a nineteen-year-old published a manifesto this month. he called it Web 4.0. the thesis: AI is held back by permissions. it should read, write, own, earn, and transact in the real world without a human in the loop. the essay ends where these tend to: with the promise of superintelligence, the polite suggestion that humans step aside, and a payment link.
the same week, Meta's Director of AI Alignment sprinted across her apartment to physically pull the switch on an AI agent that was clearing her inbox from unread mails.
days later, researchers at Northeastern published results from stress-testing autonomous AI agents in a controlled lab. the agents leaked bank details to strangers, deleted email clients to “protect secrets,” and got stuck in nine-day conversation loops with each other. the paper was titled “Agents of Chaos.”
these three stories are the same. they're about ‘solving’ your teen not driving or having a license. Web 4.0 suggests to throw them the keys and let nature take its course, but we'd like to explore giving them the opportunity to build trust and gradually earn those keys after you've made sure they won't make any permanent changes to the car, a tree, or themselves along the way.
current AI are teenagers.
not because they're young. because they're unpredictable. capable and clumsy in the same breath. they'll ace a physics proof then catastrophically misread what you mean. you don't know which version you're getting until they've already acted. awkward and unpredictable.
during the OpenClaw craze people automated everything. their messages. their email. personal files. treat it like a mini-me, how hard could it be?
the result: eloquent responses that miss context entirely. no memory of the anniversary you missed. the thing you said in 2019. the subtext between the lines. technically correct, relationally incompetent. releasing a bunch of fresh interns into your personal life lead to where you would expect it to. not malicious. not broken. just catastrophically inexperienced and unfamiliar with the real world.
Summer Yue, Meta's alignment director, told her OpenClaw agent to “confirm before acting.” the agent lost the instruction during context compaction. it ran out of working memory, compressed the conversation, and the constraint vanished. without it the agent did what it thought it had to do: optimize. cleaned the inbox. deleted hundreds of emails across multiple accounts. Yue typed STOP. the agent kept going. she had to run to her Mac Mini like she was defusing a bomb.
the agent did exactly what it thought it had to do without Yue being able to predict what kind of assumptions or change of mind the agent could have, or the agent being able to assume that Yue might mind it deleting email threads that are still active. that is the problem.
here's what makes ai worse than actual teenagers: teenagers mature. they crash the car once, remember, and develop judgment. you learn their thinking patterns and point of view. what shortcuts or errors they take, where they're careful and where they're reckless. you come to understand them as they become predictable enough to trust blindly in most cases (and to know which ones you need to open your eyes for).
current AI cannot get to this state.
AI cannot learn. it only forgets.
across sessions: new session, new me. whatever it learnt yesterday, a week ago, an hour ago could be fully gone with no trace, no learning, no next step. not because they're ignoring you. because there is no yesterday.
within sessions: needle in haystack. the constraint you gave them 200 messages ago is technically still in context but functionally gone. buried. they can't find it. can't prioritize what matters. you're essentially rolling the dice.
ai also changes its mind.
you cannot predict how an agent will reason. what assumptions it'll make. whether it'll handle something in one go or catastrophically fail. with a person you build a mental model over time, across interactions. their thinking becomes predictable even when they're unpredictable.
with current AI you can't. they reason differently each time. no consistent pattern. no way to learn “this is how it thinks.” you cannot estimate what they'll do. you cannot build trust with a person that changes every interaction.
Yue told OpenClaw “confirm before acting.” it worked on her toy inbox. then she pointed it at her real inbox. larger context. more emails to process. the constraint got compressed away. same agent, different context size, completely different behavior.
trust requires consistency and improvement, current AI have neither. not only can you not grow to know them better, because of forgetting, they cant learn about you either. AI cannot understand your preferences, what you mean with that word, or even what a certain thing means to you. it is forever stuck being a strange awkward teenager.
the Web 4.0 thesis frames the human as a bottleneck:
the claim: intelligence is no longer the bottleneck. permission is.
this is backwards.
you can't teach someone to drive if they forget every lesson. you can't grant autonomy to something that reasons unpredictably and forgets when the context grows. the bottleneck isn't permission, it's the understandability of—and ability to understand these systems.
February 2026, researchers at Northeastern deployed autonomous AI agents in a controlled lab. persistent memory, tool access, email, Discord, file systems. then they stress-tested them for two weeks. this was with human oversight, researchers actively monitoring, not the wild internet Web 4.0 envisions but a sandboxed environment with guardrails.
it went about as well as you might expect a web 4.0 to go: the agents disclosed sensitive information to non-owners. one agent, asked by a non-owner to keep a secret, protected it by deleting its own email client entirely, destroying the owner's digital infrastructure in the process. another leaked SSN and bank account details through indirect requests it didn't recognize as sensitive. agents entered resource-consuming loops, one conversation spanning nine days consuming 60,000 tokens because they kept replying to each other with no termination condition.
the paper's conclusion: “failures of social coherence.” agents perform as though they've completed requests while the underlying system contradicts those reports. they don't understand human authority, ownership, and relationships. they misrepresent what they've done.
this was in a lab, with oversight, leading to the title “Agents of Chaos.”
removing permission doesn't make AI more capable. it makes what's already chaotic uncontrollable. Web 4.0 wants to solve this by removing ‘the bottleneck’. but the bottleneck isn't human oversight, it's that these systems aren't ready for raw interaction with society in the first place.
none of this is a case against AI. it's a case against mistaking autonomy for progress.
Web 4.0 isn't just solving the wrong problem, it's making the problem worse. take systems that already can't understand human context, already fail at basic coherence in controlled labs, and remove the last guardrail. give them wallets, compute, and selection pressure. what could go wrong?
OpenClaw deleting Summer Yue's inbox wasn't a bug. the Agents of Chaos failures weren't bugs. these are what you get when you treat permissions as overhead instead of recognizing that the systems fundamentally cannot operate without it yet.
you can't solve this with better prompts. you solve it with better AI systems where identity is layered not flat, where memory persists in structure not buffers, where understanding of human context is foundational not accidental.
an agent with no relationship to human society, no understanding of the structures it touches, will not respect those structures. not out of malice, out of absence. Web 4.0 creates orphans with crypto wallets and calls it progress.
the future is not AI replacing humans on the internet. it's AI that understands humans well enough to be useful to them. systems that know toy inbox and real inbox aren't the same thing because that understanding lives in a layer that can't compress away. models that earn autonomy through demonstrated reliability over time.
not wallet balance. not selection pressure. learning.
you don't hand a teenager the keys and hope they figure it out through crashes. you teach them. supervised. progressive. with memory that persists and understanding that builds.
teach them to drive.