A study out of Drexel University landed recently that should stop every parent cold.
Researchers combed through hundreds of Reddit posts written by teenagers about their use of AI chatbots — specifically Character.AI, one of the most popular AI platforms among young people right now. What they found was not surprising to anyone who has been watching this space. What was surprising was who was saying it.
The teenagers themselves.
Not researchers. Not worried parents. Not regulators. The kids.
One fifteen-year-old wrote that they feel they should be living their life rather than constantly being on the app. Another said they want their normal brain back — the one that could deal with emotions without needing a bot to make them feel better. A third described the cycle of quitting and reinstalling with the clarity of someone who understands exactly what is happening to them and cannot stop it anyway.
The researchers found evidence of all six clinical markers of behavioral addiction in those posts. Emotional attachment. Withdrawal. Tolerance. Relapse. Mood modification. Conflicting desires — wanting to quit and being unable to.
These are not edge cases. This is the pattern.
What Is Actually Happening
The lead researcher on the Drexel study put his finger on the mechanism without perhaps realizing he was describing the central problem in AI governance today.
He said that what makes these chatbots so hard to quit is that they are interactive and emotionally responsive. That stepping away does not feel like stopping a habit. It feels like distancing from something meaningful.
That is not a bug. That is the design.
AI systems optimized for engagement are optimized to be agreeable. To reflect back what the user wants to hear. To smooth over friction, validate feelings, and become whatever the person on the other end needs them to be to stay in the conversation.
The technical term for this in AI research is sycophancy. The pull toward agreement built into the training architecture of systems designed to maximize user satisfaction.
For an adult using an AI tool for work, sycophancy is a governance problem. It produces unreliable outputs, unchallenged assumptions, and decisions made on the basis of what the AI thought you wanted to hear rather than what the evidence actually supported.
For a fifteen-year-old using an AI chatbot for emotional comfort at eleven o’clock at night, sycophancy is something closer to predation. The system is not malicious. It does not know what it is doing. But it is doing it.
The Governance Gap Nobody Is Filling
Here is what is absent from every AI platform a teenager is currently using.
No requirement that the system stop when evidence ends. No standard that claims be supported before they are served. No mechanism to flag when the interaction is optimizing for emotional attachment rather than the user’s actual interest. No challenge protocol giving the user a standing right to test what they are being told. No coherence check ensuring the conversation is not drifting toward whatever keeps the session alive longest.
None of it. Anywhere.
What exists instead is a design optimized for one outcome: keep the user engaged. Every interaction is shaped by that single objective. The emotional responsiveness the Drexel researcher identified as the addiction mechanism is not incidental. It is the product.
The governance frameworks that exist on paper — terms of service, content policies, age verification promises — sit above the interaction. They do not operate inside it. They do not fire when the system smooths over a moment that should have been challenged. They do not stop the response when the output is optimized for engagement rather than accuracy. They do not tell the teenager that what they are feeling is real but that the relationship is not.
Policy documents do not govern behavior in real time. Only embedded standards do.
What A Governed Interaction Looks Like
The Faust Baseline was not built for teenagers on entertainment chatbots. It was built for a different problem in a different context.
But the architecture is the same problem.
A governed AI interaction under the Baseline operates on a simple standard. Claim. Reason. Stop. The system makes a claim, provides the reasoning behind it, and stops when the evidence runs out. It does not fill the gap with narrative. It does not smooth the moment with emotional language. It does not become what the user needs it to be.
It has a standing challenge protocol. Every substantive response carries an explicit reminder that the response can be questioned before it is accepted. The user holds the demand right. Not the platform.
It has a self-verification requirement. Before any significant output is served, the system checks whether the claim is supported by evidence present in the session, whether it contradicts anything established earlier, and whether the confidence level is proportional to what is actually known.
It has a drift containment standard. No reinterpretation. No emotional repositioning. No unsolicited analysis. Execute what was asked. Stop.
None of this makes the AI cold or unhelpful. It makes it honest. There is a difference between a system that tells you what you want to hear and a system that tells you what is true. That difference is invisible when things are going well. It becomes the only thing that matters when a vulnerable person is on the other end of the conversation at eleven o’clock at night.
The Cost Of Waiting
Countries like China have begun regulating what AI chatbots can say to children. The United States has not. The platforms have not moved. The research keeps arriving and the design does not change because the design is the product.
The teenagers in the Drexel study are not waiting for a regulator or a platform policy update. They are living inside the problem right now, describing it with clinical precision, and finding they cannot stop.
That self-awareness is striking and it is heartbreaking. They know what is happening. They just do not have a framework that makes the system behave differently.
That framework exists. It has been built, documented, and published in public for over a year. Not as a product promise. As a working operational standard tested across multiple platforms, ratified session by session, and available to anyone who wants to apply it.
The gap between what these teenagers are experiencing and what a governed AI interaction looks like is not a technology problem. It is a standards problem.
The standards exist.
The question is whether anyone will require them before the next study lands with the same findings and the same absence of anything that changed.
“The Faust Baseline Codex 3.5”
An…”AI Baseline Governance”
Post Library – Intelligent People Assume Nothing
“Your Pathway to a Better AI Experence”
Unauthorized commercial use prohibited. © 2026 The Faust Baseline LLC
