“Only 16 percent of Americans think AI will have a positive impact on society.” It’s a brutal number, but it makes perfect sense. We spend all our time in the bubble—reading X, tracking GitHub stars, arguing about context windows, and debating the merits of different quantization methods—and we forget that the average person isn’t looking for a “copilot.” They’re looking for a reason why their job isn’t about to be deleted by a script written by a 22-year-old in San Francisco who thinks “efficiency” is a substitute for a living wage.
The disconnect is a feature, not a bug. While the C-suite is salivating over efficiency gains and the potential for leaner payrolls, the people actually doing the work are staring at a tool that occasionally tells them to put glue on pizza. According to this report via TechCrunch, the gap between investor optimism and public dread is a canyon. Why would the general public be optimistic? Most “AI features” added to existing software over the last year have been nothing more than a thin wrapper around a prompt that makes the UI slower and the output more generic.
This is the classic “Demo Effect” scaled to an entire industry. We’ve seen this before with the early days of the metaverse—lots of expensive headsets and glossy corporate presentations, but very few reasons for a normal person to actually wear them. The industry has focused entirely on the “magic” of the model (and probably fewer if you ask people who actually have to read the output) rather than the boring work of reliability. If you can’t trust a tool to get a basic fact right 100% of the time, you don’t see it as a benefit; you see it as a liability that requires a full-time human babysitter to prevent a corporate catastrophe.
The problem is that the labs are chasing AGI while the users are just trying to figure out why their email client is now trying to write their thoughts for them in a tone that sounds like a corporate HR manual from 1994. It’s like being sold a Formula 1 car for a commute to the grocery store. Sure, it’s fast, but it’s loud, it’s terrifying, and you can’t actually fit a gallon of milk in the back. Do we really expect people to cheer for a technology that is essentially a black box designed to optimize them out of a paycheck?
Then there is the physical reality of the thing. The energy requirements are astronomical, the H100s are priced like rare medieval artifacts, and the latency is still just annoying enough to break the flow of a real conversation. We talk about “intelligence” in the abstract, but the user experiences the friction of a spinning loading icon and a monthly subscription fee for a feature that used to be free. The cost of inference is effectively a tax on the user’s patience. Or maybe the latency is just a side effect of the current architecture—I’m not entirely sure if we’ve hit a ceiling here or if the labs are just being lazy with optimization.
This sentiment won’t flip just because a new model drops with a slightly better MMLU score or a larger context window. The public doesn’t care about benchmarks. They care about whether the tool actually solves a problem without creating three new ones. Unless there is a fundamental shift toward verifiable utility over probabilistic guessing, that 16% is going to keep sliding. By Q4 2026, we’ll see this number dip into the single digits as the novelty wears off and the actual costs—economic, social, and environmental—become the primary talking point.
The hype is finally hitting a wall.
