Why we keep using an AI we claim to hate so much

"We often say that we're a self-loathing AI..." That, according to the author, is how a call this summer began with the CEO of Springboards, a startup building an LLM designed to come up with a wider variety of responses than its mainstream rivals. The quote opens the article that MIT Technology Review published this Monday, October 5, and its headline asks the question bluntly: if people hate AI so much, why can't they get enough of it?

A caveat before going further. At the time of writing we have only been able to read the headline and the opening of the text, where the quote appears cut off, so we are not going to attribute figures or conclusions to it that we have not read. What follows is what that opening makes clear, plus our reading of why the question matters to anyone building on Claude.

What the article sets out

The starting point is an anecdote that carries more weight than it seems to. When the head of an AI company describes their own project in terms of self-loathing, and does so at the very start of an interview, it shows how far rejection has become part of the commercial landscape. Companies in the sector do not discover it in a survey. They take it for granted and build it into how they present themselves.

The second data point is in the product. Springboards is not competing by promising a more powerful model, but one that responds in more diverse ways than the mainstream models do. That choice points to one of the underlying complaints: people are bothered less by AI getting things wrong than by the fact that everything it produces looks alike.

The third is in the headline itself, which does not set two different groups against each other, those who hate and those who use. It is talking about the same people.

Why both things can be true at once

This part is our interpretation. Our hypothesis is that the rejection and the usage are aimed at different objects. The rejection is directed at AI as a phenomenon: the filler text flooding search engines and social networks, or the features that show up in an app without anyone asking for them. The usage, on the other hand, is almost never experienced as "using AI". It is experienced as finishing an email or finding a bug in the code before heading home.

There is also a technical reason behind the sense of sameness. An LLM generates text by choosing among the most probable continuations, and the most probable tends to coincide with the most seen. Without context of their own, two people asking for the same thing get very similar results. Anyone working in advertising or design notices it quickly, and it is understandable to be irritated by a tool that is useful and at the same time pushes your work toward the average.

Who this reading is useful for

For anyone putting an assistant in front of end users, and especially for product teams and agencies. We take two practical consequences from MIT Technology Review's question.

1. The label carries weight. Presenting a feature as "AI-powered" no longer adds value by default, and with some audiences it subtracts. It usually works better to describe the task that gets solved and let users decide whether they want to know what is underneath, without hiding it when they ask.
2. Variety is built with context. In the Claude ecosystem, skills exist precisely to package instructions and in-house material that the model loads on demand, and MCP servers give it access to data that lives nowhere else. An assistant that knows an organization's archive and tone stops sounding like all the others.

A third issue remains, and it is a less comfortable one: trust. Part of the unease comes from not being able to check what the system claims. Linking to sources and leaving a human review step before publishing are cheap design decisions that change the user's relationship with the output.

We find a sector that admits the rejection, as Springboards does with its line, healthier than one that treats it as a communication problem. That people keep using what they say they hate suggests usefulness is running ahead of trust, and that gap is better closed with a better product than with better slogans.