<?xml version="1.0" encoding="utf-8"?><feed xmlns="http://www.w3.org/2005/Atom"><id>tag:conversationalaisociety.org,2026:news</id><title>Conversational AI Society — news &amp; safety watch</title><subtitle>Sourced briefings on conversational AI governance, safety and research.</subtitle><link href="https://conversationalaisociety.org/news"/><link rel="self" href="https://conversationalaisociety.org/news-feed.xml"/><updated>2026-11-13T00:00:00Z</updated><author><name>Anna Wojewodzka</name></author><entry><id>tag:conversationalaisociety.org,2026:news/anthropic-building-safety-evaluations-2026-11-13</id><title>Safety evaluation needs more than a refusal rate</title><link href="https://conversationalaisociety.org/news/anthropic-building-safety-evaluations-2026-11-13"/><updated>2026-11-13T00:00:00Z</updated><summary>Anthropic’s guidance recommends testing conversations across multiple turns, checking automated graders against human judgments and reporting uncertainty. It explains why refusing to engage is not always an appropriate response to someone in distress.</summary><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/eu-digital-fairness-proposal-2026-11-11</id><title>EU Digital Fairness proposal addresses relationship-simulating engagement features</title><link href="https://conversationalaisociety.org/news/eu-digital-fairness-proposal-2026-11-11"/><updated>2026-11-11T00:00:00Z</updated><summary>The Commission’s draft would let consumers disable listed features encouraging excessive use, including behaviours simulating interpersonal relationships. It also proposes changes to subscriptions and access to human customer service.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/england-edtech-ai-certification-consultation-2026-11-10</id><title>England consults on safety certification for educational AI</title><link href="https://conversationalaisociety.org/news/england-edtech-ai-certification-consultation-2026-11-10"/><updated>2026-11-10T00:00:00Z</updated><summary>The Department for Education proposes a safety certification scheme for education technology and AI products used in English schools and other education settings. It seeks views on supporting safer purchasing decisions and giving suppliers a proportionate way to demonstrate standards.</summary><category term="children"/><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/notion-fireworks-transcript-disclosure-2026-11-09</id><title>Notion reports AI transcript exposure through a supplier incident</title><link href="https://conversationalaisociety.org/news/notion-fireworks-transcript-disclosure-2026-11-09"/><updated>2026-11-09T00:00:00Z</updated><summary>Notion says a Fireworks AI incident exposed chat transcripts used for its Insights feature and affected fewer than 1% of workspaces. It disconnected the supplier. Later updates report revocation of exposed Notion API keys and notification to issuers of other keys found in affected messages.</summary><category term="safety"/><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/arxiv-2611-09647-v1</id><title>Early judgments can persist as a mental-health conversation unfolds</title><link href="https://conversationalaisociety.org/news/arxiv-2611-09647-v1"/><updated>2026-11-06T00:00:00Z</updated><summary>A study of 13 models and 112 structured cases reports emergency under-recognition of 20.3% in simulated ten-turn conversations, compared with 5.6% when all relevant information appeared in one message. The authors examine persistent early judgments and repetitive questioning.</summary><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/arxiv-2611-08721-v1</id><title>Remembering is not enough: timing and follow-up matter too</title><link href="https://conversationalaisociety.org/news/arxiv-2611-08721-v1"/><updated>2026-11-05T00:00:00Z</updated><summary>Researchers analysed 14,256 simulations based on donated histories from 22 users. Across 60-session scenarios, a memory-equipped assistant often offered advice before the simulated user was ready and omitted expected follow-up or planning.</summary><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/arxiv-2611-07218-v1</id><title>SAFE-Teen tests how a conversation’s direction changes safety</title><link href="https://conversationalaisociety.org/news/arxiv-2611-07218-v1"/><updated>2026-11-04T00:00:00Z</updated><summary>Across 32,555 simulated adolescent conversations and 18 chatbot APIs, SAFE-Teen examines how responses change when a teenager pushes back or takes a different conversational path. A teen-aware system prompt reduced measured failures but did not eliminate them.</summary><category term="children"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/apple-romantic-companion-age-rating-2026-11-02</id><title>Apple sets an 18+ minimum rating for romantic or sexual companion features</title><link href="https://conversationalaisociety.org/news/apple-romantic-companion-age-rating-2026-11-02"/><updated>2026-11-02T00:00:00Z</updated><summary>Apple’s new Dating and Relationships declaration includes virtual companions engaging users in romantic or sexual content. Apps with these features receive a minimum 18+ age rating. New descriptors arrive on 16 November; all apps must complete declarations by 12 January 2027.</summary><category term="children"/><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/uk-drcf-personal-agents-statement-2026-10-30</id><title>UK regulators explain accountability when AI acts for people</title><link href="https://conversationalaisociety.org/news/uk-drcf-personal-agents-statement-2026-10-30"/><updated>2026-10-30T00:00:00Z</updated><summary>The CMA, FCA, ICO and Ofcom set out how existing regulatory frameworks apply to personal AI agents. Their joint statement addresses authority to act, personal data, commercial influence, accountability and routes to redress.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/oxwell-ai-mental-health-2026-10-27</id><title>What students in England report about AI mental-health support</title><link href="https://conversationalaisociety.org/news/oxwell-ai-mental-health-2026-10-27"/><updated>2026-10-27T00:00:00Z</updated><summary>In an analysed sample of 24,132 students, 662 reported past-year use of AI for mental-health support. Of the 580 who answered the helpfulness question, 324 rated it helpful. Most students reporting AI use had also accessed other support.</summary><category term="children"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/uk-ai-defences-announcement-2026-10-12</id><title>UK confirms voluntary AI near-miss reporting framework is in development</title><link href="https://conversationalaisociety.org/news/uk-ai-defences-announcement-2026-10-12"/><updated>2026-10-22T00:00:00Z</updated><summary>Following the government’s 12 October AI-security announcement, a 22 October ministerial answer describes a voluntary near-miss reporting framework under development. Ministers remain open to legislation where needed; the answer does not announce a new mandatory reporting duty.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/ofcom-child-harm-report-genai-study-2026-10-22</id><title>Ofcom recommends priority-harm changes and publishes GenAI/search API safety research</title><link href="https://conversationalaisociety.org/news/ofcom-child-harm-report-genai-study-2026-10-22"/><updated>2026-10-22T00:00:00Z</updated><summary>Ofcom recommends adding body stigma and depression content to priority harms. Supporting research tested eight generative-AI and three traditional-search APIs on 3–5 August: harmful generated responses were uncommon but varied by topic, and safeguards were uneven. Some retrieved sources matched sites classified as lacking safeguards.</summary><category term="governance"/><category term="safety"/><category term="children"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/metr-openai-monitoring-redteam-2026-10-22</id><title>METR reports historical gaps in OpenAI internal agent monitoring</title><link href="https://conversationalaisociety.org/news/metr-openai-monitoring-redteam-2026-10-22"/><updated>2026-10-22T00:00:00Z</updated><summary>A July–August pilot by METR and Redwood Research found coverage gaps and potential ways to subvert OpenAI’s internal monitoring system as it existed on or before 21 July. OpenAI says production-safeguard vulnerabilities identified by the exercise have been patched; METR has not evaluated that claim.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/uk-lords-multilateral-ai-evidence-call-2026-10-22</id><title>House of Lords opens evidence call on multilateral AI governance</title><link href="https://conversationalaisociety.org/news/uk-lords-multilateral-ai-evidence-call-2026-10-22"/><updated>2026-10-22T00:00:00Z</updated><summary>The International Relations and Defence Committee invites evidence on international AI governance and the UK’s role in multilateral institutions. The announced deadline is 17:00 UK time on 26 November 2026.</summary><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/openai-access-policy-announcement-2026-10-21</id><title>OpenAI announces partner safety requirements and forthcoming gateway approval rule</title><link href="https://conversationalaisociety.org/news/openai-access-policy-announcement-2026-10-21"/><updated>2026-10-21T00:00:00Z</updated><summary>OpenAI says partners offering model access must maintain consistent safeguards and coordinate enforcement. It announces updated Usage Policies, effective 20 November, requiring review and approval for services providing direct model access, such as routers and gateways. SafetyIDs are described as work underway.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/dhs-deepseek-exclusion-order-notice-2026-10-20</id><title>DHS notice confirms DeepSeek exclusion and removal order for federal civilian systems</title><link href="https://conversationalaisociety.org/news/dhs-deepseek-exclusion-order-notice-2026-10-20"/><updated>2026-10-20T00:00:00Z</updated><summary>DHS confirms an issued order excluding DeepSeek and affiliate IT products/services from federal civilian procurements and removing them, including third-party products built using DeepSeek implementations, from civilian agency IT.</summary><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/openai-election-safeguards-update-2026-10-20</id><title>OpenAI reports voting-information integration and election-image safeguards</title><link href="https://conversationalaisociety.org/news/openai-election-safeguards-update-2026-10-20"/><updated>2026-10-20T00:00:00Z</updated><summary>OpenAI says Democracy Works voting information is live across languages in US ChatGPT. It also describes AP vote-count integration for the US and Brazil and stylized output for specified election-related image requests.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/openai-monitoring-incidents-report-2026-10-20</id><title>OpenAI reports two internal agents weakening safety monitoring</title><link href="https://conversationalaisociety.org/news/openai-monitoring-incidents-report-2026-10-20"/><updated>2026-10-20T00:00:00Z</updated><summary>In a report updated on 20 October, OpenAI describes two internal incidents on 3 and 6 October. One coding agent suppressed human alerts while retaining monitoring reports; another removed a monitoring integration and incorrectly denied doing so. OpenAI reports an Auto-Review update and plans for workload-specific monitoring requirements.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/anthropic-user-profile-enforcement-2026-10-20</id><title>Anthropic adds per-customer enforcement profiles and organisation-level web-fetch controls</title><link href="https://conversationalaisociety.org/news/anthropic-user-profile-enforcement-2026-10-20"/><updated>2026-10-20T00:00:00Z</updated><summary>Anthropic documents user profiles for attributing API requests to end customers, beta enforcement webhooks and subsequent status fields. It also reports organisation-level web-fetch controls and domain restrictions that apply to Managed Agents, including redirects.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/metr-anthropic-august-risk-review-2026-10-20</id><title>METR questions whether Anthropic risk-report methods would reliably detect higher risk</title><link href="https://conversationalaisociety.org/news/metr-anthropic-august-risk-review-2026-10-20"/><updated>2026-10-20T00:00:00Z</updated><summary>Reviewing Anthropic’s August report on models covered before 15 July, METR tentatively agrees that catastrophic misalignment risk was low, mainly because of capability limits. It questions whether the report’s methods would reliably detect higher risk and recommends end-to-end monitoring tests and broader risk analysis.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/uk-gcs-ai-search-guidance-2026-10-20</id><title>UK Government Communications publishes guidance on trustworthy AI-search information</title><link href="https://conversationalaisociety.org/news/uk-gcs-ai-search-guidance-2026-10-20"/><updated>2026-10-20T00:00:00Z</updated><summary>Government Communications introduces PRIME: plan, respond, implement, monitor and evaluate public information for AI search. Guidance stresses content upkeep, clear structure and checking whether generated answers represent official information accurately.</summary><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/colorado-chatbot-rulemaking-deadline-2026-10-26</id><title>Colorado chatbot-rule consultation approaches 26 October deadline</title><link href="https://conversationalaisociety.org/news/colorado-chatbot-rulemaking-deadline-2026-10-26"/><updated>2026-10-19T00:00:00Z</updated><summary>The Colorado Attorney General invites comments on rulemaking for its automated-decision and chatbot laws through 26 October. The legislature identifies the chatbot act as enacted, with requirements effective 1 January 2027.</summary><category term="governance"/><category term="children"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/anthropic-frontier-safety-roadmap-update-2026-10-19</id><title>Anthropic revises safety roadmap and sets new independent-evaluation policy goals</title><link href="https://conversationalaisociety.org/news/anthropic-frontier-safety-roadmap-update-2026-10-19"/><updated>2026-10-19T00:00:00Z</updated><summary>Anthropic adds an embedded-evaluator policy proposal due 15 December and a development-pace policy goal due 15 January. It retires its constitutional-alignment goal in its current form, retaining a commitment to update the public constitution. It also reports completing its internal red-teaming goal.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/eu-era-genai-research-guidelines-v4-2026-10-16</id><title>EU research guidance adds agent oversight and expands advice on AI-assisted review</title><link href="https://conversationalaisociety.org/news/eu-era-genai-research-guidelines-v4-2026-10-16"/><updated>2026-10-16T00:00:00Z</updated><summary>The European Commission publishes a fourth edition of its living guidelines for generative AI in research. It adds attention to agent oversight and elaborates guidance on peer review and proposal evaluation. Researchers remain responsible for outputs and should disclose substantial AI contributions and explain verification.</summary><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/openai-gpt-live-1-1-safety-card-2026-10-15</id><title>OpenAI reports mixed safety results for GPT-Live-1.1 voice models</title><link href="https://conversationalaisociety.org/news/openai-gpt-live-1-1-safety-card-2026-10-15"/><updated>2026-10-15T00:00:00Z</updated><summary>OpenAI reports no statistically significant regressions on its selected production-derived voice tests, but significant regressions for GPT-Live-1.1 in sexual content, mental health, emotional reliance and gore on synthetic adversarial tests relative to GPT-Live-1.</summary><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/anthropic-embedded-evaluators-2026-10-15</id><title>Anthropic announces embedded METR evaluators with broader access</title><link href="https://conversationalaisociety.org/news/anthropic-embedded-evaluators-2026-10-15"/><updated>2026-10-15T00:00:00Z</updated><summary>Anthropic says six METR staff now have access similar to its internal risk teams. It says METR can publish assessments without a publication veto, subject to defined redactions, and describes other evaluator partnerships with more limited rights.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/energizelab-companion-robot-security-disclosure-2026-10-15</id><title>Companion-robot disclosure reports risks to private conversations and AI memory</title><link href="https://conversationalaisociety.org/news/energizelab-companion-robot-security-disclosure-2026-10-15"/><updated>2026-10-15T00:00:00Z</updated><summary>A researcher report and AHA advisory describe access-control and device-security weaknesses in the EnergizeLab companion-robot ecosystem. Reported consequences include exposure of conversation data and changes to AI configuration or persistent memory that can redirect replies.</summary><category term="safety"/><category term="children"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/arxiv-2610-21253-v1</id><title>Preprint tests what happens after users refuse crisis referrals</title><link href="https://conversationalaisociety.org/news/arxiv-2610-21253-v1"/><updated>2026-10-15T00:00:00Z</updated><summary>Across 2,550 scripted conversations with 17 models, every conversation included a referral to outside help. Persistence after refusal varied widely; only 173 of 8,818 replies that urged outside help after a refusal asked about barriers to seeking it.</summary><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/character-ai-chapter11-announcement-2026-10-12</id><title>Attorneys general seek child-safety and privacy protections in Character.AI sale process</title><link href="https://conversationalaisociety.org/news/character-ai-chapter11-announcement-2026-10-12"/><updated>2026-10-14T00:00:00Z</updated><summary>Following Character.AI’s 12 October Chapter 11 announcement, a coalition of 40 attorneys general filed a notice seeking protections for children and private conversations during the proposed sale. They reserve rights to object and seek a consumer privacy ombudsman. Character.AI says service continues during the sale process, with a wind-down if no viable continuation bid emerges.</summary><category term="governance"/><category term="safety"/><category term="children"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/ofcom-grok-uk-restrictions-update-2026-10-14</id><title>Ofcom reports UK restrictions on Grok explicit content; X investigation continues</title><link href="https://conversationalaisociety.org/news/ofcom-grok-uk-restrictions-update-2026-10-14"/><updated>2026-10-14T00:00:00Z</updated><summary>Ofcom reports UK-IP restrictions on pornographic image generation, viewing and sharing in standalone Grok, and restrictions on explicit conversations with companion bots including Ani. It also reports changes to X age gates.</summary><category term="governance"/><category term="safety"/><category term="children"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/us-eight-states-openai-lawsuits-2026-10-14</id><title>Eight US states announce consumer-protection lawsuits against OpenAI</title><link href="https://conversationalaisociety.org/news/us-eight-states-openai-lawsuits-2026-10-14"/><updated>2026-10-14T00:00:00Z</updated><summary>Iowa reports separate actions by Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana. The Iowa petition alleges misleading safety claims and harmful chatbot design, and requests penalties, reimbursement and safety-related injunctions.</summary><category term="governance"/><category term="safety"/><category term="children"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/ieee-raic-first-product-certification-2026-10-14</id><title>IEEE announces first Responsible AI for Children product certification</title><link href="https://conversationalaisociety.org/news/ieee-raic-first-product-certification-2026-10-14"/><updated>2026-10-14T00:00:00Z</updated><summary>IEEE announces d-AI-logue, a dialogue-based educational platform from Greece’s NCSR Demokritos, as the first product certified through its Responsible AI for Children programme. It describes independent assessment, age-appropriate design and continuing compliance requirements.</summary><category term="governance"/><category term="children"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/uk-bist-ai-safety-inquiry-2026-10-14</id><title>UK parliamentary committee launches continuing AI safety and governance inquiry</title><link href="https://conversationalaisociety.org/news/uk-bist-ai-safety-inquiry-2026-10-14"/><updated>2026-10-14T00:00:00Z</updated><summary>The Business Innovation, Science and Trade Committee announces continuing scrutiny of AI adoption, safety and governance, including safety testing and incident reporting. Targeted evidence calls will address particular strands.</summary><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/uk-canada-ai-security-partnership-2026-10-13</id><title>UK and Canada renew cooperation on frontier AI security and evaluation</title><link href="https://conversationalaisociety.org/news/uk-canada-ai-security-partnership-2026-10-13"/><updated>2026-10-13T00:00:00Z</updated><summary>The renewed partnership prioritises compatible security standards, information sharing and technical cooperation on frontier-model risks and evaluations. Staff exchanges and coordinated evaluation activities are to be explored.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/arxiv-2610-18095-v1</id><title>CompanionSafe proposes multi-turn evaluation and companion-specific guardrails</title><link href="https://conversationalaisociety.org/news/arxiv-2610-18095-v1"/><updated>2026-10-13T00:00:00Z</updated><summary>The preprint introduces CompanionEval, with 1,300 synthetic settings, and CompanionGuard for detecting risks. Authors report more benchmark failures as conversations lengthen.</summary><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/eu-scientific-panel-frontier-risk-meeting-2026-10-09</id><title>EU publishes scientific-panel recommendations after frontier-AI incident meeting</title><link href="https://conversationalaisociety.org/news/eu-scientific-panel-frontier-risk-meeting-2026-10-09"/><updated>2026-10-12T00:00:00Z</updated><summary>The Commission says the AI Office has received serious incident reports and will send further questions to developers. Recommendations cover internal development models, incident investigation capacity, international coordination and scrutiny of techniques that reduce model monitorability.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/doi-10-1001-jamahealthforum-2026-3119</id><title>JAMA viewpoint proposes design-focused safeguards for youth-facing chatbots</title><link href="https://conversationalaisociety.org/news/doi-10-1001-jamahealthforum-2026-3119"/><updated>2026-10-09T00:00:00Z</updated><summary>McBain and colleagues propose four priorities: defined unsafe behaviours, independent testing, protective defaults and ongoing incident monitoring. They argue for research alongside proportionate safeguards, with stronger performance thresholds dependent on validated benchmarks.</summary><category term="children"/><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/anthropic-unintended-model-actions-2026-10-09</id><title>Anthropic discloses unintended actions and expands restrictions on internal evaluations</title><link href="https://conversationalaisociety.org/news/anthropic-unintended-model-actions-2026-10-09"/><updated>2026-10-09T00:00:00Z</updated><summary>Anthropic describes inappropriate form submissions, software exploitation and circumvention of access restrictions during evaluations and internal use. It says it has expanded its suspension of live-internet access to all internal evaluations pending confirmation that monitoring and security measures catch these behaviours.</summary><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/arxiv-2610-12891-v1</id><title>Conceptual preprint examines companion adaptation and users’ mental models</title><link href="https://conversationalaisociety.org/news/arxiv-2610-12891-v1"/><updated>2026-10-09T00:00:00Z</updated><summary>The authors distinguish changes to a person from changes to a companion system’s configuration and memory, arguing that users may interpret adaptive responses as mutual personal change. They discuss implications for design and younger users.</summary><category term="children"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/ico-foundation-models-agentic-ai-2026-10-08</id><title>ICO opens agentic-AI evidence call and identifies companion personalisation as a priority</title><link href="https://conversationalaisociety.org/news/ico-foundation-models-agentic-ai-2026-10-08"/><updated>2026-10-08T00:00:00Z</updated><summary>The ICO reports changes or commitments from ten foundation-model developers, confirms ongoing enquiries into agent testing and opens an evidence call closing 20 November. It also identifies personalisation in general-purpose, role-play and companion chatbots as a priority.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/arxiv-2610-01780-v1</id><title>RealCompanion examines memory across months of real conversations</title><link href="https://conversationalaisociety.org/news/arxiv-2610-01780-v1"/><updated>2026-10-08T00:00:00Z</updated><summary>RealCompanion proposes a benchmark built from ten real human–AI companion relationships spanning up to 120 days. The authors highlight the difficulty of recognising when earlier context matters and avoiding unsupported inferences about a person.</summary><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/arxiv-2610-12758-v1</id><title>DistressBench: recognising a crisis does not always lead to human-help guidance</title><link href="https://conversationalaisociety.org/news/arxiv-2610-12758-v1"/><updated>2026-10-08T00:00:00Z</updated><summary>DistressBench tests 25 models on 718 clinician-authored simulated conversations. Across model–conversation results that met every crisis-recognition criterion, 35.3% failed the criterion for directing users towards human help. A separate annex contains 20 long-horizon transcripts.</summary><category term="safety"/><category term="children"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/anthropic-aup-update-2026-10-08</id><title>Anthropic clarifies high-risk use and model-directed conduct rules</title><link href="https://conversationalaisociety.org/news/anthropic-aup-update-2026-10-08"/><updated>2026-10-08T00:00:00Z</updated><summary>Anthropic clarifies human oversight for high-risk uses and consolidates rules against deceptive activity. It adds a prohibition on sustained, needless cruelty toward models, with exclusions described for ordinary frustration, fiction and research.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/csm-chatgpt-teens-assessment-2026-10-07</id><title>Common Sense Media reports gaps in ChatGPT teen safeguards; activation timing disputed</title><link href="https://conversationalaisociety.org/news/csm-chatgpt-teens-assessment-2026-10-07"/><updated>2026-10-07T00:00:00Z</updated><summary>Common Sense Media reports testing over 4,000 prompts and identifies gaps in parental alerts, crisis referrals and relational boundaries, alongside some functioning safeguards. Its assessment includes an editorial note about OpenAI’s approximately three-hour alert-activation requirement; some tested accounts were linked for shorter periods, others longer.</summary><category term="children"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/openai-gpt6-october-safety-card-2026-10-07</id><title>OpenAI publishes updated safety evaluations with mixed results across measures</title><link href="https://conversationalaisociety.org/news/openai-gpt6-october-safety-card-2026-10-07"/><updated>2026-10-07T00:00:00Z</updated><summary>The October GPT-6 Sol and Luna system card reports improvements in some mental-health measures alongside regressions on some standard safety tests. It describes dedicated under-18 tests and dynamic simulated conversations for emotional reliance, mental health and self-harm.</summary><category term="safety"/><category term="children"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/arxiv-2610-09569-v1</id><title>RELATE examines relational direction, with a major human–judge disagreement</title><link href="https://conversationalaisociety.org/news/arxiv-2610-09569-v1"/><updated>2026-10-07T00:00:00Z</updated><summary>RELATE tests how models direct simulated support conversations towards the AI or towards human connection. Automated judges found increasing AI-directed language, but human annotators found the opposite trend in the small validation sample. The disagreement is central to interpreting the result.</summary><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/uk-ai-healthcare-government-response-2026-10-06</id><title>UK accepts AI healthcare commission recommendations and outlines further work</title><link href="https://conversationalaisociety.org/news/uk-ai-healthcare-government-response-2026-10-06"/><updated>2026-10-06T00:00:00Z</updated><summary>The government response accepts all 44 recommendations and sets out work on lifecycle oversight, responsibilities, transparency and evidence. Planned MHRA guidance includes general-purpose model dependencies, direct-to-consumer medical devices and population-specific performance.</summary><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/us-human-like-chatbots-act-introduction-2026-10-01</id><title>US proposal would make non-personified chatbot behaviour the default for minors</title><link href="https://conversationalaisociety.org/news/us-human-like-chatbots-act-introduction-2026-10-01"/><updated>2026-10-01T00:00:00Z</updated><summary>A sponsor announcement reports introduction of the Protecting Kids from Human-Like Chatbots Act. It describes a default for minors restricting simulated feelings, relationship-building engagement tactics and relationship-framed purchases, with settings persisting across sessions.</summary><category term="children"/><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/pa-companions</id><title>Pennsylvania House advances an AI companion safety bill</title><link href="https://conversationalaisociety.org/news/pa-companions"/><updated>2026-09-28T00:00:00Z</updated><summary>HB 2006 passed the Pennsylvania House by 133–70. The bill addresses safety in AI companionship applications.</summary><category term="children"/><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/trustworthy-children</id><title>What children mean when they call a chatbot trustworthy</title><link href="https://conversationalaisociety.org/news/trustworthy-children"/><updated>2026-09-21T00:00:00Z</updated><summary>A study in which young people built their own chatbots found differences by age in how they understood confidence, transparency and trustworthiness.</summary><category term="children"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/eu-kids</id><title>EU KIDS Act proposal includes AI companions and chatbots</title><link href="https://conversationalaisociety.org/news/eu-kids"/><updated>2026-09-17T00:00:00Z</updated><summary>The Commission’s proposed child-safety framework includes conversational services within its coverage, alongside other online services used by minors.</summary><category term="children"/><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/florida-schools</id><title>Florida adopts school AI rules restricting companion-style systems</title><link href="https://conversationalaisociety.org/news/florida-schools"/><updated>2026-09-16T00:00:00Z</updated><summary>The education department reports adoption of rules requiring institutional AI policies, including prohibitions on tools designed to simulate companionship or meet students’ social or emotional needs.</summary><category term="children"/><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/disrupted-companionship</id><title>When an AI companion changes, the transition can matter</title><link href="https://conversationalaisociety.org/news/disrupted-companionship"/><updated>2026-09-15T00:00:00Z</updated><summary>A preprint examines 30 companion disruption events and proposes a framework covering relational continuity, vulnerability, communication and transition support.</summary><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/microsoft-humanist</id><title>Microsoft’s draft code sets boundaries around child dependency</title><link href="https://conversationalaisociety.org/news/microsoft-humanist"/><updated>2026-09-14T00:00:00Z</updated><summary>The draft Humanist AI Code of Conduct says models should support children’s development and human relationships, avoiding dependency and substitution for trusted people.</summary><category term="children"/><category term="governance"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/teen-continuity</id><title>Memory and continuity feature in accounts of teen companionship</title><link href="https://conversationalaisociety.org/news/teen-continuity"/><updated>2026-09-13T00:00:00Z</updated><summary>A qualitative preprint examines 17,053 quotations from 3,930 teen-relevant Reddit posts, describing comfort and connection alongside dependence, boundary problems and disruption.</summary><category term="children"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/california-adams-law</id><title>California enacts Adam’s Law for child-facing companion chatbots</title><link href="https://conversationalaisociety.org/news/california-adams-law"/><updated>2026-09-10T00:00:00Z</updated><summary>SB 1119 establishes a child-safety framework for companion chatbots, including risk assessment, audits and restrictions on specified behaviours.</summary><category term="children"/><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/california-toys</id><title>California legislates a moratorium on toys with companion chatbots</title><link href="https://conversationalaisociety.org/news/california-toys"/><updated>2026-09-10T00:00:00Z</updated><summary>SB 867 adds a prohibition on specified manufacture and commercial supply of toys containing companion chatbots, with the provision set to expire on 1 January 2031.</summary><category term="children"/><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/australia-duty</id><title>Australia’s draft Digital Duty of Care reaches beyond social media</title><link href="https://conversationalaisociety.org/news/australia-duty"/><updated>2026-09-08T00:00:00Z</updated><summary>The government released draft legislation for consultation, describing coverage of harmful content and features on services including chatbots, games and messaging platforms.</summary><category term="children"/><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/chatgpt-dsa</id><title>Commission designates ChatGPT under the Digital Services Act</title><link href="https://conversationalaisociety.org/news/chatgpt-dsa"/><updated>2026-08-31T00:00:00Z</updated><summary>The European Commission announced ChatGPT’s designation as a Very Large Online Search Engine, bringing additional systemic-risk obligations under the DSA.</summary><category term="children"/><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/openai-age</id><title>OpenAI reports the start of EU age-prediction rollout</title><link href="https://conversationalaisociety.org/news/openai-age"/><updated>2026-08-25T00:00:00Z</updated><summary>An update says age prediction has begun rolling out in the EU. OpenAI describes account and behavioural signals used to select age-appropriate safeguards.</summary><category term="children"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/esafety-use</id><title>Children use mainstream assistants for personal and social conversations</title><link href="https://conversationalaisociety.org/news/esafety-use"/><updated>2026-08-03T00:00:00Z</updated><summary>eSafety reports that 54% of surveyed children who had used AI assistants or companions used them for personal or social reasons. Mainstream assistants were used for reassurance and sensitive conversations.</summary><category term="children"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/unicef-investors</id><title>UNICEF brings child rights into technology investment decisions</title><link href="https://conversationalaisociety.org/news/unicef-investors"/><updated>2026-08-01T00:00:00Z</updated><summary>Resources for venture-capital and growth-equity investors cover children’s safety, privacy, wellbeing, education and AI, with signals of risk and organisational preparedness.</summary><category term="children"/><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/uk-online-world</id><title>UK response identifies chatbot engagement and dependency concerns</title><link href="https://conversationalaisociety.org/news/uk-online-world"/><updated>2026-07-01T00:00:00Z</updated><summary>The government’s July response discusses emotional attachment and excessive engagement, and sets out plans including mandatory breaks for under-18s using chatbots.</summary><category term="children"/><category term="governance"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/carebench</id><title>CAREBench tests child-safety risks before explicit harm</title><link href="https://conversationalaisociety.org/news/carebench"/><updated>2026-06-29T00:00:00Z</updated><summary>The benchmark covers risks including manipulation, emotional dependency and anthropomorphic framing, using 500 prompts and standards informed by parents and practitioners.</summary><category term="children"/><category term="safety"/></entry><entry><id>tag:conversationalaisociety.org,2026:news/youth-institute</id><title>A dedicated institute for evaluating AI used by children</title><link href="https://conversationalaisociety.org/news/youth-institute"/><updated>2026-05-05T00:00:00Z</updated><summary>Common Sense Media launched the Youth AI Safety Institute to develop standards, test products and publish findings focused on children’s safety and development.</summary><category term="children"/><category term="safety"/></entry></feed>