#130: Impact Signals #130 — FOIA Records Open Up Medicare's AI Prior-Authorization Pilot, Gates Foundation Pledges $1 Billion, Manhattan DA Seizes 12 Deepfake Domains
AI for Impact Daily Briefing, September 15, 2026
Top Stories
EFF Wins 1,000 Pages on Medicare's WISeR AI Pilot, Including a Request That Sat 83 Days Against a 72-Hour Promise
The Electronic Frontier Foundation obtained roughly 1,000 pages of documents and data from the Centers for Medicare and Medicaid Services through a Freedom of Information Act lawsuit, describing the launch of the WISeR model from the inside. WISeR, the Wasteful and Inappropriate Service Reduction model, began on 1 January 2026 across six states (Arizona, New Jersey, Ohio, Oklahoma, Texas and Washington) and runs through 2031, routing prior-authorization decisions for a narrow list of traditional Medicare services through AI-assisted review by contracted vendors. The program promised a 72-hour turnaround; EFF's analysis found requests that blew through it, including one that went unanswered for 83 days, with clinicians describing procedures postponed or cancelled while approvals sat in the queue. The sharpest detail in the records is about schedule rather than model quality: a WISeR vendor told CMS weeks before launch that its system would not be ready, and the agency went live on the original date. Why it matters: This is the first time the internal record of a live government AI deployment in health coverage has been forced into public view, and it locates the failure in procurement and schedule discipline rather than in the model. The 83-day figure against a published 72-hour commitment converts a policy dispute into a measurable service-level breach, which is the kind of evidence both the EFF litigation and a Congressional Review Act vote can act on. The program arrives at this moment already contested: the Government Accountability Office ruled earlier in 2026 that WISeR constitutes rulemaking rather than guidance, twenty senators signed a resolution in May 2026 seeking to end it, and the pilot is scheduled to run in those six states until 2031.
Gates Foundation Commits $1 Billion to AI Over Two Years, With $100 Million for Languages the Models Cannot Yet Speak
The Gates Foundation announced on Tuesday 15 September 2026, alongside its annual Goalkeepers report, that it will spend $1 billion over the next two years on AI for global health, education, agriculture and language access. The split is published rather than thematic: roughly $400 million for healthcare applications, roughly $400 million for education tools, roughly $100 million for agricultural technology, and roughly $100 million for building language datasets in communities the major models currently serve poorly. Named partners come with the pledge, including OpenAI, which has committed $50 million toward training health workers in Rwanda, Anthropic on vaccine development and crop data, Google.org, and Microsoft's AI for Good Lab. The framing in the Goalkeepers report is defensive: the foundation projected that 2025 would be the first year this century to show an increase in child deaths, a reversal it attributes largely to cuts in global health funding. Why it matters: This is the largest private AI-for-development pledge on record, and committing it to published proportions rather than a theme makes the Rwandan health-worker programme a checkable first deliverable instead of an intention. The $100 million language line is the most revealing item inside the pledge, because it buys an input the frontier labs have not found commercially worth building and concedes that the models as they exist do not work for the populations the other $900 million is meant to serve. The foundation has published no application route or deadline for the commitment, so it is a funding programme rather than an open call as of today. Details at geekwire.com.
India's 1033 Highway Helpline, 15,000 Calls a Day Across 22 Official Languages, Goes Out to Tender for AI
The National Highways Authority of India issued an Expression of Interest on 10 September 2026 to rebuild its 1033 National Highway helpline around AI, real-time caller location tracking and automated emergency dispatch, with bids due by 26 October 2026. The helpline launched in 2018, runs 24 hours a day, covers more than 146,000 km of national highways, and takes about 15,000 calls per day. The three capabilities map onto three specific failure points: location tracking for callers who cannot describe where they are on a highway, multilingual handling for a call centre serving a country with 22 official languages where caller and operator frequently do not share one, and smart dispatch for deciding whether an ambulance, a crane or a patrol vehicle should be sent. NHAI has stated that human operators stay in the loop, particularly in emergencies where judgement is required. Why it matters: At 15,000 calls a day this is one of the largest single emergency-call surfaces anywhere being tendered for AI in the open, with the requirements published rather than negotiated privately. The unusual clause is the multilingual one, which makes language coverage a procurement condition in a country where the barrier between caller and operator is a routine part of the failure mode rather than an edge case, and the human-in-the-loop carve-out is written into the tender rather than added afterward. Nothing is running yet: this is a procurement document with the bid window open until 26 October 2026, at nhai.gov.in.
Manhattan DA Seizes 12 Domains Built on AI Images of Roughly 1,200 Real People
Manhattan District Attorney Alvin Bragg announced on 14 September 2026 that his office had seized the domain names of 12 websites used to publish and sell non-consensual sexual deepfake videos, which the office describes as the largest known seizure of AI-generated deepfake websites to date. The domains were taken under court orders as part of an ongoing investigation by the office's Cyber Crime Bureau. The scale sits in the victim count rather than the site count: individuals under investigation are alleged to have used AI image and video generation tools to convert existing photographs and footage of approximately 1,200 real people into synthetic sexual material, overwhelmingly women and primarily public-facing, including actors, politicians, athletes, musicians, social media influencers and activists. Why it matters: The mechanism is notable precisely because it is ordinary: no new AI statute was invoked, and domain seizure under court order, a tool that predates generative models entirely, reached synthetic-media harm at a documented scale of roughly 1,200 identified people across 12 sites. The victim profile the office published is a dataset about who this technology is actually aimed at rather than an estimate. Because the investigation is described as ongoing and the office has asked people who believe they were depicted to contact the Cyber Crime Bureau directly, the identified-victim count is a floor rather than a total.
Leaked EU Commission Draft Would Put AI Chatbots Behind the Same Age Gate as Social Media for Under-15s
A European Commission document obtained by RTE, Euronews and Reuters sets out an "EU Kids Act" that would restrict under-15s from social media, video-sharing platforms, online games and AI chatbots. RTE reported the draft on 14 September 2026 and named Commission President Ursula von der Leyen and EU tech chief Henna Virkkunen as presenting it on Thursday 17 September 2026. The draft is tiered rather than a single cut-off: no access under 3, fully parent-controlled accounts limited to child-friendly services for ages 3 to 12, parent-opened introductory accounts with limited contacts and strict time limits at 13 and 14, and self-opened accounts from 15. Platforms would have to verify children's ages, avoid addictive design patterns, provide reporting tools for harmful content and implement parental controls, and would pay a supervisory fee that funds the regulators enforcing the rules. Why it matters: This is the first major regulatory text to place AI chatbots in the same age-gating category as social media rather than treating them as a software product, which changes what compliance means for any conversational AI reachable from the EU. The industry-funded supervisory fee is the structural detail, because it addresses the enforcement-capacity gap that has limited earlier EU digital rules by funding regulators from the firms they oversee. This is a leak and not yet a proposal: the 17 September presentation is the date on which these thresholds either survive into an official text or change.
Upcoming Events & Opportunities
European Commission Presentation of the "EU Kids Act"
- September 17, 2026
- Location: Brussels, European Commission
- The date on which the leaked under-15 thresholds, including the AI chatbot age gate, either survive into an official proposal or change
- Register: ec.europa.eu
Humanity AI Open Call (Funding)
- Amount: $10 million total; individual grants $75,000 to $1,000,000, projects up to three years
- Deadline: October 21, 2026, 4:59 pm PT / 7:59 pm ET
- Eligibility: US-based nonprofits as lead applicant, proposals focused on US-based communities; non-US organisations may join as collaborators but not lead. Aligned to education, labor and economy, and the humanities, arts and culture.
- Pooled fund co-led by the Ford Foundation, MacArthur Foundation, Omidyar Network and seven other US foundations
- Apply: humanityai.ai
NHAI 1033 Helpline Expression of Interest, Bid Submission Closes
- Expression of Interest issued September 10, 2026
- Deadline: October 26, 2026
- Location: India (national)
- The open bid window on an AI rebuild of a helpline handling roughly 15,000 emergency calls a day
- Register: nhai.gov.in
ASEAN Committee on Disaster Management Ministerial Meetings, Manila
- Convened the week of September 14, 2026; closing date not established
- Location: Manila, Philippines
- Philippine officials pressed ASEAN toward anticipatory action and improved warning systems
- Register: pna.gov.ph
Active Disaster Monitoring (GDACS/OCHA)
- China:** Flooding, GDACS Orange alert, onset July 31 with the last update September 7. 4 deaths and 318,341 people displaced as of that update; labelled ongoing. Source: gdacs.org.
- Indonesia:** M6.5 earthquake 126 km NNE of Teluknaga on September 11, USGS significance 657. PAGER alert green, meaning low expected casualties and damage given the offshore location and distance from dense population. No major response reported. Source: earthquake.usgs.gov.
- Liberia:** Floods with onset in August 2026, listed as an ongoing situation on ReliefWeb. Casualty and displacement figures are not established from the feed. Source: reliefweb.int.
- Note: only major or ongoing-major events are carried; low-severity alerts and preparedness exercises are excluded.
Sources: See individual stories above for full attribution.