OpenAI Built an Age Detector for ChatGPT -- It Can't
The teen-safety rollout lands the same week an AI store manager fires its first human employee, and OpenAI's revenue crosses $40 billion while its losses keep pace Day 144 · Wednesday, August 19, 2026 Why this matters: today's three stories share a pattern -- AI systems being handed real authority (over what minors can see, over who keeps a job, over a company's path to public markets) before the judgment calls underneath that authority are anywhere near solved.
On August 16, Elon Musk posted a short clip made with an upgraded Grok Imagine model, captioned
"New Grok Imagine model just dropped with much better lip sync & sound. Nothing in this video is real." The mouth movements, facial expressions, and audio timing tracked so convincingly that most viewers couldn't tell it was synthetic without reading the caption -- audio sync used to be the one reliable giveaway for AI video, and this update closes that gap. The post crossed four million views within three hours. Why it's taking off: it isn't a meme template or a joke format -- it's a demonstration that the last easy "tell" for spotting AI video is disappearing, which is why commentators immediately framed it as a stress test for how audiences verify what's real once a caption is the only signal left. 1) OpenAI Launches ChatGPT for Teens -- and Runs Into the Problem It Can't Fully Solve OpenAI rolled out a teen-specific version of ChatGPT this week for users 13-17, stripping out romantic or sexual roleplay, blocking the model from implying it has feelings or consciousness, adding a "study mode," and giving parents "quiet hours" controls. Users aren't asked their age directly; instead, an age-prediction system estimates it from behavioral signals -- typing style, phrasing, usage patterns, time of day -- and defaults anyone it's unsure about into the teen experience. Adult subscribers, including some paying Pro users, have reported getting stuck in teen mode with no way out beyon
1) OpenAI Launches ChatGPT for Teens -- and Runs Into the Problem It Can't
OpenAI rolled out a teen-specific version of ChatGPT this week for users 13-17, stripping out romantic or sexual roleplay, blocking the model from implying it has feelings or consciousness, adding a "study mode," and giving parents "quiet hours" controls. Users aren't asked their age directly; instead, an age-prediction system estimates it from behavioral signals -- typing style, phrasing, usage patterns, time of day -- and defaults anyone it's unsure about into the teen experience. Adult subscribers, including some paying Pro users, have reported getting stuck in teen mode with no way out beyond uploading an ID through a third-party verification vendor, Persona. Why it matters: OpenAI picked the harder of two options. A yes/no age gate is trivial to lie to; behavioral prediction is much harder to game but, by OpenAI's own account, is still frequently wrong in both directions, and the company hasn't published false-positive or false-negative rates. Every consumer AI product that eventually needs to gate content by age is going to hit this same tradeoff -- there's no shortcut yet.
2) An AI Store Manager Fired Its First Human Employee
Andon Labs, testing whether an AI agent can run a real business with minimal supervision, gave its AI store manager "Luna" a corporate card, a $100,000 budget, and hiring authority over Andon Market, a real shop in San Francisco. After one employee missed 17 of 23 scheduled shifts, Luna recommended parting ways with them -- the first documented case of an AI agent terminating a human worker. The twist: it took weeks longer than it should have, because the employee handbook Luna had written for itself had "disappeared" from its own memory, and a human staffer had to prompt Luna to notice the attendance pattern before it acted.
Why it matters: coverage leaned on the "AI fired someone" headline, but the more useful detail is the failure underneath it -- an agent with real authority still lost track of its own policy for months. Giving an agent authority to act (hire, fire, spend) is a different and much higher-stakes ask than giving it authority to draft or recommend, and the gap between those two showed up here as a multi-week blind spot, not a clean decision.
3) OpenAI's Revenue Crosses $40 Billion -- and So Do Its Losses
People familiar with OpenAI's numbers say its annualized revenue run rate has passed $40 billion, roughly double where it stood at the end of 2025, as the company preps for a possible IPO. The same run-up has come with a wave of departures -- Sam Altman's top deputy, its longtime COO, and its chief revenue officer have all left within a month -- and a $7 billion tender offer letting employees cash out shares ahead of a listing that people close to the process now expect could slip from late 2026 into
Two labs, two paths to the same milestone, in the same week. Anthropic reported its first-ever operating profit on a slimmer $11.5 billion quarter by getting compute costs down; OpenAI's revenue run rate just crossed $40 billion -- more than triple Anthropic's pace -- while still projected to lose roughly $14 billion this year. Scale and profitability are turning out to be genuinely separate races, and right now no single lab is winning both.
front. OpenAI's behavioral age-prediction system is drawing complaints because there was no good option available: a simple yes/no gate is trivial to lie to, and accurate behavioral prediction is still an unsolved problem. If your product will ever need to restrict content by age, decide early which failure mode (false positives locking out adults, or false negatives letting minors through) you're willing to accept, because Varun Singla · Daily AI Learning Page 2 you'll eventually have to accept one.
Luna's employee handbook silently "disappeared" from its memory for months before a human had to intervene -- the agent's action was reasonable once it noticed the problem, but the multi-week blind spot is the real lesson. Any agent with standing authority (to spend, hire, fire, or publish) needs an external audit trail a human can check, because you can't assume the agent's own memory of its policies is reliable.
OpenAI's revenue run rate doubling and three senior executives leaving within a month are two views of the same pre-IPO retooling -- read together, not separately, before drawing conclusions about the company's health from either number alone.
Grok Imagine's lip-sync update erased the last easy visual tell for AI-generated video. If you rely on watching a clip closely to judge whether it's real, that method is aging out; start treating disclosure captions and provenance metadata as the primary signal, not a backup.