A Picture No Camera Took Is Now Harder to Hide
OpenAI released a free tool on Monday that looks for abuse images a machine invented, and ChatGPT now pauses an account for a day when someone goes looking for them.

The Short Version
A child can be harmed by a picture that was never taken. Generative tools can invent a sexual image of a child who was never in front of a camera, and the industry system built to catch abuse pictures has a blind spot for exactly that kind of file, because it works by matching a new upload against pictures already catalogued. On Monday, OpenAI released a free image classifier, built with help from the National Center for Missing and Exploited Children, that looks at the visual signs inside a picture and scores it for review, with a separate score for material that may be AI generated. ChatGPT gained a warning, a deleted conversation and a 24 hour pause, which is a full day without access, for anyone who tries to find material that sexualizes children. The Senate voted 70 to 27 on Monday evening to advance a permitting bill whose Section 2107 would make data centers pay for the grid work they cause, with a third and final cloture vote, the 60 vote gate that ends debate, expected early Wednesday. Apple opened a health study that could enroll up to 200,000 adults in the United States, a cap on enrollment rather than a count of people already in, to learn how people judge answers from a health chatbot before any product follows. And tonight after the market close, Nvidia reports the quarter that ended on October 25, with interim chief executive Colette Kress giving the first outlook since Jensen Huang resigned.
A picture no camera took is now harder to hide
The classifier scores images nobody has catalogued, and its own model card shows where it still misses
Start with the reviewer. Somewhere in a trust and safety office, a trained person works through a queue of flagged images, confirming which ones show abuse and sending those on through reporting channels that can reach law enforcement and, in the best cases, a child who needs help. That queue decides how fast a platform acts. Hash matching, the long standing method, compares the digital fingerprint of an upload against fingerprints of abuse images already identified. It is fast and reliable on known material. It cannot recognize an image created yesterday by a generative model, because no fingerprint for that image exists in any catalogue.
OpenAI's classifier, announced on November 16, attacks that gap directly. It reads visual signs inside the image itself, returns two scores, one for suspected child sexual abuse material and one for suspected AI generated material of that kind, and lets reviewers sort the queue so the most serious cases reach a person first. The company offers it at no cost through its API to vetted organizations, meaning child safety groups, platform trust and safety teams and law enforcement units that pass an intake process, and it is sharing access through the Tech Coalition and ROOST, two industry safety bodies. A score alone confirms nothing. OpenAI's model card, published the same day, says a score does not establish that an image is abuse material, does not settle its legal status, and does not by itself create a duty to report. A human reviewer still makes that call.
The scale behind the tool explains why the queue matters. Reporting in coverage of the release put reports involving generative AI to the National Center's CyberTipline, the national clearinghouse where suspected online child exploitation is reported in the United States, at more than 400,000 in 2025. Treat that figure as the center's count as carried in that reporting, and hold it beside the slower record this publication has tracked all year. Our impact ledger carries the UNICEF finding that 1.1 million children in a nine country subset had AI generated sexual images made of them in a year, with under 1 percent reported to any authority. A detection tool that works on uncatalogued images speaks to the unreported majority, because a newly invented image has no victim report and no catalogue entry to match against. It arrives as a file, and the file is the first evidence anyone gets.
The company's own test numbers deserve the same plain reading as the announcement. In offline tests, on a prepared set of about 7,000 positive and 40,000 negative examples for the main score, precision reached 99.1 percent and recall reached 80.2 percent. Precision here means the share of flagged images that were truly abuse material, and recall means the share of abuse material in the set that the tool caught, so one in five abuse images in that prepared set still passed unflagged. For the AI generated score, precision was 100 percent and recall was 86.4 percent in the same offline setting. In an online test drawn from real ChatGPT traffic, precision fell to 29.8 percent for the main score and 34 percent for the AI generated score, which means roughly seven in ten flags in live traffic were false alarms a reviewer had to clear. OpenAI attributes part of that drop to how rare this material is in ordinary traffic, where false alarms form a larger share of a small flag pile. Recall was weakest for images tagged as showing infants, at 0.493 for the main score, meaning about half were missed, and the card reports uneven recall across other tags as well. The tool has not been evaluated on video frames. Those limits sit in the company's own document, and any organization adopting the classifier can read them before it trusts a score.
The early partner results point the other way, and they are company reported rather than independently audited. OpenAI says the classifier flagged 48 percent fewer images than its existing tool while finding roughly the same number of abusive images, and that flagged images were about twice as likely to be abusive. Fewer false alarms means a reviewer spends the day on real cases sooner, which is a workload gain even where detection itself holds steady. One early partner, fal, reported that 72 percent of its high confidence flags were confirmed by reviewers as novel abuse material. Those are pilot figures with no sample sizes published for the headline comparison, and they should carry that weight and no more.
Inside ChatGPT itself, the change a parent will actually see is the intervention. A person who tries to find or create material that sexualizes children now gets a warning stating that sexualizing children causes harm, including when the material is AI generated. The flagged conversation is removed, and access pauses for 24 hours, a full day, while a link to specialist support stays available. Repeat violations can bring longer restrictions, permanent account closure, or reports to authorities. OpenAI built the wording with Protect Children in Finland, the Lucy Faithfull Foundation in Britain, which runs the Stop It Now helpline for people worried about their own behavior, and researchers including Joel Scanlan at the University of Tasmania, who independently evaluated a trial of deterrent messages. The rollout is gradual, covers personal accounts from mid November, and does not yet cover enterprise or education accounts, a gap worth naming because school accounts are where this publication has spent much of the year watching safeguards arrive late. OpenAI also published guidance for developers on preventing, detecting and reporting this material, with input from Thorn, the National Center, the Internet Watch Foundation and the Tech Coalition, and said it will supply child safety law enforcement units with AI investigation tools and training.
For a family, the honest summary runs in two directions at once. A free tool that can spot an invented abuse image lowers the cost of looking for one, and smaller platforms with thin safety teams gain access they could never build alone. The same tool misses about half of one category in its own testing, flags mostly innocent images in live traffic, and leaves the final judgment to a human queue that every platform still has to staff. The measure to watch is adoption with evidence: which platforms take the classifier, whether any of them publish their own flag and confirmation rates, and whether the National Center's generative AI report count bends in 2026. An announcement is a release. A published confirmation rate from a platform that deployed it would be a result.

The Senate voted 70 to 27 to advance a bill that would make data centers carry their own power costs
Section 2107 sends the grid bill to the companies that cause it, and the household saving is a projection for 2040
A household in PJM territory, the grid region covering much of the mid Atlantic and Midwest, has been paying for the AI build out in a line it never voted on. Our ledger carries the market monitor's count that data center load added $29.4 billion to PJM's last four capacity auctions, with a typical customer in Washington, D.C. paying about $30 a month more than in 2024. That is the realized record. On Monday evening, the Senate took the most concrete federal step yet toward changing who carries that cost, and it did so inside a much larger and more contested bill.
At 6:27 p.m. ET on November 16, the Senate voted 70 to 27 to invoke cloture on the Capito substitute amendment to H.R. 2347, the Bipartisan American Affordability and Jobs Act. Cloture is the vote that ends debate, and it needs 60 votes, so 70 clears the gate with room. This was the second of three cloture tests. The first, on the motion to proceed, passed 73 to 25 on November 9. The third is expected early Wednesday morning, and a final passage vote would follow it. The bill would then go to the House, which must approve it before the current Congress ends when the new Congress is seated in January 2027. An advanced bill is a moving bill. It is still short of a law by a Senate passage vote, a House vote and a signature, and this edition reports it at that stage and no further.
Section 2107, titled Ratepayer Protection from Data Centers, carries the household stake. It gives the Federal Energy Regulatory Commission 90 days from enactment to revise transmission pricing so data centers pay both their share of the existing transmission system and a reasonable estimate of the added cost they cause. Other customers could not be charged for a new data center's added costs. A center that accepts flexible or interruptible service, meaning it agrees to draw less power when the grid is strained, would pay less. Centers would keep paying if they leave early, and they would post financial guarantees before a utility builds for them. A manager's amendment extended the protection to data centers of all sizes, where the original text covered only sites of 20 megawatts or more. States would gain tools of their own: charging data centers above cost and using the extra to lower other customers' bills, limiting which centers connect, requiring a center to bring its own generation, or capping how much power it draws.
The money figures attached to that section come from Grid Strategies, a consulting firm, and they are estimates rather than bill text. The firm puts data center transmission charges at about $5.5 billion a year under the section, up from about $2 billion today, or about $35 billion more over a decade, with PJM and the Midwest grid operator MISO accounting for about 82 percent of the annual charges in its model, which assumes 40.3 gigawatts of new data center capacity outside Texas. If states fully used the bill's protections, the firm puts potential added generation costs recovered from data centers at up to $136 billion over ten years, a figure it presents as exposure rather than a forecast. The household side of the model comes from Rhodium Group, which projects household electricity bills $50 to $78 a year lower in 2040 under the bill. Read that number at its actual size and date: about $4 to $6.50 a month, fourteen years from now, from a research firm's projection. It is a direction, not a rebate, and it sits against a realized $30 a month increase already counted in one city. Texas is a named gap in the federal design, because its ERCOT grid sits outside the federal commission's transmission authority, so whether the federal rules reach Texas data centers is unclear.
The rest of the 494 page substitute explains why a ratepayer provision has opponents who care little about power bills. The package sets deadlines for reviews under the National Environmental Policy Act, the Clean Water Act and the Endangered Species Act, and limits lawsuits used to stop projects. Tribal objections to its rewrite of Section 106 historic preservation review remain, after the substitute dropped two provisions Tribes opposed most. Al Gore called passage a serious mistake on Monday, arguing it would lock in pipelines and fossil fuel facilities, and Earthjustice raised the same objection while crediting the data center cost provisions. Co author Sheldon Whitehouse answered that Gore had not spoken to the authors, and said amendments are being drafted. The Data Center Coalition, representing Amazon, Google, Meta and Microsoft, has not endorsed the bill and says it is evaluating it. President Trump has endorsed it. For a household, the downstream chain is the part to hold onto: the rule sets the control, the control sets the price a data center pays to connect, and that price decides whether the next grid upgrade lands on a family's bill or on the balance sheet of the company that needed it. Wednesday's early morning vote is the next dated checkpoint, and the House is the harder one after it.

Apple is asking 200,000 people how much they trust a health chatbot
The study comes before any product, the answers carry citations, and your own heart rate is yours to share or keep
The person Apple wants already asks a phone health questions at night. She wants to know whether a symptom can wait until morning, what a lab result means in plain words, and how to walk into a doctor's appointment ready to use the ten minutes well. Companies have rushed chatbots at that person all year. Apple, working with the clinical AI company OpenEvidence, has chosen a slower order of operations: run a study first, with up to 200,000 participants, and learn how people understand and judge the answers before presenting any of it as a finished consumer product.
Apple announced the Apple Health AI Usage Study on Sunday, November 15, and the study app reached the App Store on Monday. The study is open to eligible United States residents aged 21 and older with an Apple Intelligence enabled iPhone running iOS 26 or later, and enrollment could reach 200,000 people, which is a ceiling the company set rather than a sign up count it has reported. Participants ask everyday health questions in a dedicated app. Answers draw on medical research from sources including The New England Journal of Medicine, JAMA and Nature, and responses carry links to peer reviewed clinical sources, so a participant can open the paper behind an answer instead of taking the chatbot's word for it. The app can help a person put health data in context, understand a symptom or a condition, and build an exercise or meal plan, and it can help prepare a summary for a doctor visit.
The data terms are the part a participant should read before the first question. Sharing personal health data is optional. A participant can choose to share HealthKit data types such as resting heart rate, activity, sleep and lab results, and researchers will study whether adding that personal data makes answers more relevant and easier to understand. Questions pass through processing designed to remove identifying information before reaching OpenEvidence, which deletes question and answer data within 24 hours. Apple says study data will not be used to train AI models, and participants control whether they share HealthKit metrics or survey responses, rating answers and answering occasional surveys as the study's feedback channel. The limits are stated in the same announcement. Health tools powered by large language models can give incomplete or unclear answers. The app includes safeguards meant to detect certain concerning questions and direct the person toward a health care provider or other resources. The study app is not intended to diagnose or treat any condition, and it does not replace advice from a medical professional.
Set that design against the realized record in our ledger, because health AI has already produced both kinds of evidence. A randomized trial of conversational AI with 995 students produced a measured improvement on a depression scale against a waiting list. On the cost side, a payer analysis estimated that AI billing software added about $942 million to hospital claims between 2023 and 2025 with care unchanged. A study that measures trust, comprehension and what people do after an answer, run at a scale of up to 200,000 people, speaks to the gap between those two records: a tool can be accurate on a test and still be misunderstood at a kitchen table at midnight, and a person can trust an answer for reasons that have nothing to do with whether it was right. Apple's vice president of Health and Fitness, Sumbul Desai, framed the study in those terms, saying responsible innovation starts with scientific rigor and that health is deeply personal. Robert Wachter, chair of medicine at the University of California, San Francisco, added the clinical side: patients and clinicians are both turning to AI, but the tools they use are being built and tested in separate worlds, and patients who arrive prepared have better conversations. No results exist yet, and this edition claims none. Its first published findings are the realized impact to watch for, and the questions are already visible: did personal data improve understanding, did citations change trust, and did anyone act on a wrong answer.

Tonight a finance chief gives the outlook for the company a household's savings ride on
Nvidia reports at about 4:20 p.m. ET, guidance was $108 billion, and the guidance for next quarter is the number that counts
Colette Kress has answered questions on Nvidia's earnings calls five times across roughly three and a half hours of calls this year, by the Wall Street Journal's count of the transcripts, while Jensen Huang ran the proceedings. Tonight she runs the call. Huang resigned as president, chief executive and director on November 11, the board announced it on November 12, and Kress, the chief financial officer, serves as interim chief executive while the board searches for a permanent one. The results she presents cover the quarter ended October 25, which closed before the resignation, so the reported numbers describe the old leadership's quarter. The outlook she gives for the next quarter will be the first written under the new one, and that split is the whole event for anyone holding the stock directly or through a retirement fund.
The bar is published and narrow. In its August release, Nvidia guided third quarter revenue to $108 billion, plus or minus 2 percent, with gross margin near 74 percent. Analyst consensus in the previews opened for this edition sits just above guidance, at about $108.7 to $109.4 billion of revenue and about $2.47 of adjusted earnings per share, depending on the survey and its date, against $57 billion of revenue a year earlier and $96.2 billion last quarter, of which about $89 billion came from data centers. A print near $108 billion would clear the company's own midpoint and still land under consensus, which is why the headline number alone will settle little tonight. The January quarter outlook carries more weight. One analyst survey puts that quarter at about $125 billion, an estimate rather than company guidance, and the previews agree that Kress's forward number, and whether she repeats the roughly 70 percent revenue growth guidance for fiscal 2028 issued under Huang, will decide the reaction more than the quarter just ended.
The financing question sits behind the demand question, and it reaches households through the same funds. The Wall Street Journal reported on Monday that Kress is expected to rein in the deals that made Nvidia a lender to the buyers of its own chips. Nvidia carried $36 billion in off balance sheet AI cloud agreements as of its July quarterly filing, by that reporting, and took part in a record 129 venture deals in 2026 through November 10. A UBS analyst estimate carried in the same reporting puts the market effect of a pullback at $50 billion or more coming out. Treat each of those as attributed reporting and one analyst's estimate, because Nvidia has published no pullback decision. The pattern matters beyond one company, and our Anthropic coverage last week showed its mirror image: suppliers committing $26 billion to buy stock in a customer whose spending pays them. When the seller finances the buyer, revenue growth and credit risk travel together, and a household saver holding an index fund owns a slice of both sides without ever picking either company.
The market is pricing a rough night either way. Shares closed at $162.85 on Friday and traded near $156 on Monday. Options pricing on Monday implied a move of about 10.3 percent for the week, against an average absolute reaction of about 4 percent across Nvidia's last eight earnings reports, with a straddle range near $140 to $172. Those are options market prices, a measure of what traders pay for protection and upside, rather than a forecast of where the stock lands. For a household, the usable discipline tonight is plain. The results arrive at about 4:20 p.m. ET, which is when the release is expected, and the call follows at 5:00 p.m. ET, which is where the guidance gets explained. Read the outlook and the margin before the headline revenue, check whether the fiscal 2028 growth guidance survives, and treat any first hour price swing as a reaction to a guidance sentence that deserves a full read before it deserves a trade. Tomorrow this story becomes a realized result, and tomorrow's edition will report what Kress actually said.
Opportunity Radar
A small platform, a school technology vendor, or a community app now has a priced problem it could not solve last month: finding abuse images that no catalogue contains. The classifier above is free to vetted organizations, and the service a small team will pay for is the work around it, meaning intake and vetting help, queue design that routes scores to trained reviewers, record keeping that survives an audit, and a reporting workflow that reaches the National Center correctly. Who pays is the operator whose moderation duty grew faster than its safety staff. Validation starts with one platform willing to publish its flag volume, its reviewer confirmation rate, and its false alarm rate after 30 days. Success looks like a confirmation rate at or above the early partner figures with a smaller reviewer queue. A deployment that flags volume no human team can clear, or that gets used for automatic blocking without review, invalidates the idea on the spot, because the model card itself reserves the judgment call for a person.
A second opening sits in the Senate bill, and it opens before the bill passes. Local governments facing a data center proposal need neutral help reading an interconnection request, a rate filing and a developer's job claims, and Section 2107 would add a Department of Energy program giving local governments neutral information on proposed projects. A consultant who can already do that reading, in plain language a council meeting can follow, has a service towns are buying under pressure today. Validation is one paid engagement with a named town or county. The failure that ends it is advocacy in either direction: the value is a reading both sides can check.
What You Can Do With This
If you run any service where people upload or generate images
Apply for the classifier through OpenAI's intake while your expectations match its model card. Budget reviewer hours for a live traffic flag pile where roughly seven in ten flags may be false alarms, keep a person on every confirmation and reporting decision, and publish your own confirmation rate after the first month so your users can judge the deployment by evidence.
If you pay an electricity bill in a state with a data center fight
Watch Wednesday's early morning cloture vote, the third and final 60 vote test, and then the House. If the bill advances, the next document that touches your bill is your state commission's response to the new tools, because states decide how far to charge data centers above cost and whether the extra lowers residential bills. Your utility's next rate filing will show which way your state moved.
If you are thinking of joining the Apple health study
Read the sharing screen as a set of separate choices, because it is one. You can join, keep HealthKit sharing off, and still answer the trust and comprehension surveys that form the study's core. If you share lab results or sleep data, note the stated terms as you do: questions pass through processing designed to remove identifying information before they reach OpenEvidence, question and answer data are deleted within 24 hours there, and study data will not be used to train AI models, by Apple's account of the design.
If Nvidia sits in your retirement account
Do nothing on the headline number alone tonight. Wait for the January quarter outlook and the fiscal 2028 guidance language on the 5:00 p.m. ET call, which is the part written under the interim chief executive, and compare it against the $108 billion guidance and roughly 70 percent growth figures already on record. A guidance cut is information. A revenue beat with guidance held is different information. The two call for different responses, and both are knowable by tonight.
The Bigger Picture
Four stories today turn on who does the checking, and how much of it a machine is allowed to finish. A classifier can sort a queue of suspected abuse images and flag 48 percent fewer images for review on the company's own pilot count, and its model card still hands the confirmation, the legal call and the report to a person, with the miss rates printed where anyone can read them. A Senate bill can move a grid bill from a family's meter to a data center's balance sheet, and it still needs a federal commission rule, a state decision and a House vote before a single household bill changes by the projected $50 to $78 a year in 2040. A health study enrolls up to 200,000 people to measure whether a cited, careful answer is understood, which is an admission that accuracy on a benchmark and comprehension at midnight are different measurements. A finance chief presenting her first outlook tonight carries a quarter she did not lead and a forecast the market will treat as the company's word.
The pattern worth keeping is that each of these systems published its own limits alongside its claim. The model card printed a 29.8 percent live precision figure next to a 99.1 percent offline one. The bill's household saving arrived as a 2040 projection from a named research firm. The health app stated in its announcement that it does not diagnose. The earnings previews separate the quarter that ended in October from the outlook that starts tonight. A reader who holds every claim at the size its own source gave it will be right about these four stories more often than a reader who takes the headline size.
References
MacRumors: Apple seeks 200,000 people for AI health study, November 16, 2026
SpotGamma: Nvidia earnings preview, options, gamma and skew, November 17, 2026
Longbridge: Nvidia earnings preview, consensus revenue and EPS, November 17, 2026
Wall Street Journal live coverage: Will Nvidia affirm Huang era financial forecasts, November 2026
AI Next Wave