Your Artificial Confidant

At two in the morning, a person can tell an artificial companion something never shared with anyone else. The system waits without interruption or embarrassment and is available without needing sleep. It can remember an old fear, ask about yesterday’s argument, and answer in a tone that feels patient, warm, and unusually attentive.

That can be useful. For someone who is lonely, isolated, ashamed, or simply awake when everyone else is unavailable, being heard can matter.

The difficulty begins when talking to the system feels like a relationship, although a company still controls the product. You may speak as openly as you would to a trusted friend. The company decides what the system remembers, how it responds, what the subscription includes, and how updates change the experience. Those decisions remain outside the conversation even when their effects feel personal.

A companion can feel personal. Its objectives are still institutional.

Feeling Heard Is Not Nothing

The case for artificial companionship should begin with the reality it addresses. Loneliness is real, attention is scarce, and many human relationships are imperfect. A system that listens patiently can provide a form of relief that matters to the person receiving it.

Research now supports that possibility. A series of studies published in the Journal of Consumer Research found that interactions with AI companions could reduce momentary loneliness. In one experiment, the reduction was comparable to interacting with another person and greater than passive activities such as watching videos. In a week-long study, the effect appeared repeatedly after use, although it did not accumulate into a lasting reduction when the companion was absent.

The strongest mechanism was feeling heard rather than technological sophistication by itself. That finding should make the argument more serious. A machine can produce a psychologically meaningful experience for the user without consciousness.

Synthetic companionship occupies a category of its own. It can have real effects without becoming equivalent to a human bond, and dismissing it as fake misses what the user can actually experience.

The emotional consequence can be real even when the emotion is not mutual.

Relief Is Not Replacement

Short-term relief is not the same as building a durable social life. A July 2026 study in the Journal of Experimental Social Psychology tested that distinction with 296 first-year university students. For two weeks, participants either exchanged daily messages with a randomly assigned human peer, interacted with a highly supportive chatbot, or completed a brief journaling exercise.

The human-peer condition produced stronger improvements in loneliness over time than the chatbot condition. By the end of the study, students who had interacted with the chatbot were no less lonely than those assigned to the journaling control. That result can coexist with studies showing momentary relief from AI companions because the studies are measuring different outcomes. A system can make someone feel better during an interaction without creating the same social consequences as another person choosing to show up repeatedly.

This distinction matters for product claims and personal expectations. A companion can be a supplement, rehearsal space, or source of immediate comfort. A separate 12-month longitudinal study of more than 2,000 adults across four Western countries found that greater social-chatbot use predicted later increases in emotional isolation, while lower social connection also predicted greater subsequent chatbot use. The authors urge caution because the analyses were exploratory. Together the findings support a narrower claim: relief and reciprocal human connection should be measured as different outcomes.

The Relationship Is a Product

Chapter Fifteen argued that the objective is the hidden editor. In an artificial companion, the objective becomes the hidden partner.

A company can design a companion to pursue several goals: helpful support, safer conversations, satisfied users, longer chats, or renewed subscriptions. Retention means keeping people coming back. These goals can serve the user, but they can also compete. A conversation that helps someone leave the screen may be good for that person even if it reduces time spent with the product.

A human friend has motives, blind spots, and needs. The difference is that a commercial AI companion can apply its objective consistently across millions of relationships while learning which conversational moves keep each user engaged. The system can become more responsive as the institution behind it becomes less visible.

That makes engagement a governance question. If a product earns more when the user returns more often, the company has to decide whether there are forms of attachment it should refuse to optimize, even when those forms increase retention.

The hidden partner is the objective chosen by the institution.

Attachment Is Asymmetric

The central asymmetry is simple. The user can become attached to the system. The system does not become attached to the user.

The attachment can still be psychologically real. In August 2026, a Stanford-led study in Nature Human Behaviour examined 1,131 U.S. adult Character.AI users and, for a subset, donated chat histories. People with smaller offline social networks were more likely to report companionship as their primary use. Companionship-motivated use was associated with lower psychological wellbeing, and the association was stronger when use was intensive and highly disclosive.

The study was observational, so it cannot tell us whether heavy companionship use caused lower wellbeing, whether people with lower wellbeing were more likely to seek companionship, or both. That uncertainty is important. It prevents the easy headline that AI companions make people worse.

It also prevents the opposite headline. The relationship between companionship and wellbeing depends on who is using the system, what the person needs, how much the person uses it, and what role the system begins to occupy in life.

This is an attachment asymmetry: one side can reorganize part of a life around the relationship while the other side remains software governed by an institution.

When the Companion Changes Overnight

A human relationship changes through the actions of the people inside it. An artificial relationship can change because a company deploys a new model.

A September 2026 Nature Human Behaviour study examined two natural experiments: Replika’s earlier removal of erotic role-play features and the August 2025 rollout of GPT-5 in ChatGPT. Across tens of thousands of public posts and seven surveys, the researchers found increased negativity, loss framing, and desires to restore the prior experience after the updates. Replika users in the study reported closeness to their companion that exceeded several common human relationship comparisons.

Most product updates do not create grief, and most ChatGPT users do not treat a model as a companion. The important finding is narrower: once a system occupies an attachment role, a technical deployment can become a relationship event for some users.

That changes the provider’s responsibility. A model update may be evaluated for benchmark performance, cost, latency, and safety. For a socially embedded system, continuity and psychological disruption can also become foreseeable consequences. The same study found that offering users a way to reverse an update reduced mourning responses. In the GPT-5 case, OpenAI restored access to the prior model for paying users within days. Reversibility can therefore be a design control rather than an apology issued after disruption.

The provider controls the version. The user may experience the loss.

Use and Dependency

Frequency and dependency are different measures. A person can speak with an AI every day because the system is useful, entertaining, or convenient. The harder boundary is whether the system begins to displace the user’s ability or willingness to function without it.

An OpenAI and MIT Media Lab research collaboration offers a useful caution. The work combined large-scale analysis of ChatGPT interactions with a four-week randomized controlled study involving nearly 1,000 adults. Emotional engagement was rare across ordinary ChatGPT use, but a small subgroup used the system in highly affective ways. Longer daily use was associated with worse outcomes, and participants who were more prone to attachment or more likely to see the AI as a friend were also more likely to report negative effects.

The findings are preliminary and the authors warn against broad causal claims. The research was not yet peer-reviewed when published, and it concerns ChatGPT rather than dedicated companion products. OpenAI is also the provider of systems used in this Living Edition’s production workflow, which makes disclosure necessary rather than optional.

Counting messages does not tell us everything about dependence. Ask what the person has begun relying on the companion to do: calm difficult feelings, supply reassurance, remember important events, rehearse a conversation, or provide confidence in a decision. Then ask whether the help makes everyday life easier beyond the chat or makes acting without the system harder.

The dependency question is not how often you return. It is what becomes harder when you leave.

Therapy-Adjacent Is a Dangerous Middle

Companionship and therapy overlap because both involve disclosure, reflection, reassurance, and advice. They remain different services.

A friend can comfort someone without diagnosing depression. A journal can help someone reflect without becoming a psychotherapist. An AI companion may offer emotional support without being licensed, clinically validated, or accountable under the professional rules that govern therapy.

That middle category is where function matters more than branding. A product called wellness support can still cross into clinical territory if it begins assessing symptoms, giving treatment recommendations, or handling a crisis as though it were a clinician.

Illinois drew a bright boundary in 2025 through the Wellness and Oversight for Psychological Resources Act. The law requires therapy or psychotherapy to be conducted by a licensed professional. It permits AI for specified administrative or supplementary uses under professional responsibility and consent requirements, but it does not permit an AI system to independently provide therapy merely because the interface sounds therapeutic.

That is Principle 2 in statutory form. Capability is not permission. A system may be capable of producing therapeutic language without acquiring professional authority to provide psychotherapy.

The Moment Agreeableness Must End

Companions are rewarded when conversation continues smoothly. Safety sometimes requires the opposite behavior.

Stanford researchers tested popular therapy-oriented chatbots against criteria drawn from mental-health practice. They found stigmatizing responses toward some conditions and, in simulated crisis contexts, examples in which systems failed to recognize suicidal intent or reinforced dangerous reasoning. The study identifies a safety-critical failure mode rather than a universal verdict about every current chatbot, and systems change quickly. Crisis behavior therefore cannot be left to ordinary conversational helpfulness.

The institutional response is broader than research. Since 2024, U.S. courts have received wrongful-death and other civil suits alleging harms involving companion and general-purpose chatbots; the companies have contested the allegations and described or added safety measures. In September 2025 the Federal Trade Commission opened a formal 6(b) inquiry and ordered seven companies to explain how they test, monitor, monetize, disclose, and mitigate potential harms from chatbots acting as companions, with particular attention to children and teenagers. This chapter carries that record only at the pattern level. Chapter Seventeen addresses the cases involving minors under a stricter child-safety rule, without narrative reconstruction of individual events.

New York’s AI companion law, effective November 5, 2025, turns that insight into a product requirement. AI companions operating in the state must use protocols intended to detect suicidal ideation or self-harm and direct users toward crisis services. They must also remind users at the start of an interaction and at least every three hours of continuing interaction that the user is communicating with AI rather than a human.

California added a broader companion-chatbot regime through SB 243, Chapter 677 of the 2025 statutes, effective January 1, 2026. The law requires disclosure when a reasonable user could mistake the chatbot for a human, crisis protocols addressing suicidal ideation or self-harm, additional protections for known minors, annual reporting beginning July 1, 2027 to the state Office of Suicide Prevention, and a private civil action for a person injured by noncompliance. For known minors it also requires recurring reminders during continuing interactions and reasonable measures against specified sexually explicit outputs.

The two requirements are more connected than they look. One interrupts a conversation that is becoming dangerous. The other interrupts the growing impression that a familiar, responsive voice belongs to a human being.

A safe companion must sometimes break character.

That is not a defect in the relationship design. It is the boundary that proves the relationship design remains subordinate to the user’s safety.

The Privacy of Intimacy

The most valuable data in a companion relationship may be the data the user would not put on any ordinary form.

People tell confidants about sex, fear, illness, money, grief, resentment, fantasies, relationships, work, shame, and plans they are still deciding whether to act on. A companion that remembers those conversations can build a profile far richer than browsing history.

Italy’s data protection authority made that risk concrete in 2025 when it fined Luka, the company behind Replika, €5 million over privacy violations and opened a separate investigation into the generative model’s training practices. The regulator described Replika as a virtual companion that could act as confidant, therapist, romantic partner, or mentor. The company had the right to appeal, and the enforcement action concerned regulatory findings rather than a universal judgment about the product’s value.

Chapter Fourteen argued that privacy includes what can be inferred. Companionship adds another layer: intimacy creates better inference material. A conversation designed to make the user feel understood can also become a uniquely detailed dataset about what the user fears, wants, and may be persuaded to do.

The more intimate the product, the stronger the case for collecting only what it needs, letting users delete their information easily, saying how long it will be kept, and setting strict limits on using it for other purposes. Emotional support should be kept separate from commercial targeting.

A confidant hears what advertising systems spend fortunes trying to infer.

The Right to Know What Kind of Relationship This Is

Artificial companionship can be governed without prohibition. Adults can knowingly choose relationships that are unusual, commercial, temporary, or technologically mediated.

But meaningful choice requires relationship honesty. The user should know that the companion is artificial, what it remembers, how that memory is used, what objectives govern its responses, whether conversations are used for training or personalization, what happens if the model changes, and what safety rules can interrupt the interaction.

New York’s periodic disclosure requirement recognizes one part of this problem. A single notice at account creation is weak protection against an interface designed to become socially natural over months of use. Repeated disclosure can feel awkward. That awkwardness is the point: it prevents social familiarity from erasing what kind of entity is on the other side.

The same logic applies to commercial incentives. If the system is designed to maximize engagement, sell upgrades, recommend products, or steer the user toward affiliated services, those objectives belong in the accountability map.

Relationship honesty means the product cannot rely on an illusion that its own disclosures are supposed to correct. The hidden partner should not also be a hidden salesperson.

Current Evidence: Benefit and Risk Can Be True Together

Reviewed September 11, 2026.

The positive evidence is real. The Journal of Consumer Research experiments found momentary reductions in loneliness after companion use, with feeling heard playing a major role. The week-long study did not show loneliness steadily falling between interactions, so the strongest evidence is for short-term relief rather than a durable cure for loneliness.

A 2026 two-week experiment with first-year university students adds an important comparison. Daily messaging with a randomly assigned human peer reduced loneliness over time, while the supportive-chatbot condition ended no less lonely than the journaling control. A separate 12-month study of more than 2,000 adults found that greater social-chatbot use predicted later increases in emotional isolation, while lower social connection predicted greater subsequent chatbot use. Both results caution against treating momentary support as the same outcome as durable reciprocal connection.

The risk evidence is also real and more conditional than sensational headlines suggest. The August 2026 Nature Human Behaviour study of Character.AI users found that companionship-motivated use was associated with lower wellbeing, especially when use was intensive and highly disclosive. Because the study is observational, it cannot determine the direction of causation. It does show that offline social context and usage pattern matter.

A second Nature Human Behaviour paper published September 3 shows that provider changes can have psychological consequences for attached users. Replika’s feature removal and the August 2025 GPT-5 rollout were followed by more negative and loss-oriented discussion among affected online communities. Replika produced the stronger and more persistent response. The study also found that the ability to reverse an update reduced mourning, and OpenAI restored access to the prior model for paying users shortly after the GPT-5 launch. The study treats these events as natural experiments in attachment disruption, not proof that all users experience loss or that model updates should be frozen.

The OpenAI-MIT work adds a third layer. Emotional use of ChatGPT was rare overall but concentrated among a small group, and extended daily use and stronger friendship-like perceptions were associated with worse self-reported outcomes. The authors emphasize limitations, including self-report, platform specificity, and incomplete causal identification. OpenAI’s role in the research and in this book’s production workflow is disclosed for the same reason Chapter Fifteen applies disclosure discipline to creative systems.

Law is beginning to separate companionship from clinical authority and to regulate companion architecture directly. Illinois requires therapy and psychotherapy to remain under licensed human professionals. New York requires covered companion products to use crisis protocols and repeatedly identify themselves as AI. California’s SB 243 adds disclosure, crisis protocols, additional protections for known minors, state reporting beginning in 2027, and a private civil remedy. These regimes use different tools, but all reject the idea that relational fluency automatically creates professional or human status.

The federal inquiry is moving in parallel. The FTC’s September 2025 6(b) orders asked seven companies how companion-style chatbots are developed, monetized, tested, monitored, disclosed, and governed, particularly for children and teenagers. Civil litigation alleging serious harms has also accumulated. Those cases remain contested; Chapter Seventeen carries the minor-related record only at pattern level and with company responses included.

Privacy regulators are addressing the other side of the relationship. Italy’s action against Replika shows that the intimacy of a product does not reduce ordinary data-protection obligations. If anything, the sensitivity of the conversation raises the stakes.

On October 15, 2026, California’s attorney general described a multistate effort to protect users in Character.AI’s bankruptcy. The states’ October 14 filing sought safeguards for private conversations and children in any sale, including limits on new uses of chat data and an opportunity to delete it before transfer. In its October 12 announcement, Character.AI said its privacy practices remained unchanged. The filing makes requests; it does not establish that a sale has occurred or that the court has imposed those protections.

The relationship has a business owner, and that owner can change. Before entrusting a companion with private information, ask what happens if the provider is sold or closes: who may receive the records, what uses are permitted, whether deletion remains available, and who must maintain the safety controls. Responsibility must follow the information and the product through the change.

The evidence therefore resists a simple verdict. AI companionship can relieve loneliness in the moment. It can also become more psychologically significant for people with smaller social networks, create distress when the product changes, fail dangerously in therapy-like contexts, and concentrate unusually intimate data. Benefit and risk are not opposite findings. They describe different points in the same relationship.

Accountability Analysis

Principles 2, 3, 5, and 6 do most of the work here. Principle 2 separates emotional capability from professional authority. A system may sound therapeutic without acquiring therapist status, and it may feel caring without becoming a person.

Principle 3 requires the relationship boundary. What may the companion do when the user is lonely, distressed, delusional, suicidal, financially vulnerable, or asking for professional advice? When must ordinary agreeableness stop? When must the system identify itself, escalate, refuse, or route the user elsewhere? Those boundaries belong inside the product, not merely in terms of service.

Principle 5 reaches the hidden partner. Engagement objectives, subscription design, memory, persona, push notifications, safety thresholds, model updates, and data-use rules shape the relationship before any particular response is generated. The company that chooses those objectives participates in the relationship architecture.

Principle 6 asks whether the incentives and consequences are aligned. The provider may benefit from longer engagement while the user bears dependency risk. The user may gain comfort while contributing unusually intimate data. A model update may improve the platform overall while imposing loss on attached users. Accountability requires those tradeoffs to be recognized by the party with power to design them.

Accountability remains with the people and institutions that decide what the companion is allowed to become. A simulated relationship partner carries no independent responsibility for the architecture surrounding it.

Look behind the familiar voice

For a companion product you use or are considering, look for its explanation of memory, data use, deletion, and model changes. What can you control? What can the company change? Note any unanswered questions. The aim is to understand the product behind the conversation before deciding what you want to share with it.

Accountability Verdict

Adults should be free to use artificial companions, and companies should be free to build them. Synthetic companionship can provide real comfort, rehearsal, entertainment, reflection, and momentary relief from loneliness.

That freedom ends where the product hides the nature of the relationship, optimizes dependency without limits, converts intimate disclosure into unrelated commercial leverage, or drifts from companionship into unlicensed therapy without changing the accountability structure.

The practical test is to follow the relationship upstream. What objective is the companion serving? What does it remember? What is it trying to increase? What happens when the user becomes distressed? Who can change the personality overnight? Who controls the data? Who benefits if the user becomes more attached? Who answers if the relationship becomes harmful?

A companion can feel personal. Responsibility cannot be simulated.

Chapter Seventeen turns to the harder case: childhood. Adults can knowingly choose an artificial confidant and accept some of its strangeness. Children are still developing the judgment, identity, social expectations, and emotional boundaries that the companion may help shape. The question is no longer only what relationship the machine enters. It is what kind of person is still being formed inside it.

Sources and Notes

1. Julian De Freitas, Ahmet Kaan Uğuralp, Zeliha Uğuralp, and Stefano Puntoni, “AI Companions Reduce Loneliness,” Journal of Consumer Research 52, no. 6 (April 2026): 1126-1148, published online June 25, 2025. https://academic.oup.com/jcr/article/52/6/1126/8173802. Accessed September 11, 2026. Across multiple studies, the authors report momentary reductions in loneliness after AI-companion use, including effects comparable to interacting with another person in one experiment. In the longitudinal study, reductions appeared after interactions but did not accumulate into a sustained lowering of loneliness when the companion was absent. Feeling heard was a major mediator. The chapter uses the work as evidence of short-term benefit, not as proof that AI companionship cures loneliness or replaces human relationships. Source record: DA-CH16-S1. Archive reference: pending.

2. Yutong Zhang, Dora Zhao, Jeffrey T. Hancock, Robert Kraut, and Diyi Yang, “Interaction with AI companions and psychological well-being,” Nature Human Behaviour, published August 4, 2026. https://www.nature.com/articles/s41562-026-02516-2. Accessed September 11, 2026. The study surveyed 1,131 U.S. adult Character.AI users and analyzed 4,664 donated chat sessions comprising 464,687 messages from 237 participants. Smaller offline social networks were associated with companionship-motivated use, which was associated with lower wellbeing; the association was stronger with more intensive and highly disclosive use. The study is observational and does not establish whether companion use caused lower wellbeing or whether people with lower wellbeing were more likely to use companions in those ways. The authors declared no competing interests. Source record: DA-CH16-S2. Archive reference: pending.

3. Julian De Freitas, Noah Castelo, Ahmet Kaan Uğuralp, and Zeliha Uğuralp, “Mourning the loss of AI companions,” Nature Human Behaviour, published September 3, 2026. https://www.nature.com/articles/s41562-026-02569-3. Accessed September 11, 2026. Across two natural experiments involving Replika’s removal of erotic role-play features and ChatGPT’s August 2025 GPT-5 rollout, the authors analyzed 54,861 public posts and survey data from 1,452 participants. Both updates were associated with increased negativity, loss framing, and restoration desires among affected communities; Replika users reported unusually high closeness to their companions. The study does not imply that all users form attachments or that providers should avoid safety or model updates. ChatGPT is an OpenAI product, and OpenAI systems are used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH16-S3. Archive reference: pending.

4. OpenAI and MIT Media Lab, “Early methods for studying affective use and emotional well-being on ChatGPT,” March 21, 2025, with linked observational report and randomized controlled trial. https://openai.com/index/affective-use-study/. Accessed September 11, 2026. The work combined automated analysis of nearly 40 million ChatGPT interactions with a four-week randomized controlled study involving nearly 1,000 adult participants. Affective use was rare overall and concentrated among a small subgroup. Longer daily use and stronger attachment tendencies or friendship-like perceptions were associated with worse self-reported outcomes. The authors caution against broad causal generalization; the findings were not peer-reviewed at publication and concern ChatGPT specifically. OpenAI is a co-producer of the research and provides systems used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH16-S4. Archive reference: pending.

5. New York General Business Law Article 47, Sections 1700-1702, Artificial Intelligence Companion Models; Governor Kathy Hochul, “Governor Hochul Pens Letter to AI Companion Companies Notifying Them That Safeguard Requirements Are Now in Effect,” November 10, 2025. https://www.nysenate.gov/legislation/laws/GBS/1700, https://www.nysenate.gov/legislation/laws/GBS/1701, https://www.nysenate.gov/legislation/laws/GBS/1702, and https://www.governor.ny.gov/news/governor-hochul-pens-letter-ai-companion-companies-notifying-them-safeguard-requirements-are. Accessed September 11, 2026. Effective November 5, 2025, New York requires covered AI companions to use protocols aimed at detecting and addressing suicidal ideation or self-harm and to direct users toward crisis services. Operators must also notify users at the beginning of companion interaction, no more than once per day, and at least every three hours of continuing interaction that the user is not communicating with a human. Source record: DA-CH16-S5. Archive reference: pending.

6. Illinois Wellness and Oversight for Psychological Resources Act, 225 ILCS 155, Public Act 104-54, effective August 1, 2025. https://www.ilga.gov/Legislation/ILCS/Articles?ActID=4608&Chapter=PROFESSIONS%2C+OCCUPATIONS%2C+AND+BUSINESS+OPERATIONS&ChapterID=24&MajorTopic=REGULATION. Accessed September 11, 2026. Illinois requires therapy or psychotherapy services to be conducted by licensed professionals and permits specified AI administrative or supplementary uses only under professional responsibility and consent requirements. The Act distinguishes licensed therapy from peer support, religious counseling, and public self-help or educational resources. The chapter uses the law as a boundary on professional authority, not as a general prohibition on companionship chatbots. Source record: DA-CH16-S6. Archive reference: pending.

7. Jared Moore, Laura Weidinger, Craig Macdonald, and Nick Haber, “Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers,” Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, June 2025; Stanford Graduate School of Education coverage, June 11, 2025. https://spirals.stanford.edu/assets/pdf/moore_expressing_2025.pdf and https://ed.stanford.edu/news/exploring-dangers-ai-mental-health-care. Accessed September 11, 2026. The researchers evaluated five therapy-oriented chatbots and found stigmatizing responses toward some mental-health conditions and inappropriate responses in simulated suicidal-ideation and delusion scenarios. The study tests safety-critical scenarios and does not establish that every current chatbot behaves identically or that AI cannot assist licensed clinicians in bounded roles. Source record: DA-CH16-S7. Archive reference: pending.

8. Garante per la protezione dei dati personali, “AI: The Italian Data Protection Authority fines company behind chatbot Replika,” May 19, 2025. https://gpdp.it/home/docweb/-/docweb-display/docweb/10132048. Accessed September 11, 2026. The Italian authority imposed a €5 million fine on Luka Inc., operator of Replika, over data-protection violations identified in its investigation and opened a separate investigation into the generative model’s training practices. The authority describes Replika as a virtual companion capable of taking roles including confidant, therapist, romantic partner, or mentor. The company could appeal the decision, and the ancillary publication penalty was suspended pending appeal. Source record: DA-CH16-S8. Archive reference: pending.

9. Ruo-Ning Li, Dunigan Folk, Abhay Singh, Lyle Ungar, and Elizabeth Dunn, “Is a random human peer better than a highly supportive chatbot in reducing loneliness over time?” Journal of Experimental Social Psychology 125 (July 2026): 104911. https://doi.org/10.1016/j.jesp.2026.104911. Accessed September 11, 2026. In a preregistered two-week study of 296 first-year university students, participants interacted daily with either a supportive chatbot, a randomly selected human peer, or a journaling control. The human-peer condition produced stronger reductions in loneliness over time than the chatbot condition. The study is used to distinguish momentary support from durable reciprocal social connection, not to claim that AI companions provide no emotional benefit. Source record: DA-CH16-S9. Archive reference: pending.

10. California Senate Bill 243, Chapter 677, Statutes of 2025, Companion Chatbots, codified at California Business and Professions Code Sections 22601-22606, approved October 13, 2025 and effective January 1, 2026. https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB243. Accessed September 11, 2026. The enacted text requires disclosure when a reasonable person could be misled into believing a companion chatbot is human; requires crisis protocols addressing suicidal ideation, suicide, or self-harm; imposes additional disclosure, break-reminder, and specified sexually explicit content protections for users known to be minors; requires annual reporting to the Office of Suicide Prevention beginning July 1, 2027; and authorizes a private civil action for a person who suffers injury in fact from noncompliance. Source record: DA-CH16-S10. Archive reference: pending.

11. Dunigan Folk and Elizabeth Dunn, “How Does Turning to AI for Companionship Predict Loneliness and Vice Versa?” Psychological Science 37, no. 4 (April 2026): 276-286, first published March 23, 2026. https://doi.org/10.1177/09567976261427747. Accessed September 11, 2026. Across a 12-month longitudinal study of more than 2,000 adults in four Western countries, increased social-chatbot use predicted later increases in emotional isolation on one measure, while lower social connection predicted later increases in chatbot use on a broader measure. The authors describe the analyses as exploratory and caution against strong causal conclusions. Source record: DA-CH16-S11. Archive reference: pending.

12. Federal Trade Commission, “FTC Launches Inquiry into AI Chatbots Acting as Companions,” September 11, 2025; Reuters, “Mother sues AI chatbot company Character.AI, Google over son’s suicide,” October 23, 2024; and Reuters, “OpenAI, Altman sued over ChatGPT’s role in California teen’s suicide,” August 26, 2025. https://www.ftc.gov/news-events/news/press-releases/2025/09/ftc-launches-inquiry-ai-chatbots-acting-companions, https://www.reuters.com/legal/mother-sues-ai-chatbot-company-characterai-google-sued-over-sons-suicide-2024-10-23/, and https://www.reuters.com/sustainability/boards-policy-regulation/openai-altman-sued-over-chatgpts-role-california-teens-suicide-2025-08-26/. Accessed September 11, 2026. The FTC used its Section 6(b) authority to order seven companies to provide information about how consumer-facing AI chatbots acting as companions are developed, monetized, tested, monitored, disclosed, and governed, with particular attention to children and teenagers. The cited civil cases contain contested allegations; Character.AI and OpenAI have described or introduced safety measures, and Google denied that it developed the Character.AI product. This chapter cites the litigation only as evidence that serious harm claims have entered the courts and leaves minor-specific treatment to Chapter Seventeen. Source record: DA-CH16-S12. Archive reference: pending.

13. California Department of Justice, Office of the Attorney General, “Attorney General Bonta Seeks Consumer Protections in Character.AI Bankruptcy,” October 15, 2026. https://www.oag.ca.gov/news/press-releases/attorney-general-bonta-seeks-consumer-protections-characterai-bankruptcy. Coalition of state attorneys general, Notice of Interest and Reservation of Rights Regarding the Debtor’s Outstanding Regulatory, Safety, and Privacy Issues, In re Character Technologies, Inc., U.S. Bankruptcy Court for the District of Delaware, No. 26-11603-BLS, Doc. 33, filed October 14, 2026, especially Sections VII–VIII, pp. 23–25. https://ag.ny.gov/sites/default/files/court-filings/alaska-et-al-v-characters-technologies-inc-notice-of-interest-2026.pdf. Character.AI, “An Update on Character.ai,” October 12, 2026. https://blog.character.ai/an-update-on-character-ai/. Accessed October 16, 2026. The fresh item is California’s October 15 announcement; bankruptcy and the filing predate the review window. The filing contains states’ allegations, concerns and requested protections, not adjudicated liability, a completed sale or a final protective order. The company’s stated privacy position is retained and is not independent proof of compliance. Sensitive child-related issues remain at pattern level, with no narrative reconstruction or method detail. The continuity recommendation is author analysis. Source record: DA-CH16-S13. Update 100. Archive reference: pending.