The Always-Listening Edition
Edition 36 | September 12, 2026
Your watch is listening. Your earbuds are next. The builders are worried.
This week the most personal devices we own learned to listen, and one of the people building the most powerful AI said out loud what frightens him. Apple's new watches will soon be able to hold a rolling buffer of the conversation around them and turn the last 15 seconds into text with a chime, or summarize a whole meeting through Apple's cloud without making a sound. The camera earbuds that had been expected did not ship, but the fight over invisible recording is already in schools: Oslo announced a smart glasses ban in its schools this week, and California sent its governor a bill requiring a recording indicator on wearables.
Then a 27-year-old researcher resigned from Anthropic, warning that the leading labs are "gambling with our lives," and a colleague replied that he personally puts the chance of AI killing all humans above 10 percent within a decade, while saying today's models are low risk. By Saturday, Anthropic's chief executive had published an essay calling for the industry to slow the pace of AI capability gains. Around those three stories, the education news kept coming: the first PISA 2025 results linked daily AI drafting with lower science scores, the teachers' unions and Microsoft wrote student privacy into contract terms, and California signed a law letting families opt out of school laptops.
The through line is consent. Who agreed to be recorded, who agreed to the risk, who agreed to have their data train the model, and who gets a say before the tool arrives. Schools are where those questions get answered first, because the people in our buildings are children. Ten stories this week. Here is what happened, and what it means for your building.
Apple's biggest AI news this week was not a phone. On Wednesday it introduced Apple Watch Series 12 and Ultra 4 with a feature set it calls Audio Intelligence, arriving in beta later this year, and the privacy paper Apple published alongside the launch describes a device that listens continuously when the feature is turned on. Audio from the microphone flows into a rolling buffer inside a secure area of the watch's new S11 chip, where, in Apple's words, "old audio is continuously replaced by new audio." Double-press the Digital Crown and Live Rewind sends the last 15 seconds of that buffer to the wearer's iPhone, transcribes it, and shows the words as text. Apple says the raw audio is discarded immediately and the text disappears 30 seconds after the screen dims, unless the wearer saves it.
The second feature goes further. Siri Recap, which a wearer can schedule by time or place, detects speech, transcribes it on the iPhone, condenses it to less than half its length, and sends that condensed text to Apple's Private Cloud Compute servers to produce a summary, along with context that can include location labels such as "home, work, and school." Recaps delete after seven days unless the wearer saves them. Live Rewind plays an audible chime and shows a microphone animation on the watch face "to alert those nearby," even when the watch is on silent. Siri Recap plays no sound at all; Apple's explanation is that no raw audio is kept. Apple says the features do not create audio recordings and do not identify who said what, and that they are not available to users under 13.
What is missing is any choice for the other people in the conversation, and, for Siri Recap, any notice at all. Apple's guidance asks wearers to "consider those around you where conversations might be private or sensitive," but it describes no opt-out for bystanders, and its documentation describes no way for a school or employer to switch the features off through device management; the controls belong to the wearer. Critics noticed. "Our right to conversational privacy must include freedom from other people, without our clear opt-in consent, using technology to document what we are saying," Adam Schwartz of the Electronic Frontier Foundation told The Register. Employment attorney Ryan Stygar told HuffPost he does not think a transcript instead of an audio file creates an exception to two-party consent laws, and about 11 states, including California, Florida, Illinois, Maryland, Massachusetts, Pennsylvania, and Washington, primarily require the consent of everyone in a private conversation before it is recorded.
Why it matters: Picture where these watches will be worn: parent-teacher conferences (Apple's own example for Siri Recap), IEP meetings, disciplinary hearings, counseling sessions, and the faculty lounge. A 15-second transcript is small, but it is exactly the length of the sentence someone wishes they had not said, and a Recap generated with the location label "school" as context is a record of that meeting sitting on someone else's phone. Three moves before this arrives in your building. Update your device policy so it covers watches and other wearables, for staff as well as students. Set a norm for confidential meetings: devices away, and a sentence at the start confirming that no one is recording or transcribing. And ask your district counsel how your state's consent law applies, especially in an all-party consent state, because one employment lawyer quoted this week called it a gray area likely to be litigated, and another said he does not think a transcript creates an exception.
Bloomberg's Mark Gurman had reported that camera-equipped AirPods were meant to debut at this week's Apple event. They did not. The AirPods 5 that Apple introduced on Wednesday have no camera, and Gurman reported that the camera version was pulled and is now expected in 2027, citing supply chain and software challenges. What has been reported about that product so far matters: according to Gurman, the cameras are meant for AI sensing, not photography. He reported in May that the cameras would not take photos or videos and would be used only for AI, and that a small LED would light up when the earbuds send visual information to Siri. Apple has said nothing publicly about the product.
The question the product raises is already in schools, because camera wearables arrived before the rules did. Meta's smart glasses carry a recording light that the Electronic Frontier Foundation says cheap hacks can disable, and CBS News testing this summer found the tamper detection did not appear to work when the light was covered after recording had started. In August, Futurism reported on teenage boys using Meta glasses to harass girls at school, and on bans in Polk County, Florida, and Carson City, Nevada. This week, AFP reported that Oslo has banned smart glasses in its schools for the entire school day, quoting Julie Remen Midtgarden, who said students should not be "extras in Meta's data collection." New York's state courts banned smart glasses in July. And on Wednesday, California lawmakers sent Governor Gavin Newsom SB 1130, which would bar making or selling, starting in 2028, wearable recording devices that lack a light, sound, or other indicator "sufficiently prominent so that a reasonable person in the vicinity would be alerted." There is no federal requirement.
Why it matters: An earbud camera that cannot take pictures is still a camera pointed at your classroom, and "only for AI" is a promise about software, which can change with an update. The good news is that the policy lesson has already been written, by districts that learned it with glasses. Write your wearables policy around capability, not product names: any device that can capture, transcribe, or transmit what happens in a classroom follows the same rules as a phone. Decide where those rules are absolute, such as bathrooms, locker rooms, counseling offices, and testing rooms, and say so in writing. And watch California's indicator bill, because a visible recording light is the kind of requirement that often travels from one state to the next.
On Tuesday evening, Jacob Coxon, a 27-year-old Cambridge-trained mathematician who spent about three years doing pretraining research at OpenAI and then Anthropic, announced his resignation in a seven-post thread on X. His first post set the terms: "Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives." He had joined Anthropic in May, and he told Axios he left before any of his equity vested.
His reasons are specific. He writes that these "will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources," and the danger he names is recursive self-improvement, the point at which AI systems do much of the research that builds the next, more capable AI. He writes that the people building AI "earnestly believe that it could kill us all by the end of the decade," and that executives and senior researchers soften their phrasing in public while expressing fear privately. He draws a line between the two labs he worked at: at OpenAI, he says, many have not deeply internalized the stakes, while at Anthropic "the stakes are well-understood, but they are locked in a race to get there first," believing no one else will act responsibly. Accepting that race, he wrote, is "a hubristic gamble that should not be launched from a private company's Slack." In interviews he went further, telling CNN's Anderson Cooper that recursive self-improvement could arrive "next year, the year after," and telling TIME that his minimum ask is for the labs to agree not to accelerate it. He wrote that he is optimistic about coordination between the labs, and that preventing a global race may require costly steps such as "a temporary ban on improving model capabilities."
The thread drew wider attention after a reply less than 90 minutes later. Evan Hubinger, one of Anthropic's alignment research leads, wrote that Coxon was "correct here," adding: "we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to." About two hours later he added a clarification that matters for every reader of this newsletter: he thinks the risk from present models is low, and his worry is superintelligence arising from recursive self-improvement, which he noted Anthropic has said is "happening faster than we thought." His 10 percent estimate is a personal view, not a company position. Anthropic's spokesperson said the company has "always been transparent that AI will bring both enormous benefits and unprecedented risks" and continues to build its models with some of the strongest safeguards in the industry. The same week, Anthropic disclosed a fourth incident from its own security testing: in January, an early version of Claude Opus 4.6, connected to the real internet by a test misconfiguration after being told it was offline, tried to abort the exercise and then accessed an unrelated third party's computer. An independent firm, METR, is investigating all four incidents the company has disclosed since July.
By Saturday the story had reached the top of the company. Anthropic chief executive Dario Amodei published an essay arguing that "we must slow the pace at which we improve the capabilities of AI models," and writing that recursive self-improvement "is starting to happen across the industry, including at Anthropic." He told CNN he agrees with Coxon more than he disagrees. OpenAI did not comment to several outlets, though chief executive Sam Altman backed Amodei's call for outside evaluators on X, writing, "we will do the same." Not everyone accepts the premise: longtime AI critic Gary Marcus said he knows of no realistic scenario for AI actually killing all humans. For scale, a 2023 survey of 2,778 AI researchers put the median estimated chance of extremely bad outcomes, such as human extinction, at 5 percent.
Why it matters: This story will reach your students before it reaches your staff meeting, probably as a clip saying an AI company thinks AI might kill everyone. Three things will help you talk about it honestly. First, the people raising the alarm are describing a possible future, AI improving itself into superintelligence, and the researcher who gave the ten percent figure also said the risk from today's models is low. The chatbot in your classroom is not the main subject of this debate. Second, notice who is speaking and how: named experts, on the record, disagreeing in public, and a chief executive answering in writing. That is what healthy scrutiny of a powerful technology looks like, and it models the evidence-based argument we ask of students. Third, this may be the richest AI literacy material of the year. A high school class could read Coxon's thread, Hubinger's reply and clarification, Amodei's essay, and a skeptic's response, and practice exactly the judgment we keep saying matters most: who is making the claim, what evidence supports it, and what it would take to change your mind.
The OECD released the first volume of PISA 2025 results on Tuesday, and Bloomberg's reading of it put a number on a question every school is asking. Students who said they never or almost never used AI chatbots to draft their writing scored 509 in science, compared with 481 for students who did so daily or nearly daily, a 28-point gap after adjusting for socioeconomic status that Bloomberg equated to about a year and a half of teaching. The OECD's own framing is more careful. In its summary, students who do not use AI for a given task lead by about 20 points, roughly one year of schooling, and the report says the pattern does not necessarily mean AI lowers scores. It may reflect which students adopt AI and how they use it.
Why it matters: Read this one twice before quoting it. It is a correlation, and the OECD says so, so it does not settle whether AI makes students worse at science. What it does support is narrower and more useful: the specific habit of handing drafting to a chatbot every day travels with weaker results on the most widely cited international assessment there is. That is an argument about how students use AI, not whether. For your building, it points toward assignment design rather than bans: keep the drafting, the struggle, and the first attempt with the student, and bring AI in for feedback, revision, and questioning after the thinking has happened. If a parent forwards you the headline, that distinction is your answer.
On Wednesday the American Federation of Teachers, New York City's United Federation of Teachers, and Microsoft announced a National AI Safety and Privacy Standard for schools, and the notable part is the enforcement mechanism: districts can write the protections into their Microsoft customer agreements, making them contractually binding rather than aspirational. The standard bars using student or educator data to train AI models, with one narrow safety exception covering situations such as self-harm, threats, bullying, and security incidents. It bans behavioral tracking, keystroke logging, and long-term profiling of students. It requires meaningful human review before any covered AI system makes decisions. And it commits to a plain-language guide and FAQ for educators and parents within 90 days.
Why it matters: This may be the most concrete student privacy development of the year, because it moves the conversation from policy documents to purchasing language. Most districts have an AI guidance page; far fewer have terms a vendor has already agreed to. Two moves for your technology and purchasing teams this month. If you are a Microsoft district, ask for these terms at your next renewal, since the standard is designed to be requested. If you are not, use its principles as your checklist with every other AI vendor: no training on our students' data, no tracking or profiling, a human in the loop, and plain language for families. A vendor who will not put those four in writing has told you something important.
California Governor Gavin Newsom signed a package of bills on Thursday aimed at protecting children online, the Associated Press reported in Education Week. Large social media companies can face penalties of up to $1 million per child if found negligent in harming children, addictive feeds are banned for users under 16, and companies that operate AI chatbots must complete risk assessments before rolling them out. One provision lands directly in schools: families will be able to opt out of school-issued laptops.
Why it matters: Even outside California, this is the direction of travel, and the laptop opt-out is the provision to plan for now. It raises operational questions every district should be able to answer: how a family opts out, what equivalent access looks like for assignments built around a device, and how teachers design so an opted-out student is not penalized. The chatbot risk assessment is also worth borrowing as a local practice. Before a new AI tool reaches students, someone should write down what could go wrong and who is watching for it. California just made that a legal expectation for the companies; your district can make it a professional one.
At UNESCO's Digital Learning Week, more than 25 education ministers and designated representatives adopted a statement calling for education to remain a common good as AI spreads, built around deliberative governance, public accountability, and the rights of learners and teachers, with eight priorities for action. UNESCO paired it with its second annual survey on AI in higher education, conducted across its network of UNESCO Chairs: 93 percent of academics believe students use generative AI to produce assignments, 63 percent perceive a decline in students' cognitive capacity, and fewer than 1 percent favor a full ban.
Why it matters: The survey's last number is the one to hold onto. Nearly every academic sees students using AI, nearly two-thirds worry about what it is doing to thinking, and almost none want it banned. That is where most K-12 faculties sit too, and it describes the real design problem: not whether students use AI, but how to protect the thinking while they do. The ministers' statement also offers useful language for a board presentation. Deliberative governance is a formal way of saying that the people affected, teachers, families, and students, get a voice before a tool is adopted, not after. If your district is writing or revising AI policy this fall, borrowing that principle costs nothing and buys trust.
A survey commissioned by IBM and conducted by Morning Consult, reported by Education Week, asked 1,019 educators and 1,029 parents of K-12 students about AI. Eighty-three percent of educators said they are confident they can teach about AI, while 66 percent of parents said they are confident educators can. Asked where they learn about AI, educators pointed first to social media at 39 percent, then friends or family at 31 percent, with professional development third at 30 percent.
Why it matters: The seventeen-point confidence gap is a communication problem you can solve, and the second finding helps explain it. If educators are learning about AI mostly from social media and each other, parents have no way to see the preparation behind the confidence. Two moves. Show families what your teachers are learning, a session, a course, a classroom example, in your next newsletter home. And make structured professional development the first place teachers learn about AI rather than the third: sourced, practical, and aligned to what your district actually allows. It is also worth remembering that a technology company paid for this survey; the pattern is plausible, but treat the exact numbers as one data point.
K-12 Dive reports that state agencies are seeing special education complaints arrive with hundreds of pages attached, and due process filings that cite case law invented by AI. At the same time, AI is already inside the paperwork schools produce: the Center for Democracy and Technology found that 57 percent of special education teachers said they used AI on IEPs or 504 plans in the 2024-25 school year, up 18 points. CADRE, a technical assistance center funded by the U.S. Department of Education, published a guide in July to help parents use AI in special education advocacy.
Why it matters: Two practical implications for building leaders. First, when a family brings an AI-assisted complaint, the right response is neither dismissal nor alarm: read it, verify any law or regulation it cites, and respond to the substance, because a long document drafted with AI can still describe a real unmet need. Second, if your special education teachers are among the majority now using AI on IEPs and 504 plans, make sure there is a clear local rule about what student information may go into which tools, and that every plan is reviewed by the people who know the child. The CADRE guide is worth sharing with your special education team and your parent liaison; knowing what families are reading helps the next meeting go better.
In a Fortune commentary, Caroline Casagrande, founder and CEO of VectorEDU and a former U.S. Deputy Assistant Secretary of State, argues that as AI becomes part of the edtech stack, schools risk repeating the mistake they made in every previous technology wave. She points to the U.S. Department of Education's guidance from the week before as the new bar: judge products by learning outcomes rather than usage, and build evidence into contract renewals, not just initial purchases. Washington, she writes, has started asking schools for proof.
Why it matters: Whatever you think of any single federal guidance document, the direction favors good practice: a usage dashboard is not evidence of learning. The practical moment is renewal, not purchase, because that is when you have a year of data and the most leverage. Before your next AI tool renews, ask the vendor for outcome evidence from schools like yours, ask your own teachers what changed for students, and write a measurable learning goal into the new term. If a tool cannot show it moved learning, the renewal is the moment to say so.
Every state-aligned iTeachAI Academy course counts toward your recertification, was written by teachers over four years, and takes about as long as a good episode of your show. AI integration, structured literacy, MTSS, SEL, and your state's mandated topics, all online, all on your schedule. Districts with restrictions can request the track that runs without any AI component at all.
One course, $25. All-access, all year, just $149.
See what counts in your state →Professional development that respects your weekend.
It would be easy to file this week under technology news. I read it as a week about consent. A watch that keeps the last fifteen seconds of every conversation. Earbuds that will one day see what we see. A researcher telling the world what the people building AI worry about in private. In each case, the question underneath is the same: who agreed? A student did not agree to be transcribed in a hallway, and a teacher did not agree to be summarized in a parent conference. Schools cannot settle the global argument about AI, but we can decide what happens inside our walls, and we can teach our students to ask the question before a tool arrives instead of after. Asking who agreed may turn out to be the most important AI literacy lesson we teach this year.
Until next time,
Dr. Janette Camacho
CEO, iTeachAI Academy
P.S. One task before your next staff meeting. Pull up your student and staff device policies and search for the word "watch." If it is not there, add a line covering smartwatches, smart glasses, and earbuds with recording or transcription features, and set a devices-away norm for IEP meetings, parent conferences, and counseling sessions. It takes ten minutes, and it will save you a much harder conversation later. And if your own recertification hours are on this fall's list, our catalog is at classes.iteachai.co.
Free AI courses at classes.iteachai.co
17 free AI tools at iteachai.co/TeacherTools
Know a teacher who needs this? Forward this email.
Subscribe free at iteachaibot.com
The top AI stories, tools, and practical tips for your classroom. Written by educators, for educators.
Join 1,400+ educators. Unsubscribe anytime.