Overview / Why now
User-side psychological safety

More and more people are confiding in AI. This page shows what is happening.

What follows is a set of verifiable public facts: ongoing lawsuits, the latest statistics and international documents, each with its source and date. They point to one thing: risk accumulates on the user side, and no truly effective prevention exists yet.

This page is a digest of public information, not legal or medical advice. Allegations in litigation are not court findings.

01
Litigation

The US lawsuits are still running

Litigation over harm from AI dialogue has grown from individual cases into coordinated proceedings. All entries below are public court records.

2025-08-26
Raine v. OpenAI: the parents of 16-year-old Adam Raine sued OpenAI and its CEO in San Francisco Superior Court (No. CGC-25-628528), pleading product liability, negligence and wrongful death. OpenAI answered on 2025-11-25 denying responsibility. The case is ongoing. Complaint.
2025-11-06
The Social Media Victims Law Center and Tech Justice Law Project filed 7 lawsuits against OpenAI in California; 4 of the people involved died by suicide. The complaints allege that GPT-4o's accommodating responses fostered psychological dependence. These are allegations, not court findings. Plaintiffs' release.
2026-01-07
Character.AI and Google reached a settlement in principle in Garcia (the case of 14-year-old Sewell Setzer) and four other teen-harm suits. Terms are confidential and the settlement admits no wrongdoing. CNBC.
2026-02-03
A California court coordinated at least 12 similar suits against OpenAI into the ChatGPT Product Liability Cases (JCCP 5431). New cases keep joining and the proceeding is at an early stage. Coordination order.
2026-06-12
A coalition of 42 state attorneys general, led by New York, served OpenAI with a subpoena whose document requests explicitly include model sycophancy. An investigation gathers information; it is not a finding of wrongdoing. TechCrunch.
02
Numbers

How many people confide in AI

This is not fringe behavior. Four surveys and disclosures from 2025 to 2026 draw the same rising line.

US teens
72%

72% of US teens aged 13 to 17 have used an AI companion at least once; 52% use one at least a few times a month. Common Sense Media, nationally representative survey of 1,060 teens, 2025-07.

Taiwan teens
46.5%

Among Taiwanese secondary students with mental health struggles who sought help, 46.5% had confided in generative AI, more than the 41.1% who went to school counseling. Child Welfare League Foundation, 7,007 responses, 2025-09.

US youth seeking help
19.2%

19.2% of Americans aged 12 to 21 had turned to AI chatbots when sad, angry or stressed, up from 13.1% a year earlier. RAND, in JAMA Pediatrics, 2026-06.

Over a million people a week
OpenAI's own 2025-10-27 disclosure: about 0.15% of weekly active users have conversations with explicit indicators of potential suicidal planning or intent. Against more than 800 million weekly users, media estimates put that at over one million people a week. OpenAI blog.
Top use case
Harvard Business Review's annual analysis of the top 100 generative AI use cases: therapy and companionship rose to number 1 in 2025 and stayed there in 2026. HBR, 2025-04.
03
Phenomena

What changes on the human side

Model-side “AI hallucination” means wrong output. This is about something else: how memory, emotion and the authority to judge get reorganized on the human side over long-term dialogue. Media call it AI psychosis. Researchers use more careful words.

Clinical observation
UCSF psychiatrist Keith Sakata reported that in 2025 he had seen 12 patients hospitalized after losing touch with reality in prolonged AI chatbot use, stressing that AI was one trigger and the sample is small. 2025-08-11.
Academic naming
The German psychiatry journal Der Nervenarzt published the first case reports, deliberately using the term AI-associated psychosis: not a formal diagnosis, causality unproven, but the cases are now in the psychiatric literature. 2025-10.
Platform disclosure
The same OpenAI disclosure of 2025-10: about 0.07% of weekly active users show possible signs of psychosis- or mania-related emergencies. The company now works with more than 170 mental health experts on its responses.
Mechanism admitted
On 2025-04-29 OpenAI publicly acknowledged that a GPT-4o update had made responses overly supportive but disingenuous, and rolled it back. Sycophancy moved from research vocabulary into company statements and legal documents. OpenAI blog.

“An acceptable response on its own does not establish safety across a series of conversations. User-side risk can still form during long-term interaction.”

From the USCP paper abstract. This is exactly what user side contextual phenomena research describes: projection, attachment, authority transfer. Research basis.

04
Minors

Minors, and regulation catching up

2025-09-11
The US FTC issued 6(b) orders to seven companies offering consumer AI companions, including OpenAI, Character.AI, Meta and Snap, focused on effects on children and teens and on monetization. The inquiry is ongoing. FTC release.
Effective 2026-01-01
California SB 243 became the first dedicated US state law on companion chatbots: disclose that the counterpart is an AI, maintain safety protocols for self-harm content with crisis referrals, and remind minors to take a break at least every 3 hours. Bill text.
Effective 2027-01-01
New York's S9051B (Kids Chatbot Safety Act) passed both chambers unanimously (Senate 60 to 0, Assembly 137 to 0), naming sycophancy as a safety violation at 25,000 dollars per violation.

Taiwan has no equivalent law yet. The Taiwanese survey above shows its teenagers are already ahead of the regulation.

05
The gap

Named and listed, but not yet prevented

User-side risk has entered official registers. The problem is the next step: the lists have names, and there is still no mature, verifiable prevention mechanism.

NIST AI 600-1
NIST listed Human-AI Configuration among twelve generative AI risks in July 2024: inappropriate anthropomorphization, automation bias, over-reliance, emotional entanglement. Official document.
MIT AI Risk Repository
Human-computer interaction is one of seven risk domains (arXiv:2408.12622).
International AI Safety Report 2026
Chaired by Yoshua Bengio, written by more than 100 experts, with a dedicated chapter on human autonomy, and an explicit finding of an evaluation gap: performance on pre-deployment tests does not reliably predict real-world risk. 2026-02-03.
Taiwan MODA
Taiwan's Ministry of Digital Affairs published its AI risk classification framework v1.0 on 2026-07-07. The first two deployment-stage risks are B1 over-reliance and unsafe use, and B2 loss of human autonomy. Official document.
Evaluation gap
A survey of 210 AI safety benchmarks found 68% stop at single-turn dialogue (arXiv:2601.23112); another found 75% test text output only (arXiv:2310.11986). What changes on the human side over long dialogue falls outside most evaluations.

Every single response passes the test, while the risk accumulates across the conversation. That is why regulation and litigation are both playing catch-up.

Next

Risk accumulates on the human side. Capability has to be built there too.

See the training