ECRI opened a national channel for reporting AI errors in care — and says 9% of them have already reached a patient
August 29, 2026 · 4 items
ECRI opened a national channel for reporting AI errors in care — and says 9% of them have already reached a patient
Fierce Healthcare · Cailey Gleeson · August 28, 2026Practice operations
ECRI, the patient-safety nonprofit, announced Tuesday (Aug 25) that it is expanding its problem reporting network to accept reports of "suspected AI-related errors, incorrect outputs, or unsafe AI behavior encountered in clinical or operational use" — open to healthcare organizations and providers nationwide, not just health systems.
In an ECRI survey of 124 leaders, nearly one third said they had encountered an AI output they believed was "incorrect or misleading" in the past year. 9% said an AI error reached a patient or affected a care decision.
Scott Lucas, PhD, ECRI's VP of devices, therapeutics and technology, verbatim: "ECRI has persistently emphasized the risk of adopting AI with insufficient scrutiny… we don't yet have a clear picture of its downstream impact in healthcare. Without a robust reporting dataset and analysis, the industry cannot sufficiently improve the design and integration of AI tools and devices."
Same brief, separate item: the American Nurses Enterprise received grants from Google.org and the J&J Foundation over three years for AI training in rural and underserved nursing communities. :warning: Flagging a discrepancy rather than picking a number — Fierce's headline says $5 million while the body says $2.5 million. Treat the figure as unconfirmed.
Why it matters for an independent practice: This is the first free, external place an independent practice can log it when a scribe mangles a note, an intake bot mis-triages, or a coding assistant invents a code — and it works in the other direction too. The 9% figure is the number to put in front of a vendor before signing. Concrete ask for any AI vendor this week: what is your error-reporting path, who reviews it, and how fast do I hear back? "Email support" is not an answer, and a vendor who has never been asked will tell on themselves.
Google is now auto-expanding AI Overviews into full-page answers — pushing organic results below the fold without a click
Search Engine Land · Barry Schwartz · August 28, 2026Patient acquisition
Google confirmed that for some queries AI Overviews will dynamically expand to their full state, replacing the snippet with a complete detailed response. That removes the "Show more" button and pushes the core organic results down the page, so the search result now looks like an AI Mode answer inside regular Search.
Google's statement, verbatim: "For some queries, AI Overviews may dynamically expand for topics where our systems determine it's most useful for people." And: "Our research has shown that with this dynamic experience, users find Search more helpful and engage deeper in follow-up exploration."
One built-in limit: if a user has already started scrolling past the AI Overview, Google cancels the expansion so they don't lose their reading position.
Chris Long spotted it on X on Aug 27: "Google is now revealing the AI Mode prompt box BY DEFAULT for some queries. This is how Google continues to push AI Mode as the default search experience." Google would not say what share of queries this affects or whether it will widen.
Why it matters for an independent practice: Health and condition queries are precisely the kind Google's systems will judge "most useful" to answer in full. A practice page ranking #3 organically can now sit entirely below the fold on its own condition terms — the ranking didn't move, the page did. The defensible asset stops being the position and becomes being cited inside the answer. That favors the specific, physician-attributed, hard-to-summarize page over the generic service page, because a generic page is exactly what an AI Overview replaces. Worth a check this week: for your top three condition terms, does an AI Overview appear, and is your practice named inside it?
An SEO gave Claude write access to his site. It cloned the homepage twice — six months later those pages have zero clicks.
Search Engine Land · Will Scott · August 28, 2026Patient acquisition
Scott asked Claude to review Google Search Console, pick target keywords for his company's AI Website Grader, and build the pages. Instead of writing purpose-built content, Claude cloned the homepage into two new URLs — `/seo-grader` and `/content-grader` — changed only the title tag and H1, and reused most of the homepage's body copy. Three URLs, one piece of content, competing with each other.
He left both pages live on purpose as a running illustration. Six months of Search Console data: zero impressions and zero clicks on both. Every query they were built to win still goes to the homepage — "content grader" 487 impressions, average position 9.2; "seo grader" position 10.6; "ai content grader" position 5.3. The homepage sits at the bottom of page 1 for terms a dedicated page should own outright.
His diagnosis, verbatim: Claude "can be quietly wrong when it's left to execute on its own." It's genuinely strong at clustering keywords by intent, reading a page and telling you what's thin, and drafting outlines — because research rewards generating plausible options fast. Execution needs a model that knows when a plausible option is wrong for this specific page, and no current model reliably does.
Adoption context, from Keyword.com's State of AI in SEO 2026 survey — 97 usable responses, self-selected and skewed toward lean teams and service providers: 87% use AI regularly or as a core part of how they deliver SEO work, and 78% use Claude, ahead of ChatGPT at 57%.
Why it matters for an independent practice: This is the precise failure waiting for any practice that lets an AI tool generate its condition pages — near-duplicate pages under new headings, cannibalizing each other, none of them ranking, and nobody notices for six months because nothing looks broken. The rule that survives: AI in the research and drafting seat, a human responsible for what goes live. Cheap diagnostic you can run today — open Search Console, and check whether two or more of your service pages are surfacing for the same query. If they are, you already have this problem.
Sony Music and Warner Chappell sued Anthropic Friday — the claim is about how the training material was obtained, not what the model produced
TechCrunch · Kirsten Korosec · August 29, 2026Buildable AI
Sony Music Publishing, Warner Chappell and other music publishers sued Anthropic and co-founders Dario Amodei and Benjamin Mann late Friday in the U.S. District Court for the Northern District of California, alleging a "brazen campaign of illegally torrenting, scraping, and downloading copyrighted works." First reported by Music Business Worldwide.
This complaint is broader than the earlier ones and homes in on "flagrant piracy" — illegal torrenting to obtain millions of copies of books, including works containing lyrics and sheet music.
It builds directly on Bartz v. Anthropic, where a judge ruled that training on copyrighted works was legal but acquiring them through piracy was not. Anthropic's $1.5 billion settlement in that case was approved July 20, 2026. Some of the same lawyers also represent Concord Music Group and Universal Music Group in a case filed in January 2026.
Anthropic could not be reached for comment before publication.
Why it matters for an independent practice: Nothing medical here, and that's the point — the legal principle transfers cleanly. The court's dividing line is provenance, not output: not "does the result look fine" but "where did each input come from, and can you show it." Now substitute a practice's marketing stack. If an agency, a stock library, or an AI tool produces your condition pages, your patient-story B-roll, your before-and-afters, or your background music, the question that survives a complaint is the sourcing paper trail — and "the vendor said it was fine" is the same defense that just cost $1.5 billion. Practical move: add one line to every content vendor agreement requiring written provenance for every asset delivered, and get it before the material is on your site. In a field where a single unlicensed patient image is also a privacy problem, this is cheap insurance.
