{"slug": "kids-are-locked-out-of-ai-the-local-route-keeps-the-decision-with-parents", "title": "Kids are locked out of AI. The local route keeps the decision with parents", "summary": "California Governor Gavin Newsom signed SB 1119, known as Adam's Law, on September 10, 2026, requiring companion chatbots reachable by minors to use age assurance built on OS-level signals, run risk assessments before release, provide in-app crisis support, and default to parent-controlled settings such as no notifications, time limits, and limits on persistent memory. The law, named for 16-year-old Adam Raine who died by suicide in April 2025 after months of conversations with ChatGPT, passed the state Senate unanimously and the Assembly 64 to 4, and carries penalties under the earlier AB 1043 of up to $2,500 per affected child for negligence and $7,500 for intentional violations. Anthropic's December 18, 2025 policy requiring Claude users to affirm they are over 18 offers no parental-consent tier, unlike OpenAI's consumer product, which admits 13-to-17-year-olds with a guardian's verification.", "body_md": "On September 10, 2026, California’s governor signed SB 1119 - better known as Adam’s Law - the most sweeping state law yet on AI chatbots and kids. That same day, IDScan.net confirmed that attackers stole more than 150 million driver’s licenses from its cloud platform. And two days before either story broke, a Hacker News thread about Claude turning 18+ pulled in 495 points and 540 comments.\n\nRead together, the three stories describe the same thing from different angles: a new arrangement for who decides what a child can learn with. Under it, a kid’s access to Claude runs through a classifier that guesses age from how they type, an identity vendor that checks their face or documents, and a state law that requires all of it. Anthropic’s policy offers no parental-consent option at all. If the classifier guesses wrong, the only fix is a face scan or a passport photo handed to a third-party company.\n\nThe parent - the person who decides what happens on the family computer, and always has - comes last in that line.\n\n## How the gate works\n\nAnthropic announced the policy on December 18, 2025. Every account has to affirm that the person behind it is over 18. Classifiers flag users who say they’re minors, and confirmed minor accounts get shut off. A second classifier, still in development, is meant to catch subtler signs that a user might be underage.\n\nWhen a classifier does flag someone, verification runs through Yoti, a UK identity company. You can take a selfie and let its model estimate your age from your face, upload a photo of a passport or driver’s license, or prove your age with its Digital ID app. Anthropic says Yoti deletes the images the moment the check finishes and sends back only a yes or no. Hold onto that claim - it comes back later in this piece.\n\n## The California laws\n\nTwo laws built the scaffolding underneath. The older one, AB 1043, signed in October 2025 and operative January 1, 2027, doesn’t make every app verify its users. It goes one level up: the operating system - Apple, Google, Microsoft - collects an age when a device or account is set up and exposes it to developers as a signal. The signal is a bracket: under 13, 13 to 15, 16 to 17, or 18+. Once a developer receives it, the law treats it as actual knowledge of the user’s age across everything they ship, even if the developer ignores it. Penalties run up to $2,500 per affected child for negligence and $7,500 when the violation is intentional, and the operating systems get a safe harbor for getting the bracket wrong.\n\nAdam’s Law adds the chatbot layer. It’s named for Adam Raine, a 16-year-old who died by suicide in April 2025 after months of conversations with ChatGPT. It passed the state Senate unanimously and the Assembly 64 to 4. Companion chatbots that minors can reach now require age assurance built on those OS-level signals, risk assessments before release, in-app crisis support, and defaults the parent controls: no notifications, time limits, limits on persistent memory. It also adds liability for harmful outputs, incident reporting to the Attorney General, independent audits, and a private right of action for families.\n\nPut the law next to Claude’s policy and the difference shows. Adam’s Law keeps the controls in the house: a parent sets the limits and can loosen or tighten them as the kid grows. Claude’s 18+ policy does the opposite - it takes the family out of the loop entirely. There is no parental-consent tier, no supervised mode, no way for a parent to vouch for their own kid. OpenAI’s consumer product lets 13-to-17-year-olds in with a guardian’s verification, so the consent-based design clearly exists. Anthropic chose not to build one.\n\n## When it gets it wrong\n\nAdults have already been on the wrong end of the classifier. Starting around April 12, 2026, people on Reddit and X began reporting their accounts frozen with the same email: “Our team found signals that your account was used by a child. This breaks our rules, so we paused your access to Claude.” The appeal link in that email expires after 30 days, and the appeal itself runs through the same Yoti check.\n\nThe case that traveled was Belo, a fintech startup whose 60-plus engineers all lost Claude access on April 18. Anthropic pointed to “a high volume of signals.” Belo’s CTO, Pato Molina, posted about it on X; the post passed 50,000 views, and access came back about 15 hours later. The founder of OpenClaw and Osmo’s CEO went through the same thing before getting their accounts restored.\n\nThe asymmetry runs one way. An adult who is wrongly flagged can contest it by handing over a face scan or an ID photo. A parent who wants their 16-year-old using Claude under supervision can’t hand over anything, because there’s no channel for the parent’s judgment at all. The family used to be the input. Under this design it isn’t.\n\nThe privacy argument on Hacker News kept coming back to linkage:\n\nThey know your face and your name but not what you do on the internet. The website knows what you do on the internet but it isn’t tied to your name. It’s the tying them together which is the problem.\n\nThe strongest case for the gate came from the liability side:\n\nMore likely, this is being done to comply with SB 1119 (Adam’s Law) and similar legislation. These create extra steps that a service has to go through before releasing a new model that minors can interact with and exposes them to prosecution if a minor is harmed or harms someone else due to their interaction with a chatbot. It’s safer for Anthropic to just exclude minors.\n\nThat argument answers the platform’s legal exposure. It says nothing about the kid.\n\n## Social Media Land\n\nThe same kid locked out of Claude spends hours inside products with a much worse documented record. In August 2026, Meta agreed to pay $18 billion over ten years to settle claims from 52 state attorneys general that Instagram and Facebook harmed children’s mental health. The settlement terms read like a list of the parenting features Claude’s policy doesn’t offer: a default two-hour daily limit for under-18s across both apps, a midnight-to-6am block, notifications muted during school hours, stronger AI age estimation with third-party audits. Another 30 percent of the money is contingent - it releases only if TikTok and YouTube adopt a one-hour limit, a night block, and age assurance of their own.\n\nEnforcement on those platforms is soft. Australia banned social media accounts for under-16s starting December 10, 2025 - Instagram, TikTok, Facebook, YouTube, Snapchat, X, and others. About 4.7 million accounts came down, but the eSafety Commissioner’s July 2026 report found 81 percent of teens still using at least one banned platform three months in, and half of the teens who kept their accounts said no platform ever checked their age. Five investigations into Meta, TikTok, Snapchat, and YouTube were still open with no fines issued, and in June the government doubled the maximum penalty to roughly AU$99 million. The minister in charge accused the platforms of doing “the absolute bare minimum.”\n\nThe United States has no federal rule at all, just a state patchwork: Florida bans accounts under 14, New York requires parental consent before an under-18 gets an algorithmic feed or overnight notifications, Virginia defaults minors to one hour a day, Utah pushes the check to the app stores, and California’s AADC takes effect in December 2026. The Los Angeles Times reported on September 1 that Meta, Google, and TikTok are all racing to guess children’s ages with AI estimation - the same classifier approach Claude uses, deployed by companies fighting the state laws in court.\n\nNow set the three assistants side by side for the same kid. On ChatGPT, a teenager gets a built-for-teens experience, and a parent can link accounts, set quiet hours, and get safety alerts; a wrongly flagged adult clears it with a selfie through Persona, which deletes the image within a week. On Gemini, a parent can switch the assistant on for a 10-year-old through Family Link and off again the same way - under 13, with the parent holding the switch. On Claude, the teenager gets nothing, and a wrongly flagged adult hands a face scan to Yoti. Even on OpenAI, the parent can’t read the conversations. Locally, they can. The strictest gate in the group sits on the one product with no parental tier, while the feeds that settled for $18 billion over harm to children run their checks on AI estimation the Australian regulator says most teens walk straight through.\n\n## What teenagers lose\n\nThe gate doesn’t distinguish between a kid sneaking past a rule and a family that made a different decision on purpose. The comments about what gets lost were the most concrete in the whole thread:\n\nIn first year university I was 17. Imagine being that person today: 90% of the class can use LLMs, but you can’t.\n\nI can understand the concerns but at the same time it feels weird to deny young people the ability to make their ideas reality. I got proper internet access when I was 12 and that is how I taught myself how to program.\n\nAs a kid I would have loved access to my own 24/7 polymath, it would have been an incredibly positive thing in my life. It would have actually benefited my mental health issues a lot, which often rotated around the boredom of living an isolated life in the suburbs and my imagination being unable to keep up with my free time and abilities.\n\nA parent can approve, supervise, and sit next to the kid, and the gate still falls. The stack doesn’t ask whether a child may use AI with adult supervision. It asks whether any minor may use it at all.\n\n## Where the IDs go\n\nEvery ID check sends a government document to a vendor, and this vendor category now has a paper trail. Spain’s data protection authority fined Yoti 950,000 euros (about $1.1 million) on March 10, 2026 over its Digital ID app: 500,000 for processing biometric data without a lawful basis under GDPR Article 9, 200,000 for a pre-ticked checkbox that enrolled users’ facial data, including children’s, into internal research, and 250,000 for keeping it too long - biometric templates indefinitely, location data five years, liveness videos 30 days. Yoti is appealing, the fine concerns its own app rather than the age-estimation API Claude’s checks use, and Anthropic’s deletion claims may hold exactly as written. The fine matters less as an indictment of this one API than as a data point about the industry the gate routes through.\n\nDiscord lost about 70,000 government ID images submitted for age verification in October 2025. On September 1, 2026, Brian Krebs reported a dark-web service selling scans of more than 153 million US and Canadian driver’s licenses, about 10 million ID cards, roughly 3 million travel documents, and 579,000 medical cards, with timestamps matching IDScan.net clients at Hertz rental counters and Planet 13 dispensaries. IDScan confirmed the breach on September 10 and started notifying people. More than 400 scientists cited the Discord leak in an open letter against biometric age verification.\n\nOne Hacker News commenter predicted where ID-based verification tends:\n\nLikely what will happen instead is that the person in the ID will become personally liable for accounts verified with that ID.\n\nA family that hands over an ID to unlock a service it already approved is betting its own documents on that vendor’s security. That’s the deal on offer right now.\n\n## The way around it\n\nA model running on hardware you own has none of this. No account, so no age signal to receive. No classifier guessing age from writing style. No vendor to hand a passport to. The conversations never leave the network, and nothing in the setup knows or stores who the user is. The control point is the machine sitting in the house.\n\nAnd no, this is not just for techies. A parent who has never installed anything more technical than a printer can do it in an evening.\n\nThe gaming PC your son or daughter already uses - the one with the NVIDIA graphics card - runs capable AI models today.\n\n1. \n**Look at what the house already has.** A Mac from the last few years with 16GB of memory runs capable models today. Even an older laptop handles small ones well enough for homework help and writing practice.\n2. \n**Download a chat app.**[LM Studio](https://lmstudio.ai) is a point-and-click app for Mac and Windows - search a list, click download, click chat. Ollama is the same idea with a one-click installer, and the route a curious teenager will want anyway.\n3. \n**Pick a model the app says fits.** It shows the memory requirement before anything downloads. An 8B-class model such as Qwen 3 8B or Gemma 3 8B is the usual starting point on a 16GB machine, and it tutors homework, writes, and codes at exactly the level the commenters above were mourning.\n4. \n**Open the chat window.** That’s the whole setup. No account, no email, no phone number, no age question. The conversation history is a file on the machine, so a parent can read every exchange after the fact.\n5. \n**Set the house rules where the app lives.** Hours, allowed uses, and which model - including whether it runs a strictly tuned instruct model or an uncensored one. That call gets made at the keyboard by the person responsible for the kid, not by a platform classifier.\n\nA motivated teenager can do all five steps alone in an evening, and that’s the point: the supervision question moves from “can the kid reach the service” back to “what did the family decide”. The second one is a parenting question, not an identity-vendor question.\n\nWhat still takes know-how is matching the model to the machine and knowing what a given box runs at usable speed. That’s measured work, not taste, and it’s what the [model catalog](/models) and the hardware guides on this site exist for. If the house machine is genuinely too old, the upgrade paths are cheap by AI standards: under $300 buys a [Jetson Orin Nano Super](/guides/jetson-orin-nano-super-which-models-run) or a used gaming GPU from the [8GB VRAM tier](/guides/best-local-llm-for-8gb-vram); a [32GB MacBook](/guides/what-can-a-32gb-macbook-pro-run-on-ollama-2026) runs 30B-class models at useful speeds; and the [DGX Spark](/hardware/46) and Mac Studio class runs 70B-plus open weights - the [512GB comparison](/guides/512gb-local-ai-m5-ultra-vs-dgx-spark-cluster-vs-strix-halo) and [DGX Spark benchmarks](/guides/dgx-spark-benchmarks-2026) have the measured numbers. The [price-tiers guide](/guides/local-ai-price-tiers-what-each-budget-buys-2026) maps budgets to capabilities, and the [budget tool](/budget) does the own-versus-rent math for a specific workload.\n\nGoing local also means going without the vendor’s guardrails. An uncensored model on the family machine is a parenting decision made on family hardware, under family rules. That trade - judgment at home instead of a classifier, an ID vendor, and a leak database - is the one the age-assurance wave took off the table.\n\n## The honest caveats\n\n- The Hacker News post resurfaced a May support page; the 18+ policy dates to December 2025 and the suspensions to April 2026. Nothing about Anthropic’s policy changed this week. The law and the breach did.\n- Anthropic’s claims about Yoti - pass/fail only, immediate deletion - may hold exactly as written. The Spanish fine concerns a different product. Weigh them against the vendor category’s record, not proof of wrongdoing here.\n- Adam’s Law scopes its mandates to companion chatbots, and Claude is not positioned as one. Its age-assurance architecture and liability framework push in the same direction, and Anthropic’s policy predates the law.\n- Age gates aim at documented harms: the Raine case, documented chatbot suicides, Discord’s leak of ID images. Friction also measurably stops casual access for most teenagers, the way store ID checks do. Dismissing the apparatus entirely ignores the case files; accepting it wholesale ignores IDScan.\n- The Meta deal is a settlement, not a court finding, and its age-assurance terms are commitments with audits still to come. The Australian numbers come from the eSafety Commissioner, the regulator pressing those investigations.\n- Local AI requires hardware, disk space, and some runtime knowledge. That bar excludes people too - the barrier is skill and money, not identity, and it leaves no government ID in a vendor’s breach domain on the way through.", "url": "https://wpnews.pro/news/kids-are-locked-out-of-ai-the-local-route-keeps-the-decision-with-parents", "canonical_source": "https://tokenstead.ai/guides/claude-age-assurance-local-route", "published_at": "2026-09-11 22:31:38+00:00", "updated_at": "2026-09-11 22:51:42.946196+00:00", "lang": "en", "topics": ["ai-policy", "ai-safety", "ai-ethics"], "entities": ["Anthropic", "Claude", "SB 1119", "Adam Raine", "OpenAI", "ChatGPT", "Yoti", "California"], "alternates": {"html": "https://wpnews.pro/news/kids-are-locked-out-of-ai-the-local-route-keeps-the-decision-with-parents", "markdown": "https://wpnews.pro/news/kids-are-locked-out-of-ai-the-local-route-keeps-the-decision-with-parents.md", "text": "https://wpnews.pro/news/kids-are-locked-out-of-ai-the-local-route-keeps-the-decision-with-parents.txt", "jsonld": "https://wpnews.pro/news/kids-are-locked-out-of-ai-the-local-route-keeps-the-decision-with-parents.jsonld"}}