In May, 404 Media wrote that Character.ai’s changes are “what AI enshittification looks like,” with tons of ads, usage limits, guardrails, and “lobotomized” models. In the bankruptcy filing, Adams essentially admits that this enshittification was the strategy: “Character has, however, taken significant steps to reduce its operating burn, including transitioning inference workloads to lower-cost GPU capacity and post-training open-weight models rather than developing proprietary foundation models. At the same time, Character reduced its employee headcount,” he wrote. “In mid-2025, Character launched in-app advertising […] in early 2026, Character also introduced in-app purchases.”
While the lawsuits against Character are significant, they are also not unusual for an AI industry that has been plagued with legal problems relating to the theft of intellectual property and suggesting self harm or violence to their users. In other words, there is nothing in Character’s bankruptcy filings that could not be said about many other AI companies that are spending a huge amount of money on compute without having a revenue model that can support it. The difference between Character AI and behemoths like OpenAI, Anthropic, Meta, and Google is that those other companies have been able to repeatedly get injections of new cash, or, in the case of Meta and Google, have been able to subsidize their AI products with revenue from their massive advertising businesses.
The company said in a blog post that it will continue operating through the bankruptcy process: “The Character.ai platform remains up and running, and you can continue to access all services at this time. We’re focused on finding the best path forward for our platform and community.”
We haven’t even worked the externalities into the price yet, we don’t even know what they all are.
There probably wasn’t a billion dollar idea in there. But there might have been a $100,000 idea in there.
I’m sure there are plenty of people who’d pay $10 once, and a smaller proportion who would do so monthly, for a desktop friend branded as their favourite media character. I’m sure we’ve been doing standalone toys like that since the 90s if not before.
But they went and nickel-plated the implementation. Was it over ambition, misjudged economics, or a belief the product had to fit into the AI blitz?
The smart plan would have been to bring up a small number of curated licensed personalities. You know Miku would sell, for example.
If you have a target character, you no longer need a LLM. Use a canned base of high quality data or curated sources, and you can drive it with Eliza-style keyword mapping and some undergraduate-project Markov Chain stuff to spin the dialogue a bit.
But they promised arbitrary characters and open-ended conversations, which all but forced dependency on a huge data model of some sort. That’s a huge expense to service the long-tail customer who wants to talk to the ghost of Robert Mugabe or the protagonist from a 25-year-old C-tier manhwa, or so you can pivot from asking Batman about the Riddler to his suggestions for replacing the muffler of a Ford Aerostar. Also, you now have a content domain far too broad to fully test and rein in, ending in tragedy for some vulnerable users.
If everything worked just right, it might outperform the local “themed Bonzi Buddy” concept on some sort of Fanboy Turing Test, but was that a battle you even need to win to be a successful product?
When LLMs first became public, I went to character.ai and opened up the Socrates AI. I’m a trekkie, and I really love how Starfleet officers are always using the holodeck to hang out with Leonardo Davinci, I wanted to do that with My favourite philosopher.
Unfortunately, LLM Socrates was really stupid and would agree with anything I said. So then I went to the Karl Marx LLM and convinced him that communism is evil for shit and gigs. And that was the last time I ever saw promise in LLM technology.
Lonely people having this as their only friend surely will be worth a whole lot of money. Imagine all the ways to manipulate such a person or just passively syphoning off data. $$$
The problem with that theory being that you’re talking about technologies and principles that have been around for more than half a century, and in all that time we’ve never yet figured out how to use them to convincingly emulate conversation the way LLMs can. If what you are describing was possible, we’d have been doing it twenty years ago. People were trying. For Character.ai to have come up with that kind of solution, they’d have to have solved the exact problem that LLMs solved, only with some kind of magical technology with none of the drawbacks of LLMs.
That’s sort of like saying that all Theranos had to do was invent a product that actually did all the magical things they claimed they could do. Sure, yeah, I guess if they’d done that everything would have been fine.
Good. Investors need to feel fear
It has arrived!
Well, they’ll sell all the user data.
