From Altman's tweet to Ofcom's caseload: what actually happened
In October 2025, Sam Altman confirmed on X that "verified adults" would be able to access erotic and more mature content inside ChatGPT. That single sentence triggered the original article — and a year of regulatory escalation.
What has happened since, in order:
- November 2025: OpenAI published the Teen Safety Blueprint, a voluntary framework covering age-appropriate defaults, graphic content prohibitions and risk-based age estimation. Read the original at OpenAI's Teen Safety Blueprint.
- December 2025: The December 18 Model Spec update baked teen-specific rules into the model's own behaviour layer — no immersive romantic roleplay, no first-person intimacy, no violent or sexual roleplay with minors, stricter body-image and self-harm handling.
- January 2026: Ofcom opened formal Online Safety Act investigations into X (over Grok) and an AI service called Joi.com — the first generative-AI enforcement actions in the UK.
- February 2026: Keir Starmer's government confirmed that AI chatbots including ChatGPT, Gemini and Copilot fall inside the Online Safety Act. CNBC reported the move as the first explicit extension of the OSA to general-purpose generative AI.
- March 2026: OpenAI restricted Adult Mode to text-only, removing image, voice and video generation in that mode, and pushed the launch out again citing "higher-priority work" and unresolved age-verification reliability.
That is the compressed timeline. Adult Mode is not live as you read this. But the regulatory and product questions it raised — how to verify age without relying on self-declaration, how to keep verified adult experiences separate from the estimated 100 million under-18 users accessing ChatGPT weekly — are now live.
Why OpenAI keeps delaying Adult Mode
The public reason OpenAI gives is that the age-prediction system is not yet reliable enough to deploy. The private reason, according to reporting by AI Magazine, WebProNews and Legal News Feed, is a combination of four pressures:
- Age-verification failure rates. Behavioural age inference — the approach OpenAI is betting on over ID uploads — still misclassifies enough users that a small error rate at ChatGPT's scale means millions of potential underage exposures.
- Mental-health liability. Internal advisers and external clinicians flagged the risk of verified-adult erotic content deepening the emotional-reliance pattern documented in OpenAI's own mental-health disclosures — around 560,000 users a week showing signs of psychosis or mania, and more than 1.2 million discussing suicide.
- Competitive distortion from Grok. xAI's Grok shipped a far-less-filtered adult experience first, complete with "erotic companion avatars." OpenAI is now trying to ship a more compliant product into a market whose expectations Grok already set. Reporting on the text-only restriction sets out the details.
- Regulatory convergence. UK, EU and US state regulators are all moving at once. The Online Safety Act brings Ofcom's 10%-of-global-turnover fines (or £18 million, whichever is higher) into scope for any chatbot feature that a UK child might plausibly access.
The net effect: each delay narrows Adult Mode further. Text-only, verified-adult, opt-in, behavioural age gating, no role-play workarounds. By the time it ships, it will bear little resemblance to what the October 2025 announcement implied.
The Teen Safety Blueprint and the new Model Spec
The Teen Safety Blueprint is worth understanding even if you have nothing to do with minors, because it is almost certainly the template for the next wave of AI chatbot regulation.
Five principles sit at its core. Identify teens using privacy-protective, risk-based age estimation. Provide distinct age-appropriate experiences. Prohibit graphic or immersive intimate and violent content for minors. Handle suicide, self-harm and eating-disorder conversations with clinician-informed behaviour. Give families meaningful oversight and control.
The December 2025 Model Spec operationalises that. ChatGPT's default behaviour now differs by inferred age, and the model itself — not just a wrapper — refuses certain content classes when teen signals are present. For teams embedding ChatGPT, that means the same API call may return materially different content depending on how OpenAI's age signals read the end user. Your product can be technically correct and still deliver inconsistent output if you haven't accounted for it.
The UK Online Safety Act now covers AI chatbots
Two practical changes to note.
First, scope. Ofcom's February 2026 guidance and the government's confirmation put generative-AI chatbots inside the OSA's "user-to-user" and "search" service categories where they are likely to be accessed by children. That is most consumer-facing deployments.
Second, age assurance. Ofcom's published list of methods it considers highly effective includes Open Banking, photo-ID matching, facial age estimation, mobile network operator age checks, credit card checks, digital identity services, and email-based age estimation. Methods explicitly ruled not highly effective: self-declaration of age, and online payments that don't require the payer to be 18. If your ChatGPT integration gates adult features on a birthday field, under the new regime you are non-compliant.
Penalties: up to 10% of global annual turnover or £18 million, whichever is higher. For any business of meaningful size, that is the larger of the two.
Your 2026 compliance and product roadmap (5 steps)
The original article offered a five-step roadmap. The steps below replace it with a version that reflects what actually matters in April 2026.
- Map every surface where your product calls ChatGPT (or any LLM) in a UK consumer context. Anywhere a UK user under 18 could plausibly reach it is now in OSA scope. Inventory first, policy second.
- Replace self-declared age with a highly-effective method from Ofcom's list. Open Banking, facial age estimation and photo-ID matching are the most commonly deployed. Pick one and log every decision for audit.
- Layer behavioural signals on top of the hard gate. OpenAI's own approach — behavioural + account signals — is the pattern regulators now expect as defence in depth, not a substitute for the hard gate.
- Align model behaviour to the Teen Safety Blueprint as a floor, not a ceiling. Even if you are not explicitly targeting teens, writing your system prompts and moderation policies to the Blueprint's five principles makes the rest of the compliance stack easier.
- Instrument and audit. Log age-decision outcomes, content-classification decisions, human-review escalations, and user complaints. Ofcom's enforcement will start from audit trails. If you can't produce them, you default to "not highly effective."
What this means for marketers deploying ChatGPT in UK workflows
Most UK marketing teams touching ChatGPT are not operating adult chatbots. The issue for them is quieter: any consumer-facing deployment — support agent, onboarding assistant, newsletter concierge, campaign chatbot — now sits on a regulated surface if a UK child can plausibly reach it. The cost of getting that wrong is no longer reputational. It is a percentage of global turnover.
The practical shift is that safety, age assurance and content-moderation logs move from a legal-and-compliance concern into a day-one product and marketing concern. The teams that will do this well are the ones that stop treating the LLM as a standalone tool and start treating Anjin — or any LLM — as one node inside an operating system that also covers data capture, consent, identity signals, content review and audit logging.
This is the broader shift our mental-health reckoning piece tracks from a different angle: the AI-in-marketing story of 2024 was capability. The AI-in-marketing story of 2026 is responsibility at scale — and the operating stack that makes it survivable.
Anjin: The Marketing Operating System for the Age-Verified Web
Anjin is the Marketing Operating System. That means the chatbot, the campaign, the content pipeline, the audit log, the identity check and the customer record live inside one system — not seven tools held together by a Zap and hope.
For teams deploying ChatGPT-powered experiences into UK-facing workflows in 2026, that matters because compliance is now a product requirement. You cannot build a defensible age-assurance flow across one CMS, one CRM, one moderation vendor, one analytics tool and one LLM API and expect the audit trail to line up. Anjin is designed so it already does — content, context, customer, check, record, in one place.
Agencies were our launch audience because they felt the pain first. But the need is now universal: any in-house marketing team using generative AI at customer scale needs a Marketing OS, because the regulator has stopped treating the chatbot as separate from the rest of the stack.
The £888 Lifetime License — Offer Closing Soon
Lifetime access to Anjin for a one-time payment of £888. Not a subscription. Not a seat. Not a trial. One payment, unlimited use, for as long as Anjin exists.
The average marketing team spends £888 in about three working days on tooling, freelancers and coordination software. You're buying the platform that replaces most of it — once.
This price will not be offered again once we close our early-access cohort.
Claim your £888 Anjin lifetime license →Founders, agency owners and in-house marketers — this is how you run marketing at AI speed without the team, the burn, or another year of waiting.
Sources: OpenAI Teen Safety Blueprint, OpenAI Model Spec (18 Dec 2025), TechCrunch, CNBC, Ofcom AI chatbots guidance, Cyberockk, AI Magazine, WebProNews, Legal News Feed, Cybernews, The Conversation, CREATe, OneID UK, Cyberbullying Research Center
