The Last Mile Nobody's Walking
Aaron Levie, the founder of Box, dropped a social media post this week that's been impossible to ignore: tech CEOs are "uniquely prone to AI psychosis." He wasn't saying they're crazy in the clinical sense. He was saying something more specific, and honestly, more damning.
The claim is that CEOs sit far enough away from the actual work — the last mile of production where AI has to generate real value — that they develop a kind of delusion about what these tools can do. They look at slides. They hear efficiency pitches from VCs who love the dream of a tiny team doing massive work. And they buy it.
Levie isn't disavowing AI tools. He's not telling anyone to throw their ChatGPT subscription in the trash. What he's insisting on is something embarrassingly simple: if you're going to bet your company on a technology, use it. Actually use it. Feel where it breaks. Understand what it gets wrong.
You can't just look at a deck and declare "incredible efficiency, let's go." Not when the people actually doing the work are watching you do it from a glass tower.
It's a relatively gentle note of skepticism compared to what's coming. But it landed because it names something everyone in tech already suspects and nobody wants to say out loud at a board meeting.
Google's Search Identity Crisis
If Levie diagnosed the disease, the TechCrunch Equity podcast episode with Anthony Ha, Kirsten Korosec, and Sean O'Kane mapped out the symptoms across the industry.
Google is the poster case. The company keeps pushing AI deeper into search — more generated answers, more commercial transactions baked into results, less of the clean information retrieval experience people have trusted for decades. And then they walk it back. Add nuance. Tell you the "10 blue links" experience is still available if you want it.
Kirsten Korosec put it perfectly: Google is "chasing that thing it feels like it has to do to keep up, but it's messing with the thing that people attach to the brand the most."
The company struggles with its own identity. Is it an information retrieval system? A commercial transaction engine? Both? It keeps trying to be both and ends up doing neither well.
And then there's the spelling thing. You know what I'm talking about. Google can't spell its own name. Ask it how many P's are in "Google" and it tells you two. Two. The company that built its entire empire on being the world's most reliable answer machine can't count letters in its own URL.
Sean O'Kane noted that Google goes onstage at IO and talks about shopping, about booking flights, about commercial transactions — stuff that feels like it's moving away from what people actually think of when they think "Google." It's a tension between the information retrieval side that built the brand and the commercial ambitions that drive the strategy. And every time they lean into one, they alienate the other.
The Consumer Revolt Nobody Saw Coming
Here's where it gets interesting. This isn't just an internal tech debate about leadership and tools. There's a real consumer backlash forming, and it's showing up in the data.
DuckDuckGo reported installs are up 30%. That's a huge leap. Yes, DuckDuckGo is a much smaller product than Google — I'm not saying Google is in immediate trouble. But 30% growth on a smaller platform signals something significant: there's a very real audience that does not like the current AI direction, and they're voting with their fingers.
What's striking is how DuckDuckGo has positioned itself. A year ago, even alternative search engines were experimenting with AI features, emphasizing them because they felt they had to keep up. Now? They're saying the opposite. DuckDuckGo is promoting itself as "anti-AI." Other alternatives are positioning around the idea that they simply weren't interested in that stuff at all, or that when they do use AI, it's locked in a separate sandbox that doesn't touch your core experience.
That shift is remarkable. It suggests the market has moved fast enough that "no AI" is now a selling point rather than a liability.
And it's not just search. Graduating college students are booing any mention of AI. The bad vibes around tech industry layoffs keep getting worse. Anthony Ha captured the paradox perfectly: AI is "incredibly polarizing" — simultaneously everyone's using it and loving it, but also no one's using it and hating it at the same time. Large contingents for whom both of those things are true.
You can feel the cultural moment shifting. The people who were going to adopt AI already have. The people who aren't? They're digging in.
The Workforce Reality Check
Let's talk about what this actually means for the people doing the work.
AI-driven layoffs are directly affecting workers. Not abstractly. Not in some future-tense scenario. Right now, people are losing jobs tied to producing code, writing content, handling customer support — the kinds of work that AI tools were supposed to augment, not replace.
But it's not just about job loss. It's about how work is changing. Sean O'Kane pointed out that on the software side, where people's jobs are directly tied to producing code, things are really changing. Meanwhile, in physical infrastructure, manufacturing, and robotics — areas like Mind Robotics, the spinout from Rivian CEO RJ Scaringe — adoption is slower. The software side is where the disruption is concentrated.
Here's what Levie's argument really exposes: the gap between how executives experience AI and how workers experience it.
Executives see productivity gains. VCs love the dream of a tiny team achieving large-team effectiveness. It's a compelling narrative, and it might even be true in some cases. But if you're not touching the end work — if you're making decisions based on slides and projections rather than hands-on experience with the tools — how do you know what's actually working?
Traditional workforce transformations were bottom-up. People liked the tools, they brought them in, and executives eventually accepted what was already happening. AI adoption feels different. It's top-down. Driven by executive belief and VC conviction rather than organic worker adoption.
That matters. Because when transformation comes from the top without genuine understanding of the last mile, it tends to break things. Not just jobs — trust.
Building for the Skeptics
Kirsten Korosec raised a question that's worth sitting with: is this anti-AI moment an opportunity for startups or other areas of business?
The challenge is obvious. If you build something tailored for a group that's skeptical of AI, you're probably going to alienate users who are gung-ho about it. You can't please everyone. But there's a lane to be "anti-AI" or at least to put AI in a separate sandbox that doesn't affect your core experience.
DuckDuckGo is already proving this works. The alternative search engines that a year ago were scrambling to add AI features are now leaning into the opposite positioning. They're saying: we built this for you, without the stuff that's making everyone uncomfortable.
It's a legitimate business strategy. Not every company needs to be an AI company. Some companies can just be good companies that happen to use technology thoughtfully rather than aggressively.
The risk, of course, is that the anti-AI lane might be smaller than the pro-AI lane. But size isn't everything. Profitability matters more. And if there's a meaningful audience willing to pay for products that don't force AI down their throats, that's a business worth building.
The moment we're living in is one where both sides coexist. Evangelists and skeptics. People who love AI and people who hate it. The companies that figure out how to serve both without trying to be everything to everyone might end up winning.
What CEOs Actually Need to Do
Levie's prescription isn't radical. It's just something that keeps getting ignored.
Use the tools. Understand what they're doing. Don't just look at a slide and declare incredible efficiency.
It sounds obvious because it is. But the distance between knowing this and doing it is where the psychosis lives. CEOs who haven't actually used their company's AI tools — who make decisions about deployment, layoffs, and strategy based on secondhand reports and vendor demos — are operating blind.
The Google example is instructive. A company that should be the world's expert on information retrieval can't even spell its own name in AI-generated responses. That's not a failure of the technology alone. It's a failure of leadership to understand what they're building.
The workforce angle makes it worse. When executives push AI adoption from the top without genuine understanding of how it affects the people doing the work, they create resentment. They create backlash. They create the exact consumer revolt we're seeing with DuckDuckGo and college students booing AI mentions.
The solution isn't to stop using AI. It's to use it honestly. To understand where it works and where it doesn't. To make decisions based on actual experience rather than slides.
Levie called it "AI psychosis." Maybe that's too strong. But the underlying point stands: when leaders lose touch with the last mile of work, they start believing things that aren't true. And those beliefs have real consequences — for workers, for consumers, and for the companies they're supposed to be leading.
The Numbers Behind the Narrative
The scale of AI-driven workforce reduction in 2026 is staggering. According to Layoffs.fyi data cited by TechCrunch, the first five months of 2026 saw 115,430 people fired from 152 tech companies — nearly matching the entire year of 2025, which saw 124,636 layoffs across 275 companies. And the bulk of these cuts cite AI as the reason.
ClickUp CEO Zeb Evans took this logic to its extreme. After deploying roughly 3,000 internal AI agents, Evans laid off 22% of ClickUp's workforce — not as a cost-cutting measure, but as a deliberate move toward what he calls a "100x org." Surviving employees were offered $1M salary bands, with the promise that those who create outsized impact using AI would be paid outside traditional compensation structures. The vision: a workforce of people who direct AI agents and review their output, rather than doing the work themselves.
But here's what makes Levie's "AI psychosis" diagnosis so apt: a Gartner survey found that while ~80% of companies using autonomous tech have cut jobs, those reductions aren't translating into meaningful financial returns. The layoffs are real. The productivity gains? Not so much.
The NBER's March 2026 working paper surveyed nearly 750 corporate executives and documented what researchers call a "productivity paradox" — perceived productivity gains are larger than measured productivity gains, likely reflecting a delay in revenue realizations. Labor productivity gains are positive and vary by sector, with the largest effects in high-skill services and finance. But these gains aren't primarily driven by capital deepening — they reflect increases in revenue-based total factor productivity, closely associated with innovation-and demand-oriented channels. In other words: companies believe they're getting more out, but the actual measured output doesn't quite match the perception.
MIT researchers who created thousands of agents for task completion concluded that agents "just aren't doing human-quality work yet in many cases." Their prediction: at the current rate of LLM improvement, models will "be able to complete most text-related tasks with success rates of, on average, 80%–95% by 2029 at a minimally sufficient quality level." Agents will need another few years to outperform humans. We are not there yet.
What the Research Actually Shows
The academic evidence paints a picture that directly contradicts the CEO narrative.
A October 2025 meta-analysis published in UC Berkeley's California Management Review by Dritjon Gruda and Brad Aeon examined 371 estimates published between 2019 and 2024. The finding: "no robust, publication-bias-free relationship between AI adoption and aggregate labor-market outcomes" once methodological heterogeneity is controlled. The results vary dramatically depending on how studies define "AI," which sectors they sample, and whether they adjust for capital deepening effects.
A comprehensive Nature Human Behaviour meta-analysis of 106 experiments found that, on average, human-AI combinations perform worse than the better of the two working solo. Performance improvements emerge only in specific contexts, particularly open-ended content-creation tasks. Decision-making and judgment tasks suffer from over-reliance on AI suggestions or confusion over authority and responsibility.
Harvard Business Review published research in February 2026 by Aruna Ranganathan and Xingqi Maggie Ye that found AI tools don't reduce work — they intensify it. An eight-month study showed AI tools made productivity surge, but also brought cognitive fatigue, unsustainable hours, and other problems. When everyone is using AI to produce more stuff, the bottleneck simply shifts to executives who must authorize all the output. "If everyone is empowered to act," they write, referencing OpenAI's experience, "things may get out of control."
The pattern is clear: CEOs who are distant from the last mile of work see prototype demos and believe they're seeing productivity. The data says something different. The perceived gains are real to those experiencing them — but they don't match measured outcomes. And the workforce reductions happening in the name of AI productivity aren't delivering the financial returns executives promised.