Springboards CEO Says Users Hate AI but Keep Using It, as Startups Chase Better LLM Outputs

The chief executive of Springboards, a startup building its own large language model, told MIT Technology Review this summer that his company often describes itself as a "self-loathing AI" — a line that neatly captures the split between how people talk about artificial intelligence and how they actually use it. The remark came at the start of a call about the startup's core product, an LLM trained to produce a wider spread of answers than the models most users already know.

That internal contradiction is not just a rhetorical flourish. Consumer AI tools have accumulated hundreds of millions of weekly users even as surveys and social feeds fill with complaints about slop, hallucinations, job displacement and privacy. The gap between stated hostility and daily habit is now one of the more commercially useful facts about the market, because it means adoption is not tracking sentiment.

Springboards is betting on a specific technical complaint: that mainstream models converge on the same few responses. Its pitch is variety — an LLM that returns more diverse outputs rather than the modal answer, which is what most assistants have been tuned to give. That is a harder engineering target than it sounds, since the training and evaluation pipelines used across the industry tend to reward consistency, safety and a narrow band of acceptable replies.

The company is small and does not carry the compute or distribution of the frontier labs, so it is competing on output behaviour rather than raw benchmark scores. Whether users notice or care about response diversity is an open question; most never see the alternative. For now, Springboards' own framing suggests its leadership reads the resentment as ambient rather than decisive.

What the anecdote really exposes is a product problem the industry has been slow to name. If dislike of AI were strong enough to change behaviour, usage curves would have bent already. Instead, the habit persists, which gives builders room to keep shipping — and gives critics a reason to argue that convenience, not consent, is doing the work.

Watch for whether smaller labs like Springboards can turn output variety into a measurable retention advantage, and whether the frontier models copy the approach once they do. The larger test is whether any of this shifts the underlying sentiment, or simply gives unhappy users a better-feeling version of the same routine.