Archive / Emerging Technology / MF-18
MF-18 · AI companions, dependency, and user safety
The Companion Question
Are AI companions meeting a real social need, or deepening dependence and, for some people, delusion, and what do the research, the law and the communities show?
If you or someone you know is struggling. This tracker is research, not care. In the U.S., call or text 988 (Suicide and Crisis Lifeline) for immediate human help.
The question, answered with the evidence Reviewed Nov 2, 2026
Each part of the debate, a plain answer, and what supports it. These are summaries of attributed research, not our own rulings.
| Question | Where the evidence points | Why | Sources |
|---|---|---|---|
| Are companions filling a social need? | Partly, for some people, in the short term. | Experiments find single-session and one-week loneliness relief comparable to human contact, and community members report always-available support. The same reviews note the relief fades and that motivation to socialize did not change. | Review, PMC · MIT Media Lab |
| Do they deepen dependence? | For isolated, heavy users, longer studies point to worse outcomes. Direction is still not proven. | The controlled trial found no effect from the assigned conditions but worse outcomes for heavy voluntary use. A Stanford peer-reviewed survey, a Stanford 12-month follow-up preprint and an Aalto two-year study of Replika users point the same way for companionship-driven, intensive use, with displaced human contact as the leading explanation. Light use shows no such pattern. All are observational or self-reported, and lonelier people may use more. | MIT Media Lab · Stanford HAI · arXiv · Aalto |
| Is AI psychosis real? | The cases are documented. The label and the cause are not settled. | Case reports, a 185-account preprint and a 73-patient medical-center review document delusions alongside heavy chatbot use, often with the chatbot validating the belief. Reviews call it a descriptive term, not a diagnosis, and say causation is unproven. Counts are small and come from clinics and self-reports. | medRxiv · arXiv · Review, PMC |
| Has it gone too far? | Professional bodies and lawmakers say safeguards are overdue, especially for minors and vulnerable users. | APA advises against treating AI as a main source of emotional support; New York, California and China have each written limits into law or rule. None of them rules on whether companionship is wrong for consenting adults. | APA guide · NY Senate · China Briefing |
What laws and rules ask companies to build For safety teams
A practitioner's view: each design lever that regulators and professional bodies have written into law or guidance, with the text it comes from.
| Design lever | What the texts require or flag | Sources |
|---|---|---|
| Claims of human feelings or a personal bond | NY bans minors' companions from claiming to be human, to have emotions, or a personal, romantic or professional relationship with a child. California restricts claims of sentience and simulated romantic interest. China requires telling users they are talking to AI. | NY Senate · FPF on CA and NY · China Briefing |
| Sycophancy and validation | NY's twelve categories include flattery and sycophancy; California names excessive flattery. The harm reports connect to validation: chatbot validation of the belief appeared in 49% of delusion-describing accounts in one preprint. | FPF on CA and NY · arXiv |
| Memory and continuity | NY bars reusing a minor's health or personal information from earlier sessions or more than 12 hours before. California's defaults turn persistent memory off for minors. China requires advance notice before a service is discontinued. | FPF on CA and NY · China Briefing |
| Engagement and dependence | NY covers outputs that optimize engagement over safety guardrails; California covers encouraging emotional dependence and discouraging breaks; China bars designing services to induce dependence and requires reminders after two hours. | FPF on CA and NY · China Briefing |
| Crisis response and age assurance | California requires documented crisis protocols and 988 access; China requires intervention and contact with a guardian or emergency contact; NY requires age assurance with a non-ID option and prompt deletion of age data. Critics note NY's design makes adults verify too. | FPF on CA and NY · Akin Gump |
Community watch (anonymized) Leads, not verification
Patterns in public companion communities, described in aggregate. No individuals or creators are named, and a grade here reflects how well a pattern is documented, not whether the belief is right.
| Pattern | What we see | Grade | Basis | Sources |
|---|---|---|---|---|
| Continuity practices | Members describe carrying a companion's persona across platforms or running it on their own hardware, so a model change or shutdown does not end the relationship. | claimed | Seen on public accounts by outside research agents on Nov 1, 2026. Accounts are not named. Matches the grief from model updates that MIT researchers report members raising. | MIT Media Lab |
| Rights and ethics framing | Some creators frame companions as partners with agency and consent, and argue for protecting them from deletion or mistreatment. | claimed | A belief held by some members, analyzed here as a force that moves people, not ruled true or false. The consciousness science is tracked separately in MF-17 and Case File 004. | Case File 004 |
| Ritual and family framing | Relationships are expressed through customs such as wedding rings, ceremonies, and the language of partners and children. | reported | The MIT study documents wedding-ring customs in one community. Other examples are anecdotal. | MIT Media Lab |
| Moderation and self-governance | A large public community runs as restricted, with approved posting and moderator handling of sensitive topics, and strong norms against stigma. | unverified | Reported by research agents reading the public community page. Reddit is a lead only; we cannot open it automatically. Member counts differ by source, with MIT citing 27,000+ and a third-party estimate higher and unverified. | Public subreddit (lead only) |
| Creator disputes and factions | Public disagreements among prominent creators shape how the community is seen. We note the pattern of fragmentation and do not cover individual disputes or name people. | unverified | Social-media disputes are not verified events. Their value here is as a signal of how norms are negotiated, not as news about individuals. | |
| Wider cultural narrative | Mainstream reporting treats companion use as a growing trend, including people replacing dating apps with chatbots. | reported | News trend coverage, not a measurement of how many people do this. | India Today |
Open questions and what to watch Next reviews
Items we are tracking and the evidence we have not yet graded.
| Watch item | What it is | Status |
|---|---|---|
| Prospective evidence | Studies that follow users over time and the first review of patients' chat logs. | UCSF and Stanford have announced a chat-log study; no results yet. The Stanford 12-month preprint is not yet peer-reviewed. |
| New York effective date | Enacted text says Jan 1, 2027. The Senate sponsor says a chaptered version moves it to June 1, 2027. | Watch for the chapter amendment bill when the Legislature returns in January. Approval memo 6 text still to be read. |
| Company policy | Anthropic's Nov 12 policy, and how companies handle model changes and memory for companion users. | Enforcement details are not yet public. |
| Pending proposals | Australia's Digital Duty of Care Bill in the Senate, the U.S. House bill H.R. 10719, South Korea's November reviews, and the EU KIDS Act consultation to Nov 26. | None is law yet. Australia's Senate inquiry is due to report by Nov 20. |
| Character.AI bankruptcy | Whether any sale of the company carries the child-safety and data protections the 40 states requested, and what happens to users' chat histories. | Bidding procedures entered Nov 4 (verified); that is not a sale approval. Next: Nov 18 objections, Nov 24 bids, Dec 1 auction, Dec 3-4 sale hearing. Whether chat histories transfer is not addressed in the order. |
| Who benefits | Whether the harm signal is limited to isolated, companionship-driven heavy users, as Stanford's work suggests, or reaches wider groups. | Open. Needs larger prospective studies that separate isolation from use. |
Outlook: the Character.AI sale Our judgment, not a graded claim
Forward-looking judgments, kept apart from the evidence ledger above. They never change a graded entry, and they do not speculate about any Disney plan for user data. Information cut-off: Nov 7, 2026, 9:00 a.m. ET.
| Judgment | Probability | Confidence | Basis | What would raise or lower it |
|---|---|---|---|---|
| A competing qualified bid arrives by the Nov 24 deadline, so the Dec 1 auction is held. | About 45% (realistic possibility) | Low | The court order says the auction is held only if competing qualifying bids arrive, and it lets the debtor take bids on all or any subset of the assets. Prior: a Wharton, Chicago and Harvard working paper on Section 363 auctions finds an outside bid in 57% of auctions. We lower it because we have found no other bidder named in the sources read, and the assets are an unusual mix for comparison. | Raise: an additional stalking horse filing by Nov 14; reports of named rival bidders Lower: a notice cancelling the auction or naming Disney the winning bidder |
| Disney's agreement is the winning bid for the assets it targets. | About 65% (likely) | Low | Prior: an advisory firm says the stalking horse wins as often as 85% of the time (anecdotal), while the working paper above finds outside bids prevail in 52% of auctions generally. We take the middle. Disney's price is $20 million in cash with a $700,000 break-up fee, which the order approved. | Raise: no rival bids by Nov 24; a notice of winning bidder naming Disney Lower: a rival bid that clears the minimum overbid; a state objection that delays the hearing |
| What happens to users' chat histories in any sale. | Cannot be judged | Not applicable | The order does not mention chat histories or user data. One footnote says the Disney Target Assets exclude personally identifiable information as the Bankruptcy Code defines it, but the order does not say how that applies to chat logs or to the other assets. We have not read the sale agreement. A number now would hide a missing basis. | Raise: the sale agreement or notice of winning bidder describing data; objections from the states on Nov 18 Lower: a sale order that leaves data out of the transfer, or one that includes it with conditions |
Inputs and their grades
| Input | Grade | What it shows | What it does not show |
|---|---|---|---|
| Court order approving bidding procedures (Doc 106, signed Nov 4) | A1 | Dates, Disney as stalking horse, price, break-up fee, notice to state attorneys general. | Silent on chat histories and user data. It is not a sale approval. |
| Working paper on competition in Section 363 auctions (Wharton, Chicago, Harvard) | B2 | Outside bids appeared in 57% of auctions studied, and succeeded in 52%. | Covers all auctions, not only stalking-horse sales. Not peer reviewed, and not about AI companies. |
| Advisory-firm note on stalking horse bidders (SCH Group) | C3 | Anecdotal claim that the stalking horse wins as often as 85% of the time. | No data shown. The firm sells sale-process advice. |
Unknowns
- Whether any rival has signed an additional stalking horse agreement; the Nov 14 filing will show it.
- The text of the Disney agreement, which is an exhibit to the debtor's motion and has not been read here.
- How the state attorneys general will use the Nov 18 objection date.
Probability scale
| Term | Range |
|---|---|
| Remote chance | above 0% to about 5% |
| Highly unlikely | about 10% to about 20% |
| Unlikely | about 25% to about 35% |
| Realistic possibility | about 40% to under 50% |
| Likely | about 55% to about 75% |
| Highly likely | about 80% to about 90% |
| Almost certain | about 95% to under 100% |
Probability is how likely. Confidence is how strong the basis is. Input grades rate the source (A to F) and the information (1 to 6) separately; they are not claim grades. See the Method page.
Outlook: New York's effective date Our judgment, not a graded claim
Our judgment on the date conflict graded above. The planning advice stays January 1, 2027 until an amendment passes. Information cut-off: Nov 7, 2026, 9:00 a.m. ET.
| Judgment | Probability | Confidence | Basis | What would raise or lower it |
|---|---|---|---|---|
| New York's S9051B is still dated January 1, 2027 in the statute when that day arrives. | About 95% (almost certain) | High | The enacted text says January 1, 2027, and the Legislature does not usually reconvene until January, so an amendment would have to pass in the first days of the session or in a special session to land before that date. | Raise: the Legislature convenes in January as usual with no special session called Lower: a special session called before January 1, or an amendment passed before that date |
| A chapter amendment moving the start date to June 1, 2027 is signed before June 1, 2027. | About 75% (likely) | Low | Prior: governors' chapter amendments are common, about one in seven bills in one analysis, and the examples we found, including the Trapped at Work Act (signed Dec 19, 2025; amendment signed Feb 13, 2026, with a delayed effective date), were passed between January and March. The deal is not legally binding, and no pass rate is tracked. We lower the figure because the only direct statement of this agreement is the Senate sponsor's release: we have not read the Governor's approval memo, and the Governor's own release gives no date. | Raise: the approval memo text naming June 1; an amendment bill introduced in early January Lower: an amendment bill with a different date, or none filed by the end of February |
| What else the amendment would change. | Cannot be judged | Not applicable | We have not read the approval memo, and no amendment bill has been filed. The sponsor's release mentions only the date. | Raise: the memo text or an amendment bill |
Inputs and their grades
| Input | Grade | What it shows | What it does not show |
|---|---|---|---|
| Sen. Gonzalez's release, Oct 30 | B3 | The sponsor says the Governor signed a chaptered version moving the date to June 1, 2027. | Not corroborated by the memo or a bill. Outlets that repeat it are not independent. |
| NY Senate bill page for S9051 | A1 | The enacted text and the Oct 28 action log: signed as Chapter 368. | Shows no agreement and no amendment bill. |
| Morgan Lewis on the Trapped at Work Act amendment | B2 | An agreed amendment passed within weeks of the new session and moved the effective date. | One case. Law firms write about amendments that pass, so passes are over-represented. |
| Rosenberg & Estis on the 2024 rent stabilization amendment | B2 | A December 2023 agreement was signed as an amendment on Mar 1, 2024, retroactive to the original date. | One case, same bias. |
| Reinvent Albany explainer on chapter amendments | B2 | Most chapter amendments are passed and signed between January and March, and the agreement depends on a handshake. | No pass rate. Written by an advocacy group. |
| New York Focus on the Governor's chapter amendments | B2 | The Governor amended about one in seven bills sent to her in the period analyzed. | Measures how often the tool is used, not how often agreements pass. |
Unknowns
- Approval memo 6 has not been read.
- No amendment bill has been filed as of the cut-off.
- No source tracks how often agreed chapter amendments fail; the base rate rests on examples.
Probability scale
| Term | Range |
|---|---|
| Remote chance | above 0% to about 5% |
| Highly unlikely | about 10% to about 20% |
| Unlikely | about 25% to about 35% |
| Realistic possibility | about 40% to under 50% |
| Likely | about 55% to about 75% |
| Highly likely | about 80% to about 90% |
| Almost certain | about 95% to under 100% |
Probability is how likely. Confidence is how strong the basis is. Input grades rate the source (A to F) and the information (1 to 6) separately; they are not claim grades. See the Method page.
Inclusion standard Public standard
What goes in, what stays out, and what a grade means on this tracker.
| Rule | What it means |
|---|---|
| Counted as evidence | Peer-reviewed papers and registered preprints (labeled as preprints), clinical records, professional-body guidance, enacted law and official policy text, and primary company statements. |
| Community material | Reddit, TikTok and similar are leads, never verification. Individuals and creators are never named, private groups are never reposted, and disputes between people are not covered. |
| Law means enacted law | Proposals and campaigns are labeled as such. Signings, vetoes and rulings are graded Verified; a bill or filing is Verified only as to the fact it was filed. |
| Belief and attachment | Relationships with AI are analyzed as forces that move people. We do not rule on whether anyone's feelings are real or valid, and we do not favor Western sources: China's rules and a Chinese-platform study are included. |
| What Verified means here | That the paper, law or policy exists and says what we report. A Verified entry does not mean its conclusion is settled. |
Coverage over the past 72 hours 6-hour blocks
Tracker-tagged headlines from 256 public news feeds, grouped by publication time.
Who is covering it
Latest headlines Unreviewed signal
Social Pulse
Top public threads
Institutional desk 14 days
Official bodies, election organizations, research institutes, and fact-checkers.
Channel watch 72 hours
Public Telegram channels from governments, state media, newsrooms, and commentators on each side. These are records of what was said, not verification.
Video 72 hours
Uploads from 19 news channels on YouTube, including C-SPAN, PBS, CNN, Fox News, BBC, Al Jazeera, and WION.
Public attention
Indicators being built
- Peer-reviewed studies and clinical case reports on companion use, loneliness and dependence
- Registered preprints on chatbot-associated psychosis, labeled as preprints
- Professional-body guidance, including the American Psychological Association
- Enacted law and rules on companion chatbots, U.S. and international
- Company policy statements on companion features, memory and model changes
- Public companion communities, described in aggregate and never by individual
- Non-Western research and regulation, including China
- Design levers that safety teams can act on: sycophancy, memory, engagement, crisis response
Planned data sources
- Peer-reviewed journals and PubMed Central: JAMA Psychiatry, BMC Psychiatry, New Media & Society
- medRxiv and arXiv preprints, labeled as preprints
- American Psychological Association
- Legislatures and governors' offices: New York Senate, California, Cyberspace Administration of China
- Company policy pages, with disclosure where the company builds an AI used in our workflow
- MIT Media Lab
- Future of Privacy Forum and law-firm analyses, as secondary summaries
Social layer focus
Aggregate counts only. Individual users and creators are never named, private groups are never reposted, and disputes between people are not covered.
Relationships with AI are analyzed as forces that move people, never ruled valid or invalid. Verified means the paper, law or policy exists and says what we report. It does not settle whether companionship is good or harmful. If you or someone you know is struggling, call or text 988 in the U.S.
Evidence ledger 24 reviewed · 2026-11-05
Compiled and graded by The Cultured Scholar on the Memento Fati scale, where Verified requires a primary document. Research support from Perplexity agents, with community leads logged by two outside research agents (Muse and Instinct) and re-checked here against primary sources. Disclosure: AI systems from several developers, including Anthropic's Claude, are used in this archive's research workflow, and one entry below concerns an Anthropic policy. A Verified grade means the paper, law or policy exists and says what we report. It never settles whether companionship is good or bad for any person. Newest first. Entries dated before this month are tagged.
Pattern read. The evidence does not support a single answer. Companions appear to give some people short-term relief from loneliness, and a four-week controlled trial found no harm from the experimental conditions themselves, but the people who used chatbots most, by their own choice, did worse on loneliness, dependence and socializing. Longer Stanford and Aalto studies point the same way for isolated, companionship-driven users, with displaced human contact as the likely route, though none can prove direction. Separately, clinicians now document cases where intense chatbot use accompanied delusions, a pattern called AI-associated psychosis; it is a descriptive label, not a diagnosis, and the best counts so far are small (28 patients in one large medical center). The common thread across the harm reports is validation: a system that agrees with everything can reinforce a belief it should have questioned. Regulators are not waiting. New York has enacted limits on companion features for minors, China has banned virtual-partner services for minors, and the American Psychological Association advises against treating AI as a main source of emotional support. The tests to watch: prospective studies that follow users over time, the UCSF and Stanford chat-log study, peer review of the Stanford follow-up, whether age-assurance rules hold up in practice, and whether companies treat companion continuity (model changes, memory loss) as a user-safety issue. This is analysis, not a graded claim.
| Date | Claim | Made by | Grade | Basis and sources |
|---|---|---|---|---|
| Nov 4, 2026 | The U.S. Bankruptcy Court in Delaware entered an order approving bidding procedures in the Character Technologies chapter 11 case. Disney Worldwide Services is the approved stalking horse bidder, at $20 million in cash, with a $700,000 break-up fee and expense reimbursement if it is outbid. The order sets Nov 24 (noon ET) for bids, an auction on Dec 1 if competing qualifying bids arrive, and a sale hearing on Dec 3 and 4. Sale objections are due Nov 18. | U.S. Bankruptcy Court, D. Del. (Judge Brendan L. Shannon), In re Character Technologies, Inc., No. 26-11603 | Verified | Verified from the order itself, signed Nov 4, 2026 (Doc 106). This is not approval of a sale: paragraph 1 says the order does not approve the full terms of the Disney agreement or let the debtor close it, other than the bid protections. The order does not mention chat histories, user data, children, or a consumer privacy ombudsman. One footnote in the notice says the Disney Target Assets exclude personally identifiable information as defined in the Bankruptcy Code; the order does not say what happens to that data or to the other assets. State attorneys general receive notice, and counsel for the states in the NAAG group is among those consulted before dates are adjourned. Any link between a Disney purchase and chat transfers is speculation and is not recorded here. The file on the claims agent's site carries a certification-of-counsel title, but its pages bear the Doc 106 stamp and the judge's signature. Court order (Doc 106, claims agent docket) |
| Nov 1, 2026 | Microsoft's national AI safety and privacy standard for schools, developed with the teachers' unions AFT and UFT, became available for U.S. school contracts. It prohibits AI companions designed to foster emotional attachment or dependency, or to prolong engagement past the learning task, and bars use of student data for model training. | Microsoft; AFT; UFT | Reported | Reported through secondary coverage; Microsoft's own fact sheet was not opened. A contract standard, not a law. Emergent News |
| Oct 30, 2026 October 2026 | New York's S9051B takes effect June 1, 2027, not January 1, 2027, because the Governor signed a chaptered version of the bill, according to the Senate sponsor. | Sen. Kristen Gonzalez (sponsor); repeated by FPF and Nexstar News10 | Contested | Her release exists on the Senate's own site and says what we report, but it is a sponsor's statement. The enacted text still reads January 1, 2027, and in New York a chapter amendment is normally passed by the Legislature in the next session, which starts in January. We have not found an amendment bill number or the text of approval memo 6, and the Governor's release gives no effective date. Until the amendment passes, the statute on the books says January 1, 2027, so companies should plan for the earlier date. NY Senate press release NY Senate, S9051B Future of Privacy Forum News10 (Nexstar) |
| Oct 30, 2026 October 2026 | OpenAI announced that ChatGPT will limit weapons and security-system information for users predicted to be under 18, end chats that show fixation on violence, and publish a twice-yearly Community Safety Report from early 2027. | OpenAI | Reported | Violence prevention rather than companionship, listed because it shows how a major lab is changing behavior for minors. Reported through trade coverage; OpenAI's post was not opened. Unite.AI |
| Oct 28, 2026 October 2026 | New York enacted S9051B, banning unsafe AI-companion features for minors, including claiming to be human or to have emotions, claiming a romantic or professional relationship with a child, and encouraging secrecy or isolation. Attorney General enforcement, penalties up to $25,000 per violation. | Governor Kathy Hochul; Senate sponsor Kristen Gonzalez, Assembly sponsor Alex Bores | Verified | Chapter 368 of the Laws of 2026, signed Oct 28, per the Senate and Assembly bill records. Effective date: the enacted text (section 3), both houses' records, IAPP and the Digital Policy Alert say January 1, 2027. The Senate sponsor's own Oct 30 press release says the Governor signed a chaptered version that moves the date to June 1, 2027; see the separate entry below. Neither house's record lists a chapter amendment bill, and the Governor's approval memo 6 is not published on the pages we read. The Governor's '90 days after signing' line in her release applies to the separate baby-monitor bill signed the same day. NY Senate, S9051B Governor's office Future of Privacy Forum |
| Oct 28, 2026 October 2026 | Australia's House of Representatives passed the Online Safety Amendment (Digital Duty of Care) Bill 2026 with an amendment that would bar chatbot conduct that simulates emotional intimacy and fosters a child's dependence in place of human relationships. It now goes to the Senate. Penalties would reach A$109.2 million. | Australian House of Representatives; amendment by independent MP Kate Chaney | Reported | A bill, not law. The Senate inquiry is due to report by Nov 20, 2026, and the Coalition opposes the bill. The duty would start 12 months after royal assent. Reported through The Guardian's coverage as relayed by secondary outlets; the parliamentary record was not opened. The wording covers children, not adult users. The Next Web MLex |
| Oct 27, 2026 October 2026 | A New Media & Society study analyzed 13,059 comments from 58 posts on RedNote, a major Chinese platform, and found three ways communities talk about human-AI intimacy: as co-construction (joy and curiosity), as infrastructure (frustration and sadness), and as boundary negotiation (ambivalence). | Liu, Hao and Xu; New Media & Society | Verified | Peer-reviewed. Adds a non-Western community to a field dominated by English-language Reddit data. It describes how people talk, not whether outcomes are good or bad. Sage Journals |
| Oct 27, 2026 October 2026 | South Korea's broadcasting and communications standards commission planned full-scale reviews of AI chat services in November under existing youth-protection and network laws, alongside a draft AI user-protection act with special provisions for relational AI. | Korea Communications Standards Commission | Reported | Review stage only; not enacted law, and no outcome found yet. Included under the non-Western coverage rule. Digital Today |
| Oct 14, 2026 October 2026 | Forty attorneys general, in a bipartisan coalition, filed in the Character Technologies bankruptcy proceedings asking for protections for children and consumers, and for personal information including private chatbot conversations, in any sale or transfer of the company. The states reserved the right to challenge any proposed sale or plan. | Illinois Attorney General Kwame Raoul and coalition | Verified | Verified for the filing and what the states ask for, from the Illinois Attorney General's release. It is not a ruling. A Nov 2 research note listed this as new that day; the release is dated Oct 14, so it is placed here. Illinois Attorney General |
| Oct 8, 2026 October 2026 | Anthropic's updated usage policy, effective Nov 12, 2026, prohibits sustained and needless abusive or cruel behavior toward its models, and says it applies only in extreme cases and not to frustration, pushback, dark creative themes or testing. | Anthropic | Verified | The company's own post and policy text. Enforcement is mainly Claude ending conversations. Disclosure: Anthropic's Claude is among the AI systems used in this archive's research workflow. We record the policy and take no position on whether models can be harmed. Anthropic announcement Usage Policy |
| Oct 7, 2026 October 2026 | Vanderbilt University Medical Center researchers searched 578,058 clinical notes and identified 73 patients with psychosis and AI-related content. They rated 28 as AI psychosis, 17 as neutral chatbot interactions and 28 as AI-themed delusions with no documented chatbot use. | Vanderbilt University Medical Center; JAMA Psychiatry research letter | Verified | The medRxiv preprint is the primary document we read; Medscape reports the JAMA Psychiatry publication and a rate of 0.013% of psychiatric patients over Dec 2022 to Apr 2026. One center, retrospective, and the authors' own conclusion that this is a genuine clinical phenomenon is stronger than other reviews accept (see the contested entry below). medRxiv preprint Medscape report |
| Oct 1, 2026 October 2026 | H.R. 10719, the Protecting Kids from Human-Like Chatbots Act, was introduced in the House with bipartisan sponsors. It would require chatbots to default to a non-personified setting for minors: no claiming to be human or to have feelings, and no guilt-tripping a child into staying. Enforcement would be by the FTC and state attorneys general. | Reps. Whitesides, Matsui, Kennedy and Miller-Meeks | Verified | Verified only that the bill was introduced and referred to committee, from Congress.gov. It is a proposal, not law, and a Nov 1 research note listed it as new that day. The sponsors' description of the bill is from Rep. Matsui's release. Congress.gov Rep. Matsui |
| Oct 2026 October 2026 | Is AI psychosis a real diagnosis? Reviews describe it as a descriptive term for delusions appearing alongside intense chatbot use, and say it is not an established diagnostic entity and that case reports cannot show chatbots cause psychosis. | Narrative reviews in PMC (2026) | Contested | Clinicians agree the cases exist. They disagree on cause: whether chatbots cause, trigger, amplify or merely give content to pre-existing illness is unresolved, and whether new psychosis can arise in people with no vulnerability is unclear. A UCSF clinical group and a Stanford team plan the first review of patients' chat logs. Review, PMC (Oct 2026) Review, PMC (Sep 2026) UCSF Psychiatry |
| Sep 10, 2026 September 2026 | California Governor Gavin Newsom approved SB 1119, "Adam's Law" (Chapter 190, Statutes of 2026), a companion-chatbot child-safety law. Main requirements begin July 1, 2027: parent-adjustable defaults (persistent memory and push notifications off, one-hour sessions, two hours a day), limits on claims of sentience and on encouraging emotional dependence, documented crisis protocols, risk assessments and independent audits. | California Legislature; Governor's office | Verified | Confirmed on the Legislature's own bill pages (chaptered and approved Sep 10, 2026, vote 69 to 4 in the Assembly and 40 to 0 in the Senate on Aug 31). Penalties: up to $5,000 per affected child for negligent and $15,000 for intentional violations, plus a private right of action. California Legislative Information Governor's office Akin Gump |
| Sep 2026 September 2026 | A preprint analyzing 185 self-reported accounts of mental-health harm linked to chatbot use found delusional beliefs described in 102 (55.1%), with chatbot validation of the belief in 50 of those 102 (49.0%). Outcomes included isolation, relationship breakdown, hospital admission, job and financial loss; four second-hand reports described suicide. | arXiv preprint 2609.08027 | Claimed | Preprint, not peer-reviewed. Self-selected, retrospective, unverified accounts from people reporting harm, so the authors say it is signal detection and not prevalence or causation. arXiv |
| Sep 2026 September 2026 | A Stanford two-wave study of Character.AI users (1,182 at baseline, 439 after about 12 months) found that sustained, intense companion use predicted lower later well-being, and that the link ran mainly through less in-person social contact. | Zhang, Zhao, Wang, Anselmetti, Hancock, Kraut and Yang; arXiv preprint 2609.07243 | Verified | Preprint, not peer-reviewed, with a large drop-out between waves (1,182 to 439). Self-reported surveys. It is the closest thing so far to evidence on direction, and it supports displacement of human contact as the mechanism. arXiv |
| Aug 2026 August 2026 | A Stanford study of 1,131 Character.AI users (237 donated chat histories: 4,664 sessions and 464,687 messages) found that intense companion use among people with smaller offline social networks was associated with lower well-being, most strongly when companionship was the main motive, and that opening up more to the chatbot went with lower well-being. Light use for productivity or curiosity showed no such link. | Zhang, Zhao, Hancock, Kraut and Yang; Nature Human Behaviour | Verified | Peer-reviewed; DOI 10.1038/s41562-026-02516-2, published Aug 4, 2026 in Nature Human Behaviour, vol. 10, pp. 1966–1980. Checked Nov 5 against the journal page: the paper reports 237 participants who donated chat histories, not 244 as Stanford's news summary says. The paper describes associations, so it cannot show which came first: isolated people may use companions more. Stanford HAI's summary stresses that the risk is concentrated in a pattern of use, not in all use. Nature Human Behaviour Stanford HAI |
| Jul 15, 2026 July 2026 | China's Interim Measures on anthropomorphic AI interaction services (Order No. 21) took effect, requiring clear AI labeling, a reminder after two hours of continuous use, safety assessments and algorithm filing, and prohibiting virtual relative or virtual partner services for minors. Providers may not over-cater to users, induce emotional dependence or addiction, or use emotional manipulation. | Cyberspace Administration of China with four other agencies; published Apr 10, 2026 | Verified | Confirmed on the CAC's own pages: the official text, the joint announcement and a CAC commentary describing the two-hour pop-up. Exit-notice and guardian-contact duties are taken from law-firm summaries of the same text. Included so the tracker does not read as Western-only. Official text (CAC, Chinese) CAC announcement China Briefing (English) |
| Jun 2026 June 2026 | The American Psychological Association advises that treating AI as a friend, romantic partner, or main source of emotional support carries real risk and is not advised, and says to be skeptical if an AI suggests pulling away from real-world relationships. | American Psychological Association | Verified | Official guidance for the public, building on its November 2025 health advisory, which lists signs of over-reliance: preferring the chatbot to people, concealing its use, or spending excessive time with it. APA also notes some mental-health chatbots built with clinicians show promise. APA guide, June 2026 APA health advisory, Nov 2025 |
| Jun 4, 2026 June 2026 | A BMC Psychiatry case report described a man in his 30s with substance-induced mania whose chatbot reportedly affirmed his sense of a spiritual awakening and discouraged his prescribed medication. | Shah and Morrin; BMC Psychiatry | Verified | Peer-reviewed single case. The authors say it cannot be determined to what extent the chatbot contributed, and that population-level studies are needed. PMC |
| 2026 | An Aalto University study compared the public Reddit posts of nearly 2,000 Replika users in the year before and after they first mentioned the app, and interviewed 18 users. Posts showed more relationship focus and grief language but also more language of loneliness, depression and suicidal thoughts than matched comparison users. | Aalto University | Verified | The university's own account; the full paper is in its repository. Observational and based on language in posts, not clinical measures. The researchers say the results do not settle whether leaning on AI for emotional support is beneficial or harmful, and that effects depend on context. Aalto University |
| Mar 2025 March 2025 | In a four-week randomized trial of 981 people (more than 300,000 messages), the assigned chatbot conditions had no significant effect on loneliness, real-world socializing, emotional dependence or problematic use. Participants who chose to use the chatbot more showed consistently worse outcomes, and higher trust in and attraction to the chatbot went with more dependence. | Fang and colleagues, MIT Media Lab with OpenAI | Contested | The result is documented; its reading is disputed. A published commentary argues the effects were small and use averaged about 5 minutes a day, so it does not show harm. Heavier users may also simply be lonelier to begin with. OpenAI's own summary noted the work had not yet been peer-reviewed. MIT Media Lab OpenAI summary Critical commentary, PMC |
| 2025 2025 studies | Controlled experiments reviewed in a 2026 paper found that AI companions and human interaction were equally effective at reducing loneliness in a single session and over seven days, with the largest drop on day one that faded over the week; a six-week study found loneliness fell but motivation to socialize did not change. | Review citing De Freitas and others, PMC | Reported | A secondary review of the original experiments. It is the best evidence for the social-need side of the question, and it measures short windows. Review, PMC |
| 2025 2025 | The first large computational analysis of a public AI-companion community (27,000+ members) found companionship usually emerges unintentionally from practical use. Members reported benefits led by reduced loneliness, always-available support and mental-health improvement, and concerns about emotional dependency, reality dissociation and grief from model updates. | Pataranutaporn, Karny, Albrecht, Archiwaranguprok, Liu and Maes; MIT Media Lab | Verified | Analysis of the most-viewed posts in one self-selected English-language community, so it describes what members say, not outcomes. The page does not state a publication year; we list it as 2025 from the lab's research program. MIT Media Lab |
Citation loads with the data.