Technology. Human experience. The conversations in between.
News & safety / The archive
Earlier coverage. Still worth reading.
A monthly library of previous briefings, with original sources and reading cautions. Dates refer to the source or stated event; status labels describe the briefing at the time it was written.
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 19 of 19 archived briefings, grouped by month.
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.
Why it matters
Relevant context for conversational products with medical functions and organisations evaluating AI for patient care.
Read with care
Policy response and implementation plans, not 44 newly operative legal duties. Many next steps are scheduled for further detail by spring 2027. The document does not establish that every companion or wellbeing chatbot is a medical device; applicability depends on the product and regulatory scope. Publication record and selected response provisions checked, not a complete legal implementation analysis.
What was checked
Source record and stated publication date checked; see reading cautions for the extent and limits of the review. Checked 2026-10-12.
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.
Why it matters
A direct development in governance of relational design, rather than only prohibited outputs.
Read with care
Introduced legislation, not enacted obligations. This describes the introduced version, not a complete legal analysis or a guarantee of legislative progress. Claimed benefits are not evidence that the proposed safeguards work.
What was checked
Sponsor announcement, 1 October introduction and default-setting provisions in the introduced H.R. 10719 text checked. Checked 2026-10-12.
A study in which young people built their own chatbots found differences by age in how they understood confidence, transparency and trustworthiness.
Why it matters
Trust calibration deserves attention alongside answer accuracy. A confident presentation may mean something different to a younger user.
Read with care
A chatbot-design study, not a population survey or a test of long-term companion safety. The arXiv record is a preprint; peer-reviewed acceptance was not verified.
The Commission’s proposed child-safety framework includes conversational services within its coverage, alongside other online services used by minors.
Why it matters
It places chatbot design within a broader debate about age-appropriate access, protective defaults and platform responsibility.
Read with care
A proposal, not enacted law. Parliament and Council must consider the text. This briefing summarises the Commission’s announcement; it does not establish final obligations or commencement dates.
What was checked
European Commission announcement checked. Detailed legal-text review remains outstanding.
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.
Why it matters
Schools are distinguishing instructional assistance from relational design in their procurement and use policies.
Read with care
This summarises the department’s announcement. It is not a ban on all educational AI; precise scope, implementation deadlines and exceptions require the rule text.
What was checked
Dated Florida Department of Education announcement inspected.
A preprint examines 30 companion disruption events and proposes a framework covering relational continuity, vulnerability, communication and transition support.
Why it matters
Model changes, persona changes and shutdowns deserve consideration in a system’s safety evaluation—not only its normal operation.
Read with care
The analysis uses community-level Reddit data and reports associations. It is not child-specific and does not establish individual clinical outcomes or causal effects on children.
What was checked
Authors’ abstract and submission history checked. This is a source note, not a full critical review.
The draft Humanist AI Code of Conduct says models should support children’s development and human relationships, avoiding dependency and substitution for trusted people.
Why it matters
It gives designers and evaluators a concrete example of intended relational boundaries, beyond prohibited-content rules.
Read with care
A provider’s proposed behaviour standard is not independent evidence that its systems consistently meet that standard. The document was published for consultation.
What was checked
Draft status, publication date and child-safety section checked.
A qualitative preprint examines 17,053 quotations from 3,930 teen-relevant Reddit posts, describing comfort and connection alongside dependence, boundary problems and disruption.
Why it matters
It identifies questions for evaluating relationships across conversations: what persists, how boundaries change, and what happens when the system changes.
Read with care
Teen-relevant posts are not verified adolescent participants. These accounts cannot establish prevalence or causation. The record states that the paper is under review for CHI 2027.
What was checked
Authors’ abstract, submission date and review status checked. A fuller source note is available in Research.
SB 1119 establishes a child-safety framework for companion chatbots, including risk assessment, audits and restrictions on specified behaviours.
Why it matters
It makes ongoing system design and assurance part of the regulatory conversation, alongside the content of individual replies.
Read with care
Signed on 10 September does not mean every duty applies immediately. Audit deadlines, exemptions and conditional provisions require separate attention; the text includes an initial audit deadline in 2029 and a revenue-based exemption before 2032.
What was checked
Chaptered text, signing date and audit provisions inspected. This is a news summary, not an implementation checklist.
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.
Why it matters
Relational capability becomes relevant to product classification, extending the discussion beyond chat windows and apps.
Read with care
The statutory definitions and commencement rules matter. This does not mean all AI-enabled toys are covered, or that all provisions became operative on the signing date.
What was checked
Chaptered text, signing date and sunset provision checked.
The government released draft legislation for consultation, describing coverage of harmful content and features on services including chatbots, games and messaging platforms.
Why it matters
The proposal encourages attention to risks created by a service’s design and features, as well as individual pieces of content.
Read with care
An exposure draft is not an enacted duty. The announcement and consultation page establish the proposal; they do not demonstrate that a particular mitigation works.
What was checked
Dated ministerial announcement and official exposure-draft page checked.
The European Commission announced ChatGPT’s designation as a Very Large Online Search Engine, bringing additional systemic-risk obligations under the DSA.
Why it matters
The announcement explicitly connects the regime with risks to minors, physical and mental wellbeing, and fundamental rights.
Read with care
Designation is not a finding that a service caused harm. The announcement provides a compliance period; this briefing does not determine which duties are already operative.
What was checked
Commission announcement and its 4 September update checked.
An update says age prediction has begun rolling out in the EU. OpenAI describes account and behavioural signals used to select age-appropriate safeguards.
Why it matters
Age uncertainty becomes part of product behaviour, with a safer default when information is incomplete.
Read with care
A company’s rollout report is not an independent evaluation of accuracy, privacy impact or protection against bypasses. The underlying article was first published in January.
What was checked
Dated EU update and explanation of age prediction checked.
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.
Why it matters
Companion-like use can emerge in a general-purpose assistant, so product labels alone are an incomplete guide to risk.
Read with care
The percentage refers to children who had used those tools, not all children. Self-reported use does not establish that AI caused harm or benefit.
What was checked
Regulator’s dated release and stated survey population checked. The complete methodology was not assessed for this note.
Resources for venture-capital and growth-equity investors cover children’s safety, privacy, wellbeing, education and AI, with signals of risk and organisational preparedness.
Why it matters
Safety questions belong in funding and governance decisions as well as product development.
Read with care
Guidance is not binding law or a certification of an investment. Detailed checklist claims from the watch are not reproduced without reviewing the downloadable primers.
What was checked
UNICEF publication page, intended audience and month checked.
The government’s July response discusses emotional attachment and excessive engagement, and sets out plans including mandatory breaks for under-18s using chatbots.
Why it matters
It provides UK context for thinking about conversation length and relational design as policy concerns.
Read with care
A policy announcement is not proof that all measures are in force. Consultation responses are not a representative prevalence study. The watch’s 19 August date was a page-update date; this links to the July response.
What was checked
Government response’s chatbot section checked; implementation and commencement were not established in this note.
The benchmark covers risks including manipulation, emotional dependency and anthropomorphic framing, using 500 prompts and standards informed by parents and practitioners.
Why it matters
It broadens evaluation beyond explicit prohibited content to the ways a response can enable or deepen a risky interaction.
Read with care
The dataset specifies single-turn prompts. This is not a longitudinal or multi-turn safety evaluation, and performance on it cannot establish real-world child safety.
What was checked
Paper abstract, publication date and authors’ dataset documentation checked.
Common Sense Media launched the Youth AI Safety Institute to develop standards, test products and publish findings focused on children’s safety and development.
Why it matters
Youth-specific evaluation creates another source of scrutiny alongside providers’ own safety claims.
Read with care
The launch was in May, although it surfaced in the August watch. An institute’s existence is not evidence that a particular product is safe; inspect each assessment’s methods and disclosures.