Everyone Complains About AI, Yet Nobody Can Quit It: Unpacking the Paradox

A "Self-Loathing" AI Company

This past summer, a CEO of an AI startup said something that stuck with me. He runs a company called Springboards, which is building a large language model unlike most others — one specifically designed to generate more diverse responses than mainstream competitors. Right at the start of the call, he offered a telling description: "We often call ourselves a self-loathing AI company."

The phrase is worth sitting with, because it captures something real about the complicated, contradictory relationship people have with artificial intelligence right now. On one hand, skepticism, criticism, and outright hostility toward AI have never been louder. On the other, usage of these tools keeps climbing. People complain constantly — and then keep using them anyway.

The Paradox of Loathing and Dependency

Public frustration with AI is well-documented. It spans homogeneous content and text with an unmistakable "AI smell," anxiety about creative work being displaced, and concerns over training data sourcing, energy consumption, and factual reliability. Social media is saturated with mockery of AI-generated content — "written by ChatGPT" has become something close to an insult.

And yet, real-world adoption keeps accelerating. Whether it's drafting emails, summarizing documents, researching topics, or assisting with code, more and more people have quietly woven AI into their daily workflows. This "complain publicly, use privately" dynamic is one of the defining psychological features of the AI era.

Why Does This Split Exist?

A significant part of the answer lies in the homogeneity of mainstream model outputs. When large numbers of users rely on the same handful of models, the results tend to converge around similar tones, structures, and word choices — producing a recognizable "AI voice" over time. This uniformity breeds aesthetic fatigue and reinforces the stereotype that AI is superficially smart but fundamentally lacks personality or genuine diversity.

From a technical standpoint, this convergence stems directly from how leading models are trained. Most mainstream large language models are fine-tuned using Reinforcement Learning from Human Feedback (RLHF) — a process where human annotators rank multiple model outputs, and those signals are used to adjust model behavior. While RLHF improves safety and usability, it also tends to reward responses that appeal to the broadest set of raters: safe, middle-of-the-road answers that inadvertently suppress stylistic and perspective diversity. Add to this the fact that top models are largely trained on similar web-sourced corpora, and it's no surprise that switching between products often feels like changing the interface but not the underlying voice. This is the technical root of why the "AI-sounding" label has spread so widely.

Diversity Might Be the Key

Springboards is targeting exactly this pain point. Rather than optimizing for a "correct answer," the company is trying to build an LLM capable of producing a wider, more varied range of responses. The underlying logic: if AI outputs stop feeling interchangeable, some of the public's hostility toward them might ease.

From a product design perspective, this is a direction worth taking seriously. Current mainstream LLMs are optimized heavily toward "safe" and "reliable" outputs — a tendency that quietly compresses response diversity. Models that deliberately pursue differentiated output may have distinctive value in contexts like creative writing and brainstorming.

The Other Meaning of "Self-Loathing"

The CEO's framing of a "self-loathing AI company" also reflects a strain of genuine self-awareness within the industry. Not everyone building AI is deaf to the criticism — some are actually trying to extract product direction from public dissatisfaction. The posture of building AI with a critical eye is itself a corrective to unchecked optimism.

In large language model development, "output diversity" is typically controlled through sampling parameters, most commonly "temperature" — the higher the value, the more likely the model is to draw from lower-probability tokens, producing more varied and creative outputs; lower values yield more stable, conservative responses. But simply cranking up temperature brings real risks: logical incoherence and factual errors. What companies like Springboards are pursuing — "structural diversity" — goes further: the goal is to bake diverse perspectives and response styles into the training process itself, rather than relying on inference-time parameter tuning. This is a more fundamental but significantly more challenging technical path, requiring a rethinking of training data composition and potentially introducing multi-objective optimization mechanisms to balance quality, safety, and diversity simultaneously.

The Deeper Question Behind the Contradiction

Why do people simultaneously resent AI and depend on it? The answer likely comes down to this: AI has become a genuinely useful productivity tool whose convenience is hard to give up, even as it fails to earn emotional trust or value alignment. The tension between pragmatic utility and emotional resistance is the source of the contradiction.

This psychological split won't resolve itself through technical progress alone. If anything, as AI capabilities grow and adoption deepens, the public's ambivalence may become more complex, not less. For developers, the real challenge isn't just building more powerful AI — it's building AI that is more trustworthy, more varied, and capable of earning genuine buy-in rather than reluctant reliance.

Springboards' approach suggests a product philosophy worth borrowing: rather than ignoring public negativity, face it head-on and turn it into a source of differentiation. When AI stops being a one-note "answer machine," people's relationship with it might finally shift from passive dependency to genuine acceptance.