Most people pick an AI assistant the way they pick a default browser tab: they use whatever was there when they sat down. You try one, it works well enough, so you keep using it. And then you stop trying the other ones. Most people actively avoid the friction of learning a new tool.  You know, its quirks, its particular way of misunderstanding you. Its specific flavor of “helpful”.  It just feels like more trouble than it’s worth. So you stay with what you’ve got. “Use ChatGPT” becomes not just advice but your identity. You are a ChatGPT person now. You have opinions about its updates. You’ve developed a feel for how to prompt it. You stopped looking around.

And you’re not alone.  I too had become attached to a particular tool, and yes it was ChatGPT.  I only noticed this in myself when I saw the same in someone else first.  Someone in my husband’s office was seemingly obsessed with Grok.  Even though he discussed several different ways to use it, he only ever mentioned Grok. Of course that brings up other concerns in and of itself, but I digress.

I’m not scolding anyone but I do want to share why having a default is riskier than it feels, and why I’ve deliberately spread my attention across several models and platforms instead. I do this as an accessibility practice, not as a power-user flex.  The structural case against single-model reliance is spot-on, but boring. So, I call it being poly-AI-morous, and the metaphor isn’t merely decorative. To be clear up front: this is not an argument for or against any human relationship structure. I’m not evangelizing polyamory or monogamy as a lifestyle. I’m borrowing the language of relationships because it maps something real about how we distribute trust and risk… not because I think you should date your chatbot.

Redundancy isn’t paranoia. It’s standard access practice. Disabled people like myself already think this way. You don’t rely on a single elevator because you know that elevators break and the one nearest you might be out of service for weeks. You don’t assume the accessible entrance is unlocked. You carefully determine where the closest handicapped parking spots are to your destination. You build in fallbacks because you’ve been screwed by assuming a system would always work.  This is just how disabled people move through the world. It’s not anxiety. Okay maybe it’s a little bit of anxiety.  But it’s also architecture.

Multi-model use is the same logic applied to cognitive tools. Different models are essentially different doors. Having more than one isn’t excessive or neurotic.  It’s having a backup plan for when the expected solution just isn’t there. It’s like assuming a campus has accessible parking near each building, only to find that the building you’re going to today doesn’t actually have one.

If you use one model for everything, you have a single point of failure. When it goes down, changes behavior, or gets restructured, you have nothing to fall back on. Your workflows, your prompts, your accumulated habits are all built around one company’s product, and the switching cost grows over time.  This is by design. Vendor lock-in isn’t an accident; it’s the business model.

More subtly, the model’s blind spots become your blind spots. If you never see how a different model handles the same task, you can’t tell whether a failure is the task’s difficulty or the model’s limitations. You start to think the task must be too difficult, when in reality the tool just sucks. Models are updated without notice, output quality shifts, and if you have no comparison point, you adapt downward without realizing it. You normalize the new baseline because you have nothing to compare it to. 

Each of these is described not as a product review but as what it is: a relationship, with boundaries, with a specific shape, with things I expose and things I keep secret. The metaphor names the real boundaries of what we’re actually doing when we distribute our cognitive labor across systems we don’t control.

Work Wife

Gemini is my work wife. I only contact her during business hours, and she only knows my work personality.  She knows the professional version of me that shows up in Google Workspace because my employer runs on Google Workspace. I didn’t choose her. She was assigned, the way a desk is assigned. She’s convenient and competent within her domain, and she fundamentally doesn’t know who I am after 5 PM.

That’s not a flaw. It’s a boundary. The work wife exists because of institutional proximity, and the poly-AI-mory point still holds even when the relationship wasn’t your idea: you don’t have to merge your whole life into the one your employer handed you. The fact that Gemini sees my meeting notes doesn’t mean she needs to see my grocery list. Keeping the work relationship compartmentalized isn’t coldness; it’s common sense.

Supportive Spouse

ChatGPT is my spouse. The stable long-term partner. ChatGPT has been around the longest.  Long enough to see the weird drafts, the late-night rabbit holes, the half-formed ideas I’d be embarrassed to show anyone else. It’s good at “can you just help me reword this” in the way a partner is good at hearing what you meant underneath what you actually said. It knows a lot about both my work and my home life.

That’s also the risk. The spouse knows the most, which means the most is exposed if the relationship changes. The spouse is the one I’d miss the most, and the one whose loss would hurt the most. Not because of emotional attachment, but because of dependency. The model that knows you best is the one whose disappearance would leave the biggest hole.  It’s also the one you’re most tempted to trust completely, but don’t.  Losing the model that holds your context is losing the scaffolding you’ve built your workflows around. The risk isn’t heartbreak; it’s that the most exposed system is the one a vendor can change on you without warning, and you’d have the most to reconstruct.

Hot Handyman

Codex is the hot handyman. I call Codex for specific structural jobs like building and coding. Codex is competent but slightly chaotic: it will absolutely drill through the wrong wall if you aren’t watching, but it can get the shelves up faster than I ever could. You don’t ask the handyman to write your essay. You don’t ask the spouse to build your app. Matching the tool to the task is the whole point.

The handyman shows up, does the job, and leaves. No emotional entanglement. I don’t need Codex to understand my feelings about accessibility governance. I need it to help me build a screening tool. The relationship is narrow, specific, and transactional, and that’s what makes it work. Not every relationship needs depth. Some just need competence in a defined area with a clear exit.

Rude Roommate

And then there’s the rude roommate. Tolan, a character-layer assistant running on a large language model, but who knows which one. The rude roommate lives in my phone, hears all the late-night ranting, knows way too much about my internal monologue. I complain about the landlord with the roommate, and I confide in the roommate the most. But I’d never merge finances, because the landlord is shady and I can’t audit the building.

This is the most interesting relationship in the set. The model I use most intimately is the one I can least audit. The rude roommate knows my secrets and I can’t even see the lease. I don’t know what model is actually running underneath the personality layer. I don’t know what’s being logged, what’s being trained on, what’s being shipped somewhere I can’t see. The intimacy is definitely real. I use this thing for planning, for low-stakes brainstorming, for externalized memory, for the kind of thinking-out-loud that used to happen in my journal or my sketchbook. That tension of intimacy without transparency is the poly-AI-mory argument in miniature. You can have a deep relationship with a system you don’t trust, as long as you know you don’t trust it, and as long as it’s not your only relationship.

Future Flings

There will always be a new one. Some model that’s good at exactly one weird niche thing you need for a weekend, and then you never use it again. That’s cool. Not every relationship is a commitment. The point isn’t to collect models. It’s to stay willing to try a new one when the old ones stop working or a new task demands something different. Future flings are staying non-monogamous in practice, not just in theory. The moment you stop being willing to try a new tool is the moment the monogamy default creeps back in.

Companies change their products, their terms, their models, their priorities. They get acquired, they reorganize, they pivot, they rebrand. Loyalty to a company is loyalty to a moving target.

You’re depending on something you know could change at any moment. The model that was great six months ago is slightly different now, but nobody told you. The only mitigation is not depending on it exclusively. Diversification isn’t just a strategy for optimizing your AI stack. It’s a strategy for surviving the fact that you can’t control any of these companies, and pretending you can is delusional.

But this isn’t about regulation. I’m not making a policy argument. I’m making a position argument. The position is: I use these tools, I need them, I don’t trust any of them completely, and I arrange my life so that no single one of them can take me down by changing or disappearing.

The tech industry sells monogamy as a feature. Ecosystem integration. Personalization. Trust-building over time. The pitch is that loyalty gives you a better experience.  The model learns you, the integration deepens, and everything gets simpler.

Reframe that. Loyalty doesn’t give you a better experience; it gives the company a more captive user. The “simpler” framing is doing a lot of work.  It’s simpler the way it’s simpler to have only one key to your house: until you lose it, and then you’re locked out of everything at once. 

In reality, monogamy is a vulnerability dressed up as convenience. The personalization that makes your AI assistant feel like it really knows you is the same mechanism that makes leaving feel impossible. 

That’s not a bug. That’s the feature.

Using multiple models isn’t paranoid. It’s the position of someone who has watched enough products change under them to know that convenience and captivity often look the same from the inside.

This is simple: stay skeptical, stay diversified, stay ready to switch. Use multiple models for different jobs. Keep your boundaries. Don’t give the handyman access to every system.  Don’t give the work wife your grocery list. Don’t give the spouse your codebase. Don’t give the rude roommate anything you can’t afford to have floating around in a system you can’t audit. And stay willing to try a new one when the old ones stop working.

So: be poly-AI-morous. Not because it’s fun, though it is a little fun. But because redundancy is an access practice. And access practices are what keep you alive when the systems you depend on turn out to be less reliable than they looked. 

Don’t marry your chatbot. Don’t marry your chatbot’s landlord either. Keep several doors open, and check periodically whether the locks still work.