Top 10 AI News This Week — September 14–20, 2026

This was the week policy stopped being hypothetical. Florida made AI policies mandatory at every one of its 28 public colleges, California wrote a community college AI retraining pilot into law while a second bill that binds every district sits nine days from becoming law without a signature, and OpenAI put a firm end date on the custom GPTs that thousands of faculty built their courses around. Here's what community college educators should know.

EdTech & Campus AI

Florida now requires an AI policy at all 28 of its public colleges

On September 16, meeting at Polk State College in Winter Haven, the Florida State Board of Education approved new Rule 6A-14.0719 for the Florida College System, along with an amended K-12 internet-safety rule. We previewed the vote last week. It passed.

Each college's board of trustees must adopt a policy that sets academic-integrity expectations (AI off-limits on graded work unless the instructor allows it), names approved tools, and protects student privacy and intellectual property. According to the Department's release, colleges must also notify parents when AI instructional tools are approved, which matters for dual-enrollment minors. The policies must bar AI built for social or emotional companionship and undisclosed behavior monitoring, and must stop vendors from selling student data or training models on it. K-12 districts face a July 1, 2027 deadline. No college adoption deadline has been published, and Florida's rule registry still lists 6A-14.0719 as not yet filed, so it isn't formally in effect yet. Commissioner Henry Mack: "An educated person can use a powerful tool without becoming dependent on it."

For community colleges outside Florida, this is the first statewide rule to require an AI policy at every institution in a two-year system. California districts revising board policies and administrative procedures now have a rule to compare their drafts against. The companion-AI ban and the no-training clause for vendors are the pieces worth borrowing. Read more →

California writes a community college AI retraining pilot into law, but funds none of it yet

On September 18, Governor Newsom signed AB 2504 (Bauer-Kahan), now Chapter 305, Statutes of 2026. It creates the Future of Creative Industries Pilot Program, run by the California Community Colleges Chancellor's Office. Up to ten colleges, spread across northern, central and southern California, would retrain film, television, music, touring and live-events workers whose jobs generative AI has displaced or put at risk. The bill names sound engineers, voice actors, concept artists and entry-level employees. Eligibility criteria are due June 30, 2027, implementation begins July 1, 2028, and the program reports completions, credentials, job placement and demographic outcomes to the Legislature by January 1, 2032, before sunsetting in 2033. The catch: it becomes operative only when the Legislature appropriates money for it. It passed 78-0 and 39-0.

The bill we led with last week is still unresolved. As of the September 20 legislative update, AB 2392 (Fong), which sets GenAI procurement and training standards for the CCC and CSU, had been neither signed nor vetoed. If it is not returned by September 30, it becomes law without a signature. On the same day he signed AB 2504, Newsom issued Executive Order N-9-26 on independent audits of frontier AI labs. It matters for the broader context but has no direct campus effect.

If you teach media arts, cinema, music or design, AB 2504 is the first state law to name your students' industries as places AI is displacing workers. Program chairs in Los Angeles-area districts should get on the Chancellor's Office's radar now, before the eligibility criteria are written. Read more →

California's community college chancellor bets on faculty rather than a vendor contract, and two other systems show what that looks like

Speaking on a Dreamforce panel in San Francisco on September 16, Chancellor Sonya Christian said faculty must rethink teaching and assessment now that AI can write essays and solve equations. "We're trying to bring a whole system along," she said, in coverage EdSource published September 17. EdSource contrasted the systems' approaches. CSU pays OpenAI about $13 million a year for a customized ChatGPT. The community colleges, serving more than two million students at 116 colleges, have put their money into an AI ambassador program that gives faculty yearlong sabbaticals for AI projects that serve their own colleges and the system.

Two other systems showed the same approach this week. On September 18, Wake Tech in North Carolina said students in 33 disciplines took AI coursework over the past year, integrated into existing programs instead of a standalone AI degree. The University of Hawaiʻi Community Colleges opened registration for their free five-week asynchronous faculty course, AI in Teaching and Learning (register by October 7, starts October 26). The course has reached 173 faculty since summer 2025, and participants' self-reported confidence with AI tools rose from 33.8% to 88%. The Hawaiʻi course is open only to UH faculty. The model is the useful part: a professional-development committee could adapt it for flex days. Read more →

Two AI court sanctions, and the second one isn't about fake citations

The written order in the New Mexico ChatGPT-brief case became public on September 14. It is State v. Oscar Renee Sandoval; In re Stephen D. Aarons, New Mexico Supreme Court No. S-1-SC-40845, dated August 21, 2026. Many outlets reported it as a September 11 ruling, but that is only when the story went national. The court found attorney Stephen Aarons in direct contempt for a brief containing testimony from witnesses who do not exist, ordered $5,000 paid to the State Bar Client Protection Fund, struck his briefing, referred him to the Disciplinary Board, and barred him from appearing before the court until that investigation concludes. A public defender was appointed September 2. No discipline outcome has been reported yet.

The second case teaches something different. As reported September 18, a Florida appellate panel in Lisandrillo v. Palozzi called counsel's AI-generated filings "gibberish" while noting that the citations were real. It ordered the attorney to show cause within ten days why she should not be sanctioned, with a possible Bar referral.

Paralegal and administration-of-justice faculty can teach these two as a pair. The first is the familiar failure, fabrication. The second is newer: every citation checks out and the filing is still useless. A citation checker catches the first. Only a human who reads the argument catches the second. Writing instructors can make the same point in a first-year composition course. Read more →

Major AI Companies

OpenAI is retiring custom GPTs on every plan, and course GPTs stop working December 11

OpenAI is retiring custom GPTs across all ChatGPT plans, and its help center lays out the schedule. Admins were notified September 11. Migration tools and an in-product banner are targeted for September 22. Creating new custom GPTs is planned to end October 26, and existing GPTs stop running December 11. Each GPT converts to a "plugin." Its instructions become a skill, its knowledge files become reference files, and its connected apps carry over. The selected model does not carry over, custom actions must be rebuilt by hand, and old conversations stay with the original GPT.

Sharing is what will hurt faculty. OpenAI says a migrated plugin "starts private" and that migration "doesn't carry over your GPT's sharing settings." A course GPT that thirty students use through a shared link will not reach them automatically once converted. The instructor has to re-share it, and each student may have to separately authorize any connected apps.

This announcement came September 11, just before this week's window. We're running it now because the deadlines land inside the fall term. If you built a tutor, a writing coach or a study bot as a GPT, plan to migrate in October, not December: test the plugin, re-share it, and tell students before finals. If you support faculty, a short migration workshop now will save a lot of help-desk tickets at the end of the semester. Read more →

Gemini's study notebooks arrive in school Workspace accounts, including lecture recording

On September 17, Google announced that Notebooks in Gemini, a dedicated workspace where a student uploads course materials (up to ten sources) and gets study guides, practice quizzes and video overviews, is available to "all Google Workspace customers," outside the European Economic Area. The rollout began September 14 and should reach both Rapid and Scheduled Release domains within about 15 days. It is on by default. Admins control it through the existing Gemini app and Gemini Notebook settings, and both must be enabled for a user's organizational unit.

A companion post on September 18 added back-to-school features. The one faculty should know about is lecture recording in the mobile app. Others include voice conversation with a notebook in about 100 languages (18+ only), reports that bundle summaries with quizzes, flashcards and mind maps, and short video overviews in 80+ languages.

At a Workspace for Education campus this will simply show up for students this month, with no action by the college. Two things to settle now. First, whether your syllabus or district policy says anything about students recording lectures, because the app now makes that one tap. Second, whether your Workspace admin has made a deliberate choice about the default. The study-aid features themselves are useful and worth showing students. Read more →

AI assistants move into the document itself: Claude Docs and Slides, and ChatGPT inside Word

Two releases this week shift AI from a separate chat window into the document. On September 16, Anthropic merged Claude's Cowork and chat into one app and added Claude Docs and Claude Slides in beta. Claude now writes and edits shareable documents and decks inside a conversation, and decks export to PowerPoint or PDF. It is rolling out to Pro and Max over the coming weeks, with Team and Free plans "coming soon." The announcement does not mention education plans.

On September 17, OpenAI released ChatGPT for Word, a sidebar add-in for drafting, summarizing, revising and reformatting. OpenAI states it "is available on all ChatGPT plans, including Free," and says the add-in will be enabled by default starting October 1. Organizational Microsoft 365 tenants still need admin approval. OpenAI's own help page warns that it can make errors and "may unintentionally alter content," and add-in conversations don't sync with the user's ChatGPT history.

If you teach writing, the free tier is the part to plan around. Any student with a free ChatGPT account and Word can now revise an essay inside the assignment file, leaving no separate chat log. Integrity policies that rely on students disclosing a chat transcript should be revisited. Assignments that grade the drafting process, such as version history, conferences and in-class drafts, hold up better than ones that grade only the final file. Read more →

For API coursework: new real-time voice models, a breaking change due October 5, and Grok 4.7 finally ships

Three API changes this week matter to anyone teaching with APIs. On September 15, Google released Gemini 3.8 Live and 3.8 Live Extended Thinking, real-time voice models that switch among 97 languages mid-conversation and watermark their audio with SynthID. Both are available now in the Gemini API and AI Studio. Base Live is free to consumers through the Gemini app. Google also replaced its Antigravity preview agent on September 17 with a version that changes tool parameters, and the old version shuts down October 5. Assignments built on it will break mid-semester unless they're updated. On September 18, xAI released Grok Voice Transcribe 2.0, claiming twice the accuracy at unchanged prices of $0.10 per audio hour batch and $0.20 streaming, with speaker labels. That makes it one of the cheapest captioning options for student projects.

An update on a thread we have followed for three weeks: Grok 4.7 shipped September 21, one day after this window, at $2 per million input tokens and $6 per million output tokens. xAI's announcement gives no parameter count. The rumored 2.5-trillion-parameter "Grok 4.8" that circulated around September 13 did not appear, so the 4.8 reports we flagged as unverified last week remain unverified.

For computer science instructors, the Antigravity deadline is the one that needs action this week. For ESL and world-language faculty, a free, multilingual, real-time voice partner is now a reasonable thing to pilot. Read more →

Free & Student Opportunities

Google's free AI educator badges can now turn into college credit, some of it free

On September 18, Google announced that badges from its free Google AI Educator Series can now be converted into college credit through three partners. Each module takes about 15 minutes, and new modules come out the first Wednesday of each month. The terms differ, and only one option is stated as free:

Undergraduate credit through College Unbound is "at no cost," fully accredited, and can count toward finishing a bachelor's degree or be converted to CEUs. One secondary source says this option opens later in 2026, so check before you promise it to anyone. Graduate credit through ISTE and Dominican University has no stated cost, so do not assume it's free. License-renewal hours through the Illinois Digital Educators Alliance are for Illinois educators only. The series' September modules include "Planning the Inquiry," which uses Gemini Deep Research to teach student research skills and is aimed at higher ed. The live Badge-a-thon was September 19, but the self-paced badges are still available.

For community college faculty, this is vendor training, and the modules teach Google's tools. Still, a no-cost path from a 15-minute module to accredited undergraduate credit is rare. It could help adjunct faculty, classified staff finishing a degree, and instructional aides. Pair it with something platform-neutral if you're building a PD sequence. Read more →

AI Competitions

Closing within ten days: an all-cash IEEE hackathon and an AI-and-law build

OneAquaHealth IEEE Global Hackathon. Eligibility: students only, at or above the legal age of majority, teams required (no individual entries), companies excluded, all countries except the standard exceptions. Nothing on the page excludes community college students. Deadline: September 30, 9:00 p.m. PDT. Prize: $3,500, all cash ($1,500 / $1,000 / $500 plus two $250 special mentions), plus IEEE certificates. Free to enter. Its seven tracks, from an EU-funded freshwater-health project, include "AI-Supported Assessment" and "Data-to-Insight," a good fit for a biology or environmental-science student paired with a programmer. Enter here.

LexHack 2026. Eligibility: students only, at or above the legal age of majority, companies excluded, all countries except the standard exceptions. Deadline: September 27, 5:00 p.m. EDT. Prize, stated honestly: advertised $74,248, of which about $5,548 is cash ($1,816 / $1,456 / $1,276 plus $1,000 in referral awards). The rest is credits, subscriptions and domains the page itself labels "THIS IS NOT A CASH PRIZE." Tracks include access to justice, AI safety and governance, and legal workflow automation.

If you teach paralegal studies, LexHack is a rare competition where a legal-domain student is the scarce member of the team. Six days is tight, but a paralegal student paired with a CS student could scope an access-to-justice tool in a weekend. Read more →

Corrections to last week's competition picks, and what's still open

Corrections. Last week we led this section with the SPEED October AI Challenge ($1,500 cash, October 14). Its Devpost page now returns "410 Gone" and it no longer appears among open hackathons, so treat it as withdrawn and don't build toward it. The same is true of GatewayHacks 2026. We also reported that UnivaBio (October 6) paid $8,345 in cash, including a $2,000 no-code award. Its prize page now lists every award, the $2,000 included, as Momen or other sponsor credits under the label "THIS IS NOT A CASH PRIZE." Whether the page changed or we misread it, the cash figure is effectively zero.

Still open, deadlines re-confirmed. OpenCV AI Competition 2026, October 26, 11:45 p.m. PDT: $12,000 cash plus $8,250 in AWS compute grants, students and faculty 18+. Amazon "Build, Ship, Shape," October 23, 12:00 p.m. PDT: $138,000 cash, Quebec excluded. Nebius x NVIDIA Global AI Hackathon, October 30, 10:00 a.m. PDT: about $41,000 cash plus Jetson hardware, 18+, with a physical-AI track. Next Art AI 2026, September 30 (Korea time): free, no coding, KRW 21 million. Verify on the organizer's site. Skip: Global Innovation Build Challenge V2 (October 1). Devpost lists $0 in prize value, and every award is a credit or subscription.

For Southern California students: LA Hacks' in-person AI Hackathon at UCLA, October 17–18, is open to college students by application. The application deadline and any cash prizes are not yet posted. Read more →

Embodied AI & Robotics

Agility's Digit 5 is built to work beside people, without a safety cage

On September 15, Agility Robotics unveiled Digit 5, its next warehouse humanoid. Agility says it carries 40% more payload than Digit 4, lifts 50 pounds repeatedly, reaches up to 7.2 feet, and runs 90 minutes on a battery that recharges in about nine. The headline claim is "cooperative safety": the robot detects nearby people, warns them with lights and sound, and runs an independent safety controller. The goal is for it to work on shared floors instead of inside a fenced cell. Agility reports more than $300 million in multi-year orders, 65,000+ cumulative deployment hours, and named customers including GXO, Schaeffler, Amazon and Toyota Motor Manufacturing Canada. Early access is planned for the first half of 2027, with general availability by the end of 2027. IEEE Spectrum's same-day coverage called it possibly the first humanoid worker that is "truly safe." Those safety claims are Agility's own and have not been independently certified.

For logistics, mechatronics and industrial-maintenance programs, the caged robot cell has shaped how robot safety is taught for decades. If humanoids move onto shared floors in 2027, industrial safety curricula will need to cover collaborative-safety standards and risk assessments for people working alongside robots, and technicians will need to troubleshoot a safety controller rather than a light curtain. This is a good item for the next advisory board meeting. Read more →

San Diego asks for local control over robotaxis, while Waymo heads to Tokyo and Singapore

On September 14, the San Diego City Council voted 8-0 for a resolution seeking local say over autonomous-vehicle deployment. Amendments address empty vehicles, safe shutdowns during outages or cyberattacks, keeping pickups out of bike lanes and crosswalks, and disability-access compliance. Waymo launched in San Diego on September 1 under state CPUC authority with no city role. Taxi and rideshare drivers spoke about lost income, especially at the airport. Waymo called the resolution "extraordinarily premature." A resolution by itself doesn't change who regulates robotaxis. In California that is the CPUC and the DMV.

The same week, Waymo announced commercial service in Tokyo in 2027 with Nihon Kotsu and the GO app (September 14) and a Singapore launch planned for 2028 (September 17). By its own count, Waymo now runs 500,000+ trips a week across 15 U.S. cities.

For automotive technology programs in Southern California, this has reached your region. The workforce questions it raises are about drivers losing income on one side and fleet technicians being hired on the other. Those make a better classroom discussion than the vehicles themselves, and civics or transportation-policy courses can use the resolution to ask who gets to decide. Read more →

This roundup was assembled with AI-assisted research by the EDUAIATLAS team. Spot something we missed? Let us know.