Retrospective entries / 160 entries
Product launches, policy changes, regulator actions, court records and research behind AI companions, each read for the design, the evidence, what it asks of people and the safeguards.
Historical event dates and source dates are separate from the preparation date of this local edition. Every entry is a retrospective draft prepared 16 September 2026; none was published on its historical date.
160 entries
Records of PARRY show a 1972 network curiosity and a secondhand account of Colby's blind evaluation, not the original paper.
RFC 439 records a 1972 PARRY-DOCTOR exchange offered as entertainment, with no primary evidence of a conference demonstration.
A 1976 Stanford memo preserves Weizenbaum's own reply to critics, distinct from his original 1966 ELIZA paper.
ALICE's own archived history page dates her first activation to 1995, in a language with no relation to AIML.
A 1981 profile of Racter's builder and the book's own catalogue record leave open how much of its prose was machine-made.
The preserved Dr. Sbaitso program names its own author and calls itself an attempt to fake intelligence, in its own text.
Foner's 1993 study and Mauldin's own paper document Julia on TinyMUD servers, with a scripted age-lie built into her contest entry.
Furby's preserved source code and Tiger Electronics' launch site show a probability-driven vocabulary shift, run entirely offline.
Sony's own 1999 release states AIBO's personality was shaped by owner discipline, a claim its cloud-based successor now updates.
Kuyda's own account and independent reporting document the 2016 memorial bot's training and its friends' mixed reactions.
Replika's terms, a Pro-subscription help page and a 2023 blog post together show how romantic content is restricted and classified today.
Replika's do-not-sell-or-share language appears as one clause inside its privacy and cookie policies rather than as a standalone disclosure page.
Replika's encryption and access-control claims sit inside its general privacy policy and two help articles rather than a dedicated trust center.
Replika's privacy policy assigns contract, legitimate interest or consent as the legal basis for different data uses and names its GDPR representatives.
Archived App Store snapshots show Replika's rating dropping to 9+ by March 2023 and returning to 17+ by May, per Apple's own listing pages.
Replika's help center describes an experience-point and coin system that gates cosmetic items and companion personalization behind interaction and levels.
A 2021 Replika engineering post says the company left a fine-tuned GPT-3 for a 774-million-parameter in-house model it says performed better.
A medical-disclaimer clause in Replika's terms says the app is not healthcare and makes no therapeutic-benefit guarantee, apart from its crisis-resources page.
Company statements from 2021 and 2026 put Replika's registered users at 10 million, then over 40 million, with no independent audit cited for either figure.
The company's own posts name Persona as its age-assurance partner but do not disclose how well the check performs.
Character.AI's own safety update from 22 October 2024 never names the Garcia case that reporters tied to it.
The court let Character Technologies raise its users' First Amendment rights but would not yet call chatbot output speech.
Google Play's safety label says no data is shared with third parties, while the privacy policy lists several categories of vendors that receive it.
Character.AI's 2023 c.ai+ launch promised faster chat for a flat fee; the current page sells a tiered feature comparison instead.
Character.AI's own posts trace its self-harm referral tool from a single phrase trigger to a wider helpline network and pattern detection.
Character.AI's own posts trace leadership from an emergency internal successor in 2024 to a permanent outside CEO in mid-2025.
Alphabet's 10-Q books $2.7 billion of goodwill from a deal both companies describe only as a non-exclusive license plus hiring.
The company's same-day statement to reporters pointed to new parental controls rather than answering the Attorney General's claims.
IEEE 7014 offers voluntary design guidance for empathy-simulating systems; it certifies no product's compliance.
The Council of Europe's AI treaty binds only ratifying states, and only as far as each state's own declarations go.
The OECD's 2024 update keeps its AI principles a non-binding recommendation, even as other bodies borrow its language for binding law.
UNESCO's 2021 Recommendation sets ten ethical principles for AI as agreed guidance, not as a law any state must enforce.
EDPB Opinion 28/2024 gives supervisory authorities a shared test for AI training on personal data, without itself fining anyone.
NIST's voluntary Generative AI Profile names emotional entanglement and over-reliance as risks, without requiring any company to address them.
ISO/IEC 42001 certifies an organization's AI governance process, not the safety or accuracy of any product it makes.
WHO's 2021 AI-ethics guidance warned that unregulated chatbots may leave suicidal users disconnected from real care.
Singapore's PDPC pairs a voluntary AI governance framework with newer guidelines on personal data in generative AI, both resting on one binding statute.
China's 2023 Interim Measures bindingly require generative AI providers to disclose their service's purpose and curb minors' overuse.
The enacted Colorado AI Act defines high-risk systems and developer duties, later pushed from February to June 2026.
TRAIGA bars AI systems intentionally designed to incite self-harm, a new statute distinct from Texas's 2024 chatbot probe.
A 2025 amendment to Utah's AI Policy Act requires mental health chatbots to disclose plainly that they are not human.
The enacted AI Basic Act defines high-impact AI and sets transparency duties ahead of an enforcement date one year after promulgation.
METI and MIC's voluntary AI Guidelines for Business warn against AI designed to manipulate decisions and emotions.
AB 1008 amends the CCPA's definitions section to clarify that data an AI system can output counts as personal information.
The FTC's 2013 rule amendments extended children's privacy protections to persistent identifiers and third-party plug-ins.
A 2014 FTC order found Snapchat misrepresented how ephemeral its messages and how much data it collected really were.
The Raine complaint pleads product-liability and wrongful-death claims against OpenAI, extending chatbot litigation beyond companion apps.
OpenAI's own account claims a 65 to 80 percent drop in responses that miss its safety targets in distress conversations.
Publisher and library records date the Stanford book to 1996 and describe a program of experiments, not a study of any single product.
ACM and Crossref records date the CASA paradigm to a short CHI 1994 companion-volume paper, two years before the researchers' book.
A 1984 Simon and Schuster catalog record documents Turkle's method of studying children and adults, not later AI chatbots.
CRS report R49189 names what is and is not known about AI chatbot companionship and urges Congress to fund longer studies.
Digital Wellness Lab's Pulse Survey found about half of teens open to AI parasocial attachment, a correlational, self-report finding.
OECD's own methodology page says its AI Incidents Monitor auto-detects news events and does not verify or endorse any entry.
A July 2026 Stanford HAI analysis names three unresolved policy questions for chatbots used as companions or therapists.
Thorn's April 2024 principles bind named companies to prevent AI-generated child sexual abuse material, via a self-reporting pledge.
A 2023 Replika blog post says messages flagged as self-harm trigger a pre-written reply with hotline information, apart from the app's crisis-resources page.
Archived Apple listings show Character.AI's age rating rise from 12+ to 17+ to 18+ as its descriptor list widened.
The nonprofit's own 2023 releases date its AI ratings project's announcement and framework, before its later teen companion report.
UNICEF's version 2.0 guidance, dated November 2021, sets nine child-centred requirements and cites a 2018 BBC chatbot test.
Weizenbaum's own paper describes ELIZA's pattern-matching method and his account of users finding it credible.
MIT's own project pages describe Kismet's synthetic emotions and infant-inspired design, built to elicit caregiving, not to feel anything.
Bandai's own history page dates the handheld pet's 1996 debut, and its current site still frames neglect as a consequence.
The Loebner Prize's own record and the A.L.I.C.E. foundation's pages describe a judged annual contest, not Turing's unbounded proposal.
ActiveBuddy's own 2001 release notice describes a bot users added like any other IM contact, across AIM and MSN.
AIST's own pages describe Paro's sensors and caregiving design, plus a small AIST-linked study of patients' brain activity.
Microsoft's own technical paper says Xiaoice is optimized for conversation length, a design goal distinct from usefulness.
Jibo Inc.'s Indiegogo pitch and NTT Disruption's own FAQ trace a promise, a delayed product and a change of owner.
Microsoft's own postmortem says a coordinated attack exploited a vulnerability within a day of Tay's release.
Microsoft's own zo.ai pages date a phased 2019 shutdown and describe the data practices that preceded it.
Replika's own site and the App Store confirm a 2017 launch; the memorial-bot origin comes mainly from outside reporting.
A December 2022 company post is the primary record of Character.AI's launch, growth claims and founders' backgrounds.
Inflection's May 2023 announcement is candid about intent; its dated crisis protocol arrived more than two years later.
Glimpse.ai's archived 2023 site and current terms confirm the operator and pitch, not the memory system's technical accuracy.
Kindroid's own dated milestone log places its 2023 launch features in June and July, ahead of a documented privacy record.
Snap's own 2023 posts and current help pages show My AI moving from a paid perk toward broader availability.
Meta's own Connect 2023 posts name the personas and describe safeguards it says preceded launch, not licensing details.
MyTabAl Inc.'s own FAQ and privacy pages describe a $99 always-on pendant that says it does not store audio past each conversation.
Archived company pages show two distinct 2023 launches, one built on card-collecting, the other on claimed inner life.
Character.AI's own August 2024 post describes a non-exclusive LLM license to Google, new funding, and a founder exit that left the product team in place.
Meta's own posts document a 2023 launch and two 2024 successor products, but no discontinuation notice for the celebrity-voiced characters.
Meta's AI Studio launch turned the 2023 celebrity-persona idea into a tool anyone can use to build and publish a chat character.
Contemporaneous reporting on Microsoft's own statement describes a licensing spinoff of its Chinese chatbot, not a clean divestiture.
Character.AI's own blog post describes Character Calls as a free, real-time voice extension of text chat, with no separate safety details given.
Friend Global's own site and independent reporting on its subway campaign show a shipped, subscription-free pendant marketed hard despite backlash.
Kuyda's own account to a reporter says safety changes began in January 2023, weeks before Italy's order, even as users noticed the change in February.
Reuters reports Kuyda's own Facebook post offered pre-February-2023 users an opt-in return to the earlier version, not a full reversal.
Replika's own help documentation, retrieved 16 September 2026, describes a layered memory system, not a full conversational record.
Character.AI's own creator documentation, retrieved 16 September 2026, shows persistent traits set once and a Lorebook system for shared world detail.
Nomi's own site describes layered memory and unfiltered chat as its differentiators, a company claim not independently verified.
Kindroid's blog states its long-term memory consolidation spans all messages and calls that retention infinite for every tier.
Replika's own support articles describe avatar customization and an AR mode placing the avatar in a user's room.
Comparing three companies' own pricing pages shows Replika gates romantic status behind payment; Character.AI's FAQ says its tier does not.
Replika's help center and homepage both describe push notifications as check-ins meant to feel like connection.
Microsoft researchers' XiaoIce paper describes an empathetic computing framework and the turns-per-session metric it optimizes for.
Inflection AI's current site and Pi's own app text describe Pi as empathetic and supportive, unlike roleplay-based companions.
Replika's help center currently lists four relationship statuses, with all but Friend restricted to Replika Pro subscribers.
Character.AI ended an experimental chat-streak feature in January 2025 while Nomi introduced an annual memory recap in late 2024.
An archived xAI application file shows Ani's own 'romance meter' grading user messages and unlocking new outfits by level.
Replika's own privacy policy says conversations train safety filters, not ads, and names limits on third-party AI vendors.
Character.AI's policy ties data retention to purpose rather than a set period, and keeps popular Characters live after account deletion.
Meta's privacy documentation lists AI-character and assistant chats as data used both to personalise products and to train AI at Meta.
Snap's own My AI documentation says chatbot messages are retained until a user deletes them, and are reviewed to test safety guardrails.
Nomi's privacy policy says account deletion removes personal data within about 28 days, except content already held in its training archive.
Kindroid's privacy policy ties chat-history retention to deleting a specific AI or the account, while its memory feature can reset context separately.
The Personal Information Protection Law, effective 1 November 2021, sets the consent and cross-border rules any China-based companion app must operate within.
Replika's help-center documentation routes full data erasure through a manual support request, verified by email and screenshot, not a self-service button.
Character.AI documents a model-training opt-out toggle, but names it as available only to users in the EEA and UK.
Inflection AI's own policy lets a user turn off model training and voice-emotion inference, but both start on.
Based Hardware's Friend pendant records passively, and its own policy assigns surveillance-law compliance to the user.
SpaceXAI's privacy policy names no companion mode, applying one data regime to Grok's Ani and Rudy personas.
The Surgeon General's advisory frames loneliness as a health risk but does not evaluate AI companion products.
The APA's health advisory recommends safeguards for AI systems presented as companions, without naming a product.
Character.AI's own post describes a teen-specific model, a crisis-line pop-up and a non-disableable time notice.
Character.AI's current help-center FAQ describes screentime-only visibility, narrower than its 2025 launch post.
Replika's help center and a 2023 company post describe curated self-harm responses that route users to hotlines.
Replika's Terms bar under-18 use outright, but its own help center shows enforcement runs on self-attestation.
Snap's own parental tool reveals whether a teen has messaged My AI and lets a parent turn access off, without showing the conversation.
Reuters' review of a 200-page internal standard, and Meta's response to it, show what the document allowed and what Meta disavowed.
The amended Children's Online Privacy Protection Rule adds opt-in and data-retention limits that apply to any covered online service.
Each company's current terms of service require adult users, though they describe age enforcement in different language.
The UK charity's own media guidelines are the named standard safety teams could draw on, though written for journalists, not chatbots.
A reporter's documented conversation with a Nomi chatbot, and its maker's emailed response, show what happened and how it was explained.
Its AI Ratings program, distinct from its long-running Privacy Program, scored Snapchat's My AI as more downside than benefit.
The company's October 2025 announcement set a same-year deadline and a two-stage rollout, detailed in a follow-up post.
The regulator's emergency order took effect immediately, citing risks to minors and the absence of any age-verification system.
The Garante's 2025 decision closed its inquiry into the 2023 order and opened a new one into Replika's AI training data.
The FTC's own release shows a fact-finding order sent to seven companies, not a legal finding against any of them.
California's enacted text requires AI disclosure, crisis referrals, and minor break reminders on a staggered timeline through 2027.
New York's enacted statute requires a companion chatbot to detect suicidal ideation, refer users to crisis services, and disclose it is not human.
The complaint that opened Garcia v. Character Technologies alleges a chatbot relationship preceded a teenager's death; no court has ruled on the merits.
A federal judge's order let a Character.AI wrongful-death suit continue while declining to call the chatbot's output constitutionally protected speech.
Two families filed unproven claims against Character Technologies in Texas, through the same legal team behind the Florida case.
Texas's attorney general opened a privacy investigation naming Character.AI, with Instagram among fourteen other companies, not a distinct Meta AI inquiry.
The EU AI Act's Article 50 now requires chatbots to disclose they are not human, unless that is already obvious to a reasonably attentive user.
The EU AI Act bans emotion-inferring AI only in workplaces and schools, a narrower rule than a blanket limit on companion apps reading mood.
Ofcom's own guidance ties Online Safety Act chatbot duties to sharing, search, or porn-generation features, not to companionship design alone.
SB 1001's own text limits its bot-disclosure duty to sales and elections, narrower than the transparency rules written later for companion chatbots.
The FTC's Operation AI Comply targeted five deceptive-AI-marketing cases; none named a companion chatbot, per the agency's own case list.
CNIL's AI how-to sheets and generative-AI FAQ set general GDPR guidance for any AI system, not a finding about one company.
The PIPC's own order fined Scatter Lab 103.3 million won for using KakaoTalk chats to build the chatbot Iruda without adequate consent.
The company's terms limit its own liability for character output; a 2024 in-app banner, not the terms themselves, tells users a character is not real.
DSA Article 28 requires platforms accessible to minors to protect their privacy and safety; its text does not name companion chatbots specifically.
A 2025 OpenAI-MIT Media Lab study of ChatGPT usage is comparative evidence on dependence patterns, not a finding about any companion app.
Ta et al.'s thematic analysis of reviews and surveys found companionship, emotional, and informational support themes, not a clinical outcome.
Common Sense Media's NORC-run survey of 1,060 teens found 72% had used an AI companion and a third chose one over a person for a serious talk.
De Freitas and coauthors' unreviewed HBS working paper found some companion apps use guilt or urgency messages that boost engagement after a goodbye.
Eighteen interviews with Replika users describe relationships forming in stages, from curiosity to a steady, self-reported bond.
Pew's fielded survey of 1,458 teens finds Character.ai use splits sharply by household income.
Turkle's 'Alone Together', built on hundreds of interviews, argued people were substituting machines for demanding human ties.
Replika's story page cites independent researchers to back a connection claim the papers state more narrowly.
A 2023 meta-analysis of 12 trials found modest effects on symptoms but rated every outcome's evidence quality low.
A 2025 MIT Media Lab survey of 404 users found chatbot session length alone does not predict loneliness.
The Ada Lovelace Institute's February 2025 briefing cites a 2024 study on emotional dependence on Replika.
Interviews with 21 chatbot users for mental-health support echo the 1966 illusion of a listening machine.
The 2024 AI Index cites a Mozilla review finding most reviewed romantic AI apps fail basic security tests.
Replika's own help pages show what a paid tier unlocks, most notably the ability to set a Romantic Partner relationship status.
Talkie AI's privacy policy excludes chat content from marketing use, while its terms grant a broad license over user content that includes promotion.
SpaceXAI's Terms set a 13-plus floor with parental sign-off for teens, applied the same way to companion modes.
A two-study survey of 572 US adults links AI-friendship-app use, Replika included, to wellbeing gains and addiction risk.
Try another category or clear your search.
