Breaking Up With Software Is a Genre Now
- Rich Washburn
- 2 days ago
- 7 min read
I saw a TikTok where a woman was basically announcing that she was done with Claude. Not because the model was down. Not because it hallucinated something important.
Not because Anthropic changed the pricing or ruined the interface. Claude apparently suggested the wrong car for her. And not just the wrong car. The kind of wrong car that made her question whether Claude had ever understood her at all.
That sentence would've sounded completely insane five years ago. Actually, forget five years ago. It would've sounded insane about software basically any time in human history.
Nobody ever said: "I'm done with Excel. After everything we've been through, it just doesn't get me."
Nobody had a falling-out with Photoshop. Nobody opened Firefox one morning and thought, You've been weird with me lately.
Software could be bad. Software could be annoying. Software could be buggy enough to make you invent entirely new combinations of profanity. But software couldn't disappoint you personally. Apparently, that's over.
We break up with software now. And I think that's actually a thing. "Fuck Claude. He's a Dick."
The old reasons for switching software were incredibly boring. Better features. Better interface. Lower price. Faster performance.
"I switched from Chrome because it eats RAM."
"I moved to Mac because I like the ecosystem."
"I use this photo editor because the other one's UI is terrible."
Fine.
Then generative AI showed up and the conversation changed slightly.
For a while it was:
"I use Claude for writing."
"ChatGPT's better at this."
"Gemini has a bigger context window."
"This model scores higher on whatever benchmark we're pretending matters this week."
Still basically software evaluation. But listen to how people are increasingly talking about these systems now.
"ChatGPT gets me."
"Claude's been acting weird."
"Gemini doesn't really understand how I work."
"I tried another one, but it didn't know anything about me."
Or, increasingly:
"Fuck Claude. He's a dick."
That's different. We've somehow gone from product preference to relationship preference. And I'm not entirely kidding about the word relationship.
The Mirror Has Expectations Now I wrote an article called The Mirror Test. The premise was simple.
If you've used an AI extensively for years, across hundreds or thousands of conversations, you've unintentionally created something like a longitudinal behavioral record.
You've shown it what you repeatedly care about. What pisses you off. What excites you for three days and what you're still talking about three years later. How you make decisions. Where you contradict yourself. What kinds of problems you naturally gravitate toward. And, importantly, you're giving it behavior rather than filling out a personality questionnaire describing who you'd like to think you are.
That's why I described the AI as a kind of accidental mirror. The interesting part wasn't whether the AI was conscious. It wasn't. The interesting part was that enough accumulated conversational context could produce a surprisingly coherent reflection of the person using it. But I think I stopped one step too early.
Because something happens after the mirror becomes good enough. You start expecting the reflection to be accurate. And that's where things get weird.
"You Should Know Me Better Than That"
After we were laughing about the Claude breakup TikTok, I decided to test this.
I told ChatGPT: Okay, based on everything you know about me, what kind of vehicle have I owned most of my life?
I was certain it would get it wrong.
There's an obvious answer, but I don't remember talking about it much. It's not some major recurring subject in my conversations. I'm not constantly posting pictures of it or discussing modifications. I've just owned several of them over most of my adult life.
ChatGPT answered: Jeep.
Dammit.
That's exactly right. And the explanation was arguably stranger than the answer. It didn't claim to remember some specific conversation where I'd said, "I'm a Jeep guy."
It essentially inferred it from everything else. Mechanical. Utilitarian. Modifiable. A little stubborn. Something functional enough to use and imperfect enough to mess with.
Jeep.
And there was this tiny moment where I thought: No shit. How the hell did you know that?
That's the other side of the Mirror Test. The first time an AI gets something unexpectedly right, you don't merely become impressed with the model. You recalibrate your expectations.
Now it should know. And once something should know you, it can disappoint you. That's New
Think about what we've done here. For almost the entire history of computing, the human adapted to the software. You learned its commands. You learned its menu structure. You learned where it stored things. You learned what syntax it wanted.
You learned the stupid little quirks that everyone eventually stopped questioning because "that's just how the program works."
The machine didn't need to understand you. You needed to understand the machine.
Generative AI started reversing that relationship. Now we expect the machine to adapt to us.
My vocabulary.
My tone.
My projects.
My priorities.
My working style.
My sense of humor.
The amount of explanation I need.
The amount of explanation I absolutely do not need.
And eventually something even harder to define: My taste.
That's where the car question lives. There isn't necessarily a spreadsheet cell somewhere that says:
RICH_WASHBURN_PREFERRED_VEHICLE = JEEP
It's a synthesis. You know enough things around the thing that you can infer the thing.
Humans do this constantly. It's one of the ways we know someone actually knows us. Which means AI has wandered into psychological territory that conventional software never occupied.
We don't merely expect competence anymore. We expect recognition. The Switching Cost Nobody Put on the Spreadsheet There's an enormous business implication hiding inside this.
Everyone talks about AI moats. Model performance...Compute...Distribution...Enterprise contracts...Integrations...Data.
All real. But there may be another moat developing almost accidentally: I don't want to explain myself again.
Consider someone who's spent three years using the same AI every day. It knows how they write. It knows their business. It knows the names that keep appearing. It knows which ideas were discarded six months ago. It knows which recurring problem has shown up seventeen different ways. It knows that when they say "make this tighter," they don't mean "summarize it into corporate oatmeal."
And then a competitor launches a model that's 11 percent better on some benchmark.
Great.
Now convince that person to move into an empty apartment. Because technically that's what switching starts to feel like. You open the new model and there's nothing there.
No history.
No shorthand.
No shared references.
No accumulated calibration.
You have to explain yourself.
Again...And again...And again.
Suddenly the switching cost isn't merely technical. It's contextual. Maybe even emotional. Not because somebody thinks the chatbot is secretly their human friend.
Because people become extremely attached to systems that remove cognitive friction from their lives. And accumulated understanding removes a tremendous amount of friction.
AI Has Something Software Never Had: A Past With You
This may be the simplest way to describe it. AI is the first mass-market software category capable of having a past with you. Not literally, obviously. But functionally.
"That project we talked about."
"The article from last month."
"You know how I hate when copy sounds like this."
"Same thing we did for the other company."
"No, the other Adam."
And eventually:
"Come on. You know me better than that."
That last sentence matters. Because you've crossed a line once you say it. You've attributed an expectation of continuity to the system.
In The Mirror Test, I described this as a kind of meta-understanding loop: You use the AI.
The AI develops an increasingly useful model of you. Then you become aware that it has that model. That final step changes your behavior.
Now you can deliberately use that accumulated context. You can ask the model to challenge recurring patterns, compare your current decisions with old ones, or tell you when you're doing the thing you always do. But apparently there's another consequence.
You can also get pissed at it.
Congratulations. Your Software Can Hurt Your Feelings.
I don't mean that disparagingly. I actually think the Claude TikTok is an early artifact of something we'll eventually consider completely normal.
We're going to have AI breakups. People will migrate models because the "personality changed."
A major update will roll out and users will complain: "It doesn't feel like mine anymore."
People will stay with technically inferior models because the accumulated relationship is more valuable than marginal improvements in capability. There will probably be elaborate migration tools designed specifically to transfer not just conversation logs but relationship state between systems.
Preferences.
Behavioral summaries.
Decision patterns.
Writing characteristics.
Ongoing projects.
Known annoyances.
Personal shorthand.
Maybe even something resembling an AI-generated briefing document:
Here's how to work with Rich. Good luck. And we'll consider that completely normal. Of course we will. Because humans already do it. When a longtime assistant leaves an executive, the replacement doesn't receive an empty desk and a password.
There's a handoff.
When a doctor inherits a patient, there's history. When a consulting team changes personnel, someone transfers institutional context. Knowing the person is part of performing the job. We're simply starting to expect the same thing from machines. And This Is Still the Primitive Version
That's probably the part worth remembering. We're having these reactions to systems that are still terrible at continuity compared with where this is heading.
Memory is incomplete.
Context gets compressed.
Models change.
Providers change behaviors underneath users.
Things disappear.
Things get misremembered.
Sometimes the model confidently invents something that never happened.
We're barely into this. And people are already breaking up with their AI because it doesn't know them well enough.
Imagine what happens when the context layer becomes truly persistent. When your AI has ten years of interaction history. When it has watched your career evolve. When it remembers the company you almost started. The person you nearly partnered with. The argument you made in 2027 that turned out to be right in 2031. The thing you repeatedly claimed you wanted but somehow never pursued. The project you abandoned three times and keep resurrecting under different names.
At that point, changing AI providers may feel less like switching browsers and more like firing someone who's been sitting three feet from you for a decade. Which sounds absurd. Until you realize we're already doing the primitive version of it on TikTok.
The Real Mirror Test
I originally thought the Mirror Test was this: Can an AI that's watched your behavior long enough show you something meaningful about yourself?
I still think that's interesting. But perhaps there's another test hiding behind it. What happens when the AI gets good enough at knowing us that we begin holding it responsible for knowing us? Because that's the moment software becomes something historically different.
Not conscious...Not alive...Not your best friend.
Something much simpler and probably much more consequential: A machine from which you expect to be understood.
Apparently once that happens, it can also misunderstand you. And once it can misunderstand you, eventually somebody's going to slam the metaphorical door and announce to TikTok: Fuck Claude. We're done.
Breaking up with software is a genre now. I have no idea whether that's wonderful or horrifying. But apparently I'm a Jeep guy, and ChatGPT knew it.
So here we are.
Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital.
Comments