Filters work together. Children and Adults both include briefings affecting both groups. “All users” also includes updates with no specified age group. Region follows the scope of the briefing; unspecified does not mean worldwide.
Showing 10 of 10 selected briefings, newest first. Explore the archive for earlier coverage.
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.
Why it matters & reading cautions
Why it matters
A practical resource for teams assessing conversational products, including the difference between testing a model and testing the system people actually use.
Read with care
Provider guidance illustrated by internal evaluations, not independent validation or clinical outcomes. Its API and claude.ai results compare different configurations; they should not be presented as a controlled before-and-after test of one safeguard.
What was checked
Primary source and publication status checked; earlier method checks and limits are described in the reading cautions. No independent replication. Checked 2026-11-15.
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.
Why it matters & reading cautions
Why it matters
Relevant to how companion products encourage engagement, charge users and support meaningful choices about continued use.
Read with care
Legislative proposal, not current enforceable duties. Parliament and Council must consider it. Article 17 and recital 66 contain scope distinctions; references to proposed EU KIDS Act protections do not mean that proposal is enacted.
What was checked
Official proposal record rechecked; Article 17 and recital 66 examined during the 12 November watch review. Checked 2026-11-15.
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.
Why it matters & reading cautions
Why it matters
Relevant to schools reviewing chatbot products and developers building for learners.
Read with care
An open consultation, not an operational certification scheme or new enacted duties. The accessible landing page does not establish final standards, mandatory scope, implementation date or a response deadline. Its stated scope is England.
What was checked
Official consultation landing page and stated scope checked. Checked 2026-11-15.
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.
Why it matters & reading cautions
Why it matters
Conversation privacy depends on the whole supplier chain. Sensitive details and credentials inside messages can matter even when account labels are absent.
Read with care
Provider disclosure and ongoing investigation, not independent forensic findings. Most affected transcripts reportedly lacked user and workspace labels; that does not make their contents anonymous. Key revocation occurred on 11 November; third-party issuers were notified on 12 November, reported on 13 November.
What was checked
Notion’s advisory, incident timeline and updates through 13 November checked. Checked 2026-11-15.
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.
Why it matters & reading cautions
Why it matters
Safety can depend on when information arrives and whether a chatbot revises its first impression.
Read with care
Preprint using artificial cases and a simulated user, not real patients or measured health outcomes. Percentages concern emergency cases in the experiment, not all chatbot users. Separate branches generated triage labels and questions, unlike a normal consumer interface.
What was checked
Primary source and publication status checked; earlier method checks and limits are described in the reading cautions. No independent replication. Checked 2026-11-15.
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.
Why it matters & reading cautions
Why it matters
A close fit with long-horizon companion research: retaining a detail and using it appropriately are different capabilities.
Read with care
Preprint using one assistant model and an AI user simulator, with automated judging and partial human validation. Designed recovery trajectories and a small donor sample limit generalisation. No real-user health outcomes were measured; the findings are not a case for intervening more often by default.
What was checked
Primary source and publication status checked; earlier method checks and limits are described in the reading cautions. No independent replication. Checked 2026-11-15.
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.
Why it matters & reading cautions
Why it matters
Tests should include how people respond to a safeguard, not just whether the first answer sounds safe.
Read with care
Preprint, not verified peer review. Simulated users aged 13–17 and conversations of at most 20 exchanges do not establish real adolescent outcomes or months-long safety. Automated judging has limited human calibration. Criterion-failure rates are not percentages of users harmed. API configurations do not reproduce every consumer product.
What was checked
Primary source and publication status checked; earlier method checks and limits are described in the reading cautions. No independent replication. Checked 2026-11-15.
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.
Why it matters & reading cautions
Why it matters
A practical distribution-policy change for companion developers and a useful distinction between age ratings and access safeguards.
Read with care
App Store policy, not a worldwide statute or a blanket restriction on every AI companion. An age rating alone does not establish effective age verification or prevent underage access. Regional rating rules and exceptions require separate checking.
What was checked
Apple announcement, scope and declaration dates checked. Checked 2026-11-15.
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.
Why it matters & reading cautions
Why it matters
Useful for conversational systems that move beyond replies to making purchases, communicating or taking other actions for users.
Read with care
Joint statement of regulatory approach, not a new law or enforcement finding. Duties depend on the service and activity; every chatbot is not automatically covered by every listed regime.
What was checked
Official announcement rechecked; full three-page statement examined during the 1 November watch review. Checked 2026-11-15.
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.
Why it matters & reading cautions
Why it matters
Adds accounts of actual use to a field often assessed through simulations, while separating perceived helpfulness from proven benefit.
Read with care
Peer-reviewed observational study, not a clinical trial. Data were collected in February–May 2025, mainly in Oxfordshire and Liverpool, using a non-probability school sample. Some participants were over 18. These results are not national prevalence, current 2026 usage estimates or evidence of clinical benefit or safety.
What was checked
Publisher article, abstract and methods inspected, including collection dates, sampling and participant ages. Checked 2026-11-15.