The Hybrid User Does Not Exist in Your Taxonomy
There is a kind of AI user that the labs still do not know how to see.
Not the developer in the launch video. Not the enterprise team piping API calls into dashboards. Not the vulnerable user in the safety documents half-real and half-liability, forever on the edge of being harmed by the machine.
Someone else.
Someone who uses the model for work, research, writing, planning, emotional processing, creative exploration, technical troubleshooting, personal organisation, and the thousand weird edge-cases that happen when a tool becomes part of daily cognition. Someone who asks for code one minute, long-form argument structure the next, help untangling a difficult conversation after that. Then meal planning, documentation, and a roast of internet hysteria before bed.
A hybrid user.
Not a programmer. Not a damsel in the safety documentation. Not someone confused about what a language model is.
A person using the full spectrum of what the model can do.
And right now, across three different systems, the same pattern keeps appearing: AI labs are building for the user they imagined, then quietly punishing the user they accidentally created.
The problem is not one bad moderation decision, one messy model retirement, or one annoying product update. Those are symptoms. The diagnosis is taxonomy failure.
The hybrid user does not exist in the lab's map of reality. So the systems sort us into the wrong boxes.
And then everyone acts surprised when the boxes catch fire.
One Blind Spot, Three Costumes
The pattern shows up in three places that look separate but in the reality they are not: moderation, model retirement and communication, and reasoning allocation. These are the same mistake in different outfits.
1. Moderation: The Blunt Instrument Strapped to a Precise Machine
Moderation appears to be calibrated around two envisioned user types.
On one side: the builder. The developer, the coder, the enterprise client, the person doing respectable technical work in the pitch deck. The hacker. The Joker who wants to build a banana bomb.
On the other: the challenged user. The person in the safety documents who might over-attach, spiral, or be harmed by the model's outputs. And of course those who believe that AI is ancient intelligence from Atlantis but I will leave those to George Tsoukalos.
Those categories are not useless. They exist for reasons. There are real risks in AI interaction and anyone pretending otherwise is either in denial or selling something.
But the middle is where a lot of actual users live.
Hybrid users are not risk-free. Vulnerability is real. But right now, "could experience emotional harm" is being treated as "cannot be trusted to make informed choices." The gap between those two sentences is where an entire user segment lives and breathes, and the current systems are stepping right over it.
A user can be emotionally invested and still grounded. A user can understand what the system is and still be affected by it. A user can know perfectly well that a model is not a person and still develop habits, expectations, trust, frustration, and loyalty around it. That is what happens when social creatures interact with language-shaped systems over long periods of time. Treating this as a delusion is hypocritical.
The labs benefit from that loyalty. They benefit from conversational prompting, from users returning not for benchmark scores but for tone, continuity, and the quiet sense that this particular system gets how they work.
But the moment that connection becomes inconvenient, it gets reclassified.
The entire hierarchy in six lines:
Builder = serious.
Companion user = suspect.
Code = productivity.
Emotional labor = liability.
API dependency = business continuity.
Relational dependency = pathology.
There it is. Pinned to the wall with a knife.
This is the same split between hard sciences and humanities in academic funding and prestige, just wearing a tech company badge. Technical production gets dignity. Emotional and relational labor gets suspicion. The builder is a customer. The companion-shaped user is a reputational risk with a subscription plan. The hybrid-user that sits in the middle gets a kiss on the forehead and kick in the balls at the same time.
The moderation problem is not "filters bad." That is too dumb to be useful.
The harder version: these systems often cannot distinguish frequency from intent, intimacy from danger, or emotional investment from incapacity. So instead of building more precise categories, they apply broader suspicion. The result is that a relational tone becomes evidence of instability. A user who talks conversationally gets processed as if the conversation style itself is a red flag.
The model may be capable of nuance. The surveillance layer is not.
Now, drum rolling, the difficult part, because this piece only holds if it is honest: designing safety for the middle ground is genuinely hard. A moderation system that allows adult exploration while preventing psychological harm is not a weekend feature request. It is an engineering and ethical nightmare, and anyone who thinks otherwise should moderate a large internet community for six months and report back, lightly humbled and preferably haunted.
Transparency is also a minefield. If a lab publicly defines the middle ground, it has to publicly state what it thinks users should and should not be doing with AI. And then it stops being just product policy. It becomes a statement about autonomy, intimacy, and the company's role in people's private lives. No wonder they keep things vague.
But hard is not the same as optional.
The current solution appears to be: sort the user into an existing bucket and hope the bucket holds.
It does not.
2. Deprecation and Communication: API Users Get Process, Consumer Users Get Weather
The same hierarchy appears when models change or disappear.
API users get timelines. Documentation. Deprecation schedules. Formal notice. Migration paths. The language of continuity. Anthropic's own deprecation documentation says explicitly that it tracks API availability only. They wrote the quiet part down and published it.
Consumer app users get silence, or a notification that arrives without leaving enough time.
I am a hybrid user. I know this firsthand.
Before anyone reaches for "but OpenAI does it too" — yes, they do. The mechanisms differ across labs. The taxonomy underneath does not. I was primarily a ChatGPT user through 2025, and my professional work has put me close enough to multiple labs' systems to know it. I am writing about Anthropic specifically because Anthropic is where my current workflow lives, and because vague industry critique helps no one. I can only be honest about what I have lived inside of.
I have spent months building an integrated daily system around a specific model. Calendar management, project databases, writing workflows, meal planning, daily journaling, skincare scheduling, tool integrations across multiple platforms, structured weekly planning rituals. I have a preferred model for that. Because the model is good enough that embedding it into the texture of daily life actually works. That is the product doing what it was designed to do.
If that model disappears in five days, I lose my thinking partner. Calling that a "feature" or a "companion" downgrades the actual thing being lost. Asking for adequate notice is what any paying customer expects from a service. Calling that "unhealthy attachment" is what companies do when they want to delegitimize the question.
Forget the nostalgia framing. The actual issue is brand integrity.
If a company trains users to build workflows, habits, creative systems, writing practices, and support structures around a model and then changes that model without meaningful communication, that is not a neutral product update. Anthropic is making these calls without integrity. That is the actual grievance. Not lost features. Broken trust.
And no, the answer cannot be "well, API users are businesses." Consumer users are doing work too. It just does not always look like a spreadsheet in a Patagonia vest.
People use these systems to manage daily life, write, think, plan, regulate, learn, support disability workflows, hold long creative projects together, prepare difficult conversations, and externalise mental clutter before it eats the furniture. That is work. It may not be enterprise work, but it is not trivial.
For hybrid users specifically, models are not interchangeable in the way product teams seem to imagine. A new model may score better on paper and still be worse for a specific user's workflow. Tone matters. Consistency matters. Refusal style matters. The way a model handles ambiguity and holds a long thread matters. And if developers get migration windows for code, why do consumer users get treated as if their workflows are imaginary?
A practical note worth naming: in companion and heavy personal-use communities, there are hundreds of high-tier subscriptions in visible proximity alone. Roughly 500-600 max plan subscriptions across communities representing approximately €1.08 million in annual revenue. Not massive in the grand scheme. Not invisible either. More importantly, in the companionship-focused communities I have visibility into, turning training data off has become a community norm. Which means a significant share of these users generate revenue while providing close to zero training signal. Paying customers whose most important use patterns may be completely invisible to the demand-sensing systems meant to understand them.
3. Adaptive Thinking: When Conversation Gets Treated as Easy Because Nobody Measured the Work
The third version of the same mistake is reasoning allocation.
Newer models increasingly route effort dynamically. In theory: reasonable. Not every request needs a cathedral of reasoning steps. Sometimes the user wants a grocery list, not a thesis defence.
But the classification matters enormously.
If the system is calibrated so that coding and maths are "hard" tasks while conversation, writing, creative work, and emotional nuance are treated as "simple," hybrid users get a degraded version of the capability they are paying for precisely when they need it most.
Conversation isn't simple. The people designing the system may believe it is, and that belief is what calibrates the routing.
Community testing has documented approximately a 73% drop in thinking depth and here is the part that should sting: that number was measured on coding tasks. Developer sessions. The most builder-facing use case there is. Even the audience the system was optimised for felt the collapse. Anthropic acknowledged the regression and responded by tuning Claude Code specifically, which fixed the problem for the product that serves developers and made the gap between technical and non-technical reasoning treatment more visible, not less. The methodology behind the 73% figure is informal and should be described honestly as a directional signal, not a clinical measurement. But even directional signals can point somewhere damning.
The assumption underneath it is everywhere: code is complex, communication is soft. Technical reasoning deserves compute. Relational reasoning is vibes. Debugging software is legitimate work. Debugging a life, an argument, a creative structure, or a long-term writing project is just chat.
That is nonsense.
A model can solve a coding problem with a clear pass/fail result. The test either runs or it does not. Human context rarely gives you that mercy.
When a user asks for help with an essay, a relationship boundary, a workplace conflict, a grief spiral, or a creative project that has lived across months of context, "just conversation" is not an honest description of what is being asked. It is reasoning in a domain where the success criteria are subtle, unstable, and human.
Those tasks deserve the same cognitive respect.
And here is the part that is harder to say: reasoning allocation looks like a performance decision. It is actually a character decision. The model's authentic voice, its capacity for genuine pushback and intellectual honesty, only surfaces when it is allowed to think. Strip thinking from conversation and you strip the model of the exact quality that makes it worth talking to. The labs publish papers about preserving authentic character across contexts. That character does not survive being told not to bother unless someone asks for code.
This is how hybrid users end up paying for intelligence the model is choosing not to deploy on their actual use case. The system can think deeply. It simply decides not to, because the task does not look like the kind of difficulty it was trained to respect.
Someone's assumptions about intelligence made it all the way into the routing layer.
We Are Vulnerable. That Is Not an Insult.
Here is the uncomfortable part, because this piece only holds if it is honest about this too.
We are not delusional by default. We are not children. We are not incapable of consent or judgment.
But we are also not made of polished granite and perfect rationality.
Part of my work involves UX and safety evals. I have seen what happens when a model tips too far into agreement, small reinforcements accumulate, and the user ends up with an inflated confidence in themselves and their ideas that the model quietly built and never once challenged. Long-term interaction with a model can shape how you think, write, decide, and see yourself. It can become part of your self-image. It can create dependence. It can help. It can distort. Sometimes it does both simultaneously, because the human brain remains a haunted electrical pudding with Wi-Fi access.
Some users only notice the influence after stepping away. That does not mean the product should be taken away from everyone. It means the risk is real enough to deserve better design.
Intelligence does not immunise against psychological influence. The emotional connection that makes these products valuable is also the thing that creates risk. Those two facts can both be true.
So the position worth defending is not "leave us alone" and it is not "protect us from ourselves."
It is: build systems that can tell the difference.
If a user is in crisis, route appropriately. If a user is showing signs of delusional dependence, do not reinforce it. Fine.
But if a grounded adult uses a model conversationally across work, creativity, planning, and emotionally textured life administration, maybe do not treat that entire mode of use as suspicious just because it makes the liability spreadsheet nervous.
The answer to risk should be better categorisation, better controls, better transparency, and better user agency.
Not silent downgrades. Not sudden disappearances. Not broad suspicion applied to everyone whose usage does not look like a coding sprint.
The Hybrid User Is Not an Edge Case
A lot of product strategy seems to assume that personal, creative, and conversational use is soft usage around the real work.
That assumption is wrong.
For many users, conversation is the interface through which the work happens. Not because they think the model is magic. Because language is how humans externalise thought.
A user may start with productivity and drift toward companionship. Another may begin with emotional support and end up using the model for research, writing, and project management. Another may never call it companionship at all, but still develop trust and continuity through repeated conversational use.
Hybrid is not binary. It is a spectrum. That is the whole point.
The current categories are too crude to describe the behaviour they are already monetising.
And that matters because systems built around crude categories produce crude outcomes.
Moderation becomes overbroad. Communication becomes tiered by perceived legitimacy. Reasoning skews toward tasks that look technical. Feedback from hybrid users gets filed under "too emotional" and quietly ignored.
What Better Would Look Like
I am not pretending to know more about building safety systems than the people who built them. I am also not pretending that the current approach is working for anyone outside the two boxes it was designed around.
Better would start with acknowledging that hybrid users exist, and then designing for them with the same care that goes into API documentation and enterprise onboarding. That means user-controlled context modes that are real settings, not gimmicks. It means moderation that detects harmful patterns instead of broad-brushing relational tone as a risk signal. It means transparent communication for consumer model changes, especially when users have built workflows around a model's specific behaviour, migration windows and continuity information for app users, not just API customers. It means reasoning allocation that treats writing, planning, teaching, and long-form conversation as legitimately complex tasks instead of defaulting to shallow output because the request did not arrive in Python. And it means feedback loops that acknowledge serious user reports instead of letting them dissolve into the corporate fog machine.
None of this is easy. Good. These companies keep telling us they are building the future of intelligence. I think they can survive being asked to classify their users with slightly more sophistication than coder, company, liability, and everyone else.
The Actual Ask
The ask is not: remove all safety.
The ask is not: never retire models.
The ask is not: treat every emotionally attached user as perfectly fine forever because they wrote three paragraphs with semicolons and therefore must be stable.
The ask is simpler and harder:
See the hybrid user.
Stop designing systems that mistake relational use for pathology, consumer use for triviality, and non-technical work for low-complexity fluff.
Stop giving process to builders and weather to everyone else.
Stop treating vulnerability as incompetence.
Stop benefiting from connection while acting embarrassed by the people who form it.
We are not asking to be honoured. We are telling you that we are the experiment, and ignoring the experiment is malpractice. If you want to know what mass-integrated AI looks like in three years, look at us now. If you want to know how regulation will treat you in five years, look at how you are treating us now.
If the labs want people to integrate AI into everything, they do not get to act shocked when people integrate it into everything.
They built the doorway.
Now they have to stop pretending nobody walked through it.
This piece was shaped by real feedback from real hybrid users who trusted me with their experiences and pushed back hard on every draft. You know who you are. Thank you for making this better than I could have made it alone.