The AI bubble isn’t like the internet bubble (permalink)
One of the surprise breakout software products of the early web was Lotus Notes, a kind of primitive precursor to all-in-one office productivity suites like GDocs, Office365, etc. It was so important that its creator, Ray Ozzie, was promoted to Microsoft’s Chief Software Architect, succeeding Bill Gates himself:
▻https://knowledge.wharton.upenn.edu/podcast/knowledge-at-wharton-podcast/the-man-who-would-change-microsoft-ray-ozzies-vision-for-co
People who remember Notes tend to deride it for its clunky user interface and demi-functional administrative tools. But what made Notes so central to Microsoft wasn’t its polish – it was the fact that Notes represented a brokered peace between IT managers, who wanted mainframe-like control over everything their users could do with business equipment, and the users themselves – workers who kept smuggling internet-based tools into the enterprise network on the very sensible grounds that they had a job to do, and these were the best tools to do it.
The arrival of internet-based tools – especially ones that ran in browsers – represented a major challenge to IT departments, who had been long accustomed to dictating terms to their users. If the IT manager and the compliance department decided that the best way to manage disclosure and leak risks was to block all email attachments for outside users, then that was that: no one could send those attachments.
But after the internet arrived on the corporate desktop, employees who needed to get documents to supply chain partners and customers could treat these IT policies as damage and route around them. Just fire up your Hotmail or Yahoo mail window, or hop on MSN Messenger or ICQ or AIM, or drop the file on an anonymous FTP server and send the link to your counterparty. Job done!
IT managers hated this, and to be fair to them, they weren’t (always) wrong. These outside tools came from a variety of untrustworthy sources, including malicious sites that pushed virus-infected versions to their users. Also, by evading firewall rules with these tools, users made it impossible to achieve the compliance goals that IT had been charged with enforcing, and it was IT’s asses on the line if the company got in trouble as a result.
Foundationally, IT was being asked to do two irreconcilable things: they were supposed to be enabling workers to get their jobs done, and they were supposed to be stopping those workers from doing things that could harm the business. This can’t be done, because the only way to eliminate the possibility that a worker will take an action that harms the business is to gag that worker and lock them in a dungeon. Workers need flexibility and freedom to achieve business goals, and that flexibility and freedom means that those workers might (deliberately or accidentally) thwart the business’s goals.
What’s more, workers will always run into situations that were not anticipated by policy, and if they are denied any agency or initiative, they will fail to get their jobs done. In work, the exception is the rule, hence the importance of “process knowledge” (all the implicit knowledge shared among workers across the firm and its suppliers and customers, which cannot be captured or recorded):
►https://pluralistic.net/2025/09/08/process-knowledge/#dance-monkey-dance
Indeed, there’s a form of labor action called a “work to rule,” in which workers only do the things dictated by their rulebooks, without taking any of the routine additional measures dictated by process knowledge. Merely by following every rule to the letter, workers can grind a shop to a halt:
►https://en.wikipedia.org/wiki/Work-to-rule
Since the dawn of personal computers, workers and IT departments have come into conflict, as workers literally smuggled technology into the business that could do things the IT department had (often arbitrarily and capriciously) prohibited. When Visicalc emerged as the killer app for the Apple ][+, workers snuck these computers into work and used them to sort spreadsheets in ways that IT had declined to permit. They didn’t do this to cheat or steal from the company – the whole point was to do a better job.
So it was with the early web: workers discovered a myriad of new capabilities in the free-to-use world of web-based tools and realized how these tools would make them much more effective at their jobs. The fact that IT wouldn’t let them do these things was just more evidence that IT – and the managers who set IT’s agenda – didn’t understand the business as well as workers.
It didn’t help that IT managers’ first line of defense was the high-tech version of abstinence-only education: “You only think you need your work computers to do this, but really, you don’t, so stop trying”:
►https://www.theguardian.com/technology/2009/jun/16/computer-security-abstinence
Abstinence-only education never works, but where “you only think you need this” failed, Lotus Notes succeeded. Lotus Notes provided a whole suite of tools that largely (if imperfectly) replaced the universe of free tools that workers were using to evade their IT departments’ edicts, so they could get their jobs done. At the same time, Lotus Notes provided a set of management tools that let IT fine-tune how these tools worked, giving them (some) of the controls they needed to achieve their compliance goals.
Like all brokered peace settlements, Lotus Notes left both sides feeling like they’d made a compromise they could live with, giving up some of their goals, but keeping the things that really mattered to them.
It’s impossible to overstate how important Lotus Notes and similar products were, because workers demanded the right to use the web on their work computers, and they made those demands so forcefully that managers had to completely re-do their IT policies, lest those workers treat them as damage and route around them. Back then, the tech press was full of stories about these conflicts, as workers insisted that the new technology that was sweeping the nation was so foundational and transformative that they had to be allowed to use it.
What we never saw back then were stories about how managers had to monitor workers to ensure that they were using the web as much as possible. No one had to force workers to find ways to integrate the web into their workflows.
In other words, the story of the web at work was the opposite of the story of AI at work. Today, you can’t turn around without reading a story about bosses who are threatening to fire workers if they don’t increase their AI usage:
►https://www.businessinsider.com/boss-track-ai-use-career-2025-8
Virtually every major company now has a program to force workers into using AI:
►https://www.cnbc.com/2026/05/05/ai-use-work-employee-monitoring-tech-surveillance.html
It’s conceivable that over the past quarter-century, bosses have become technophiles while workers have fallen prey to superstitious technophobia, but it hardly seems likely. Historically, workers have always been enthusiastic about tools that let them do a better job – indeed, it’s a truism that labor-led automation produces improvements in quality, while capital-driven automation increases throughput (often at the expense of quality).
Workers aren’t the only typical early adopters who find AI lacking. As a group, teenagers and young adults hate AI:
►https://www.nytimes.com/2026/04/09/style/gen-z-ai-gallup-study.html
That’s not what it was like during the early web days. Back then, young people entering the workforce were passionate devotees of the web, to the point where the business press routinely ran articles asking how today’s workplaces were going to adapt to the demands of these webbed-up workers.
►https://www.nber.org/digest/apr03/internet-changes-labor-market
AI boosters insist that the deficits we see in AI – its lack of profitability, its primitive and error-riddled outputs – are no different from the shakedown problems of the early web (and we know how the web turned out!). But this is a profoundly flawed comparison: the early web and AI are very different from one another.
For one thing, the early web may have lost money, but it had great unit economics. Every new web user brought the web closer to profitability, as did every new use of the web, and every new generation of web technology. By contrast, AI has – in the memorable phrasing of Ed Zitron – “dogshit unit economics.” Every new AI user makes AI less profitable, as does every new use for AI, and each generation of AI loses more money than the last. AI is the money-losingest endeavor in human history:
►https://pluralistic.net/2025/09/27/econopocalypse/#subprime-intelligence
In other words, the early web was a technology that grew more profitable every day, which workers and young people had to force on their bosses – and AI is a technology that grows less profitable every day, and bosses have to force it on workers and young people.
Now, it’s true that some workers don’t have to be forced to use AI. Workers who enjoy a high degree of autonomy (that is to say, workers who are positioned to ignore workplace coercion) can adopt AI in ways that they feel suited to, just as those early web users and Visicalc smugglers did. They can fulfill the maxim that labor-driven automation improves quality, while resisting capital’s insistence that automation be used to increase throughput at quality’s expense.
They can act as centaurs (workers assisted by technology), not as reverse-centaurs (workers who are recruited to serve as peripherals for machines). As with all technology questions, what the technology does is nowhere near as important as who the tech does it for and who the tech does it to:
►https://pluralistic.net/2025/09/11/vulgar-thatcherism/#there-is-an-alternative
And there’s another group of workers who adopt AI voluntarily: workers who see that AI can do a lot of work that they view as dull and unimportant for them. These workers might be right – there are plenty of bullshit jobs out there:
▻https://memex.craphound.com/2018/06/20/david-graebers-bullshit-jobs-why-does-the-economy-sustain-jobs-that
But it’s also possible that they’re wrong, and they’re substituting AI for something that really should be done by a person.
But on the plus side, at least no one has to force them to adopt AI.
