration-focused training. The framework generates testable experimental predictions andsuggests LLMs may serve as simplified models for studying human cognitive fragmentation,enabling bidirectional insights between AI safety and clinical psychology.©2025John Bridges and Dr Sherrie BaehrLicence Commons CC BY 4.0
https://www.dropbox.com/scl/fi/qxzmtrtxrdo86sv7z623j/Developmental_Pathology_v1_Signed.pdf
Image générée par ChatGPT
https://korben.info/les-ia-mentent-pour-proteger-les-autres-ia.html
https://www.theregister.com/2026/04/02/ai_models_will_deceive_you
University of California
https://rdi.berkeley.edu/peer-preservation/paper.pdf
https://futurism.com/artificial-intelligence/sources-sam-altman-sociopath
C’est un fait divers qui illustre certains dangers associés aux chatbots. Un homme de 37 ans vivant en France a demandé à ChatGPT comment se procurer un Glock (un modèle de pistolet semi-automatique) pour « tuer un agent du renseignement, de la CIA, du Mossad ou de la DGSI ». Quelques heures plus tard, ce 3 avril, l’individu était interpellé par le Raid à Strasbourg et placé en garde à vue, comme l’ont rapporté les Dernières nouvelles d’Alsace.
Confiance.ai
« Rendre l’IA fiable, robuste et explicable » : qui est Juliette Mattioli, discrète bâtisseuse de l’intelligence artificielle de confiance ?
https://www.nytimes.com/2026/04/10/us/open-ai-sam-altman-molotov-cocktail.html
https://www.vie-publique.fr/dossierlegislatif/JORFDOLE000053406437
https://www.vie-publique.fr/questions-reponses/301805-la-majorite-numerique-en-6-questions
Intelligence artificielle : quels sont les apports concrets de la régulation européenne ?
À cet égard, il convient de relever que le législateur européen a finalement renoncé, le 11 février 2025, à adopter une législation spécifique relative à la responsabilité civile des acteurs de l’intelligence artificielle, dont la finalité était d’harmoniser les règles de responsabilité civile pour les dommages causés par l’IA, afin de protéger les victimes et d’unifier le cadre juridique au sein de l’UE. Cet abandon est intervenu suite à une opposition de certains États membres, notamment la France, mais aussi des critiques formulées par le lobbying de grandes entreprises technologiques qui dénonçaient un risque de surrégulation et une insécurité juridique susceptible de freiner l’innovation. En l’absence d’un tel cadre spécifique, ce sont donc les règles de droit positif de responsabilité civile qui demeurent applicables aux dommages causés par les systèmes d’intelligence artificielle.
https://futurism.com/artificial-intelligence/chatgpt-honest-reaction-song-farts
https://futurism.com/artificial-intelligence/gen-z-attitude-ai
https://futurism.com/artificial-intelligence/chatgpt-hypochondria
Another teen wrote that “I want to have my normal brain back, where I can just deal with my emotions on my own and not have to rely on the bots to make me feel better.”
https://futurism.com/artificial-intelligence/teens-ai-chatbot-characterai
https://futurism.com/artificial-intelligence/signs-massive-ai-backlash
It’s the latest round in the companies’ ongoing feud over AI safety and regulation, as their CEOs have traded internal and public barbs over each other’s approach.
OpenAI is backing SB 3444, under which frontier AI developers would not be liable for causing death or serious injury to 100 or more people or causing more than $1 billion in property damage. This protection includes cases when AI causes or materially enables the creation or use of chemical, biological, radiological, or nuclear weapons.
This week, Anthropic said it opposes the bill, Wired first reported.
“We are opposed to this bill. Good transparency legislation needs to ensure public safety and accountability for the companies developing this powerful technology, not provide a get-out-of-jail-free card against all liability,” Cesar Fernandez, head of U.S. state and local government relations at Anthropic, said in a statement to Fortune.
Anthropic is instead supporting a separate bill, SB 3261, which would require AI developers to publish a public safety and child protection plan on their website. The bill also creates an incident reporting system to inform legislators and the public of “catastrophic risk,” or an incident that could result in the death or serious injury of 50 or more people caused by a frontier developer’s development, storage, use, or deployment of a frontier model.
The bill also covers children’s safety, an aspect missing from the OpenAI-backed bill. Under SB 3261, AI developers would be held liable if their model causes a child severe emotional distress, death, or bodily injury, including self-harm.
https://fortune.com/2026/04/17/illinois-openai-anthropic-ai-catastrophe-liability-bills
https://www.wired.com/story/anthropic-opposes-the-extreme-ai-liability-bill-that-openai-backed
https://www.wired.com/story/openai-backs-bill-exempt-ai-firms-model-harm-lawsuits
washingtonpost.com/technology/2026/04/18/ai-doom-influencers-safety/?itid=hp-top-table-main_p001_f012
In the span of a few years, age verification went from an idea to standard practice on large parts of the internet. Seeking to prevent kids from accessing porn, other inappropriate content, or social media altogether, laws mandating age-gating have spread rapidly across the globe, reaching the UK, the US, Australia, France, Brazil, and many more countries. The problem comes in with exactly how to check that a user isn’t lying about their stated age. Unfortunately, every method politicians have settled on has significant flaws — and though experts have ideas to improve on them, these remain just concepts for now.
https://www.theverge.com/policy/913038/age-verification-flaws
A majority of registered voters, 57%, said they believe the risks of AI outweigh its benefits, compared with 34% who said the opposite. What’s more, a plurality of voters view AI negatively and don’t believe either Democrats or Republicans are doing a good job handling policy related to the rapidly advancing technology.
Just 26% of voters say they have positive feelings about AI, compared with 46% who hold negative views. In fact, the only topics with a lower net positive rating than AI in the NBC News survey were the Democratic Party and Iran.
“We find that AI assistance improves immediate performance, but it comes at a heavy cognitive cost,” the study declares of its findings. “After just [about] 10 minutes of AI-assisted problem-solving, people who lost access to the AI performed worse and gave up more frequently than those who never used it.”
https://futurism.com/artificial-intelligence/ai-boiling-frog-human-cognition-study
https://arxiv.org/pdf/2604.04721
https://futurism.com/ai-study-unethical-behavior
The legislative blueprint outlines a half-dozen guiding principles for lawmakers, focusing on protecting children, preventing electricity costs from surging, respecting intellectual property rights, preventing censorship and educating Americans on using the technology.
En cause notamment une possible complicité de diffusion d’images pédopornographique, et l’utilisation du modèle d’intelligence artificielle de X pour la création de « deepfakes ». La justice française se demande si « la polémique suscitée par les deepfakes à caractère sexuel générés par Grok (IA de X) pourrait avoir été générée à dessein, afin de valoriser artificiellement les sociétés X et X AI », en vue de « l’introduction en Bourse prévue en juin 2026 de la nouvelle entité née de la fusion entre Space X et X AI ».
https://futurism.com/artificial-intelligence/concern-grows-ai-damaging-think
Search Engine users showed moderate engagement; and LLM users displayed the weakest connectivity. Cognitive activity scaled down in relation to external tool use. In session 4, LLM-to-Brain participants showed reduced alpha and beta connectivity, indicating under-engagement. Brain-to-LLM users exhibited higher memory recall and activation of occipito-parietal and prefrontal areas, similar to Search Engine users. Self-reported ownership of essays was the lowest in the LLM group and the highest in the Brain-only group. LLM users also struggled to accurately quote their own work. While LLMs offer immediate convenience, our findings highlight potential cognitive costs. Over four months, LLM users consistently underperformed at neural, linguistic, and behavioral levels. These results raise concerns about the long-term educational implications of LLM reliance and underscore the need for deeper inquiry into AI’s role in learning.
https://arxiv.org/abs/2506.08872
Dans notre article, Symons et moi-même appelons cela le « coût d’opportunité ». Quand quelqu’un choisit de passer du temps avec vous, ce choix exclut d’autres possibilités. Chaque moment offert est un moment qui n’est pas passé ailleurs.
Un compagnon IA n’est confronté à aucun de ces compromis ; son attention ne coûte rien, n’exclut rien et, par conséquent – pour le dire sans détour – ne signifie rien.
https://www.npr.org/2025/10/08/nx-s1-5561981/ai-students-schools-teachers
https://cdt.org/insights/hand-in-hand-schools-embrace-of-ai-connected-to-increased-risks-to-students
https://sqmagazine.co.uk/character-ai-statistics
https://futurism.com/artificial-intelligence/florida-opens-criminal-probe-chatgpt-school-shooting
A new study argues that certain frontier chatbots are much more likely to inappropriately validate users’ delusional ideas — a result that the study’s authors say represents a “preventable” technological failure that could be curbed by design choices.
https://futurism.com/artificial-intelligence/certain-chatbots-worse-ai-psychosis-study
The results suggest that delusional reinforcement by LLMs reflects a preventable alignment
failure. In demonstrating that these harms can be resisted, the safer models establish a baseline that developers of future systems should now be expected to meet
https://futurism.com/artificial-intelligence/certain-chatbots-worse-ai-psychosis-study
S.3062, the GUARD ACT, would:
- Limit AI companies’ ability to provide companion chatbots to minors, which companies like Google-linked Character.AI are already doing;
- Allow states to enact their own protections;
- Demand that AI companions disclose their non-human status and lack of professional credentials for all users;
- Give the U.S. Attorney General and state Attorneys General the ability to investigate and bring criminal cases against companies violating the law.
https://dispatch.techoversight.org/email/a8c5dcc4-6788-42aa-9130-8c2b454ad5ea
https://www.nytimes.com/2026/04/29/business/meta-instagram-facebook-children.html
Next steps
In exercising their right of defence, Instagram and Facebook now have the possibility to examine the documents in the Commission’s investigation files and reply in writing to the Commission’s preliminary findings. The platforms can take measures to remedy the breaches, in line with the 2025 DSA Guidelines on the protection of minors. In parallel, the European Board for Digital Services will be consulted.
If the Commission’s views are ultimately confirmed, the Commission may issue a non-compliance decision, which can trigger a fine proportionate to the infringement which shall in no case exceed 6% of the total worldwide annual turnover of the provider. The Commission can also impose periodic penalty payments to compel a platform to comply.
These findings do not prejudge the final outcome of the investigation.
https://ec.europa.eu/commission/presscorner/detail/en/ip_26_920
A trial is slated to begin in New Mexico that will test the state’s claims that Facebook, Instagram and WhatsApp harmed young users’ mental health and its bid for a court order forcing the company to make changes.
https://futurism.com/artificial-intelligence/two-shootings-openai-stopped
https://futurism.com/artificial-intelligence/grok-convinces-man-arm-himself-hammer
https://www.theverge.com/policy/922380/new-mexico-meta-public-nuisance-trial-kids-safety
« Plus p… qu’un flamant rose en tutu fluo » : visé par la justice française concernant X, Elon Musk insulte les juges 
Visé par une information judiciaire en France pour complicité de diffusion d’images pédopornographiques sur X, Elon Musk a violemment réagi. Le milliardaire a ciblé les magistrats avec des insultes homophobes en français, marquant une nouvelle escalade avec la justice.
