Five laws of generative AI

These laws have some exceptions, but I’ve found them broadly useful for navigating the current hype wave, and nearly always correct.

1. AI is a firehose of mediocrity.
2. AI’s advantages can’t be disentangled from its flaws.
3. It’s never “just” AI.
4. The term is more important than the definition.
5. The promise is more important than the reality.

^{For the purposes of this post, “AI” means generative AI.}

I’ve added a few brief thoughts on each law below.

1: AI is a firehose of mediocrity

• AI is a probability engine. At its best, it drags every operation toward the statistical mean. It can’t consistently be above average; improvements can only make it more average.
• No one is above average at everything. Thus, AI users experience a variation of Gell-Mann amnesia: they notice the technology’s flaws in their own area of expertise and ignore them everywhere else.
  • Programmers think AI can write well; writers think AI can create good art; artists think AI can code. Meanwhile, they nearly all recognize that it can’t do their own job very well.
  • AI looks like a cheap, fast way to automate the work of one’s colleagues without undermining one’s own value. But since it looks that way to everyone, it’s a Faustian bargain. In the end, the winners and losers are more likely to be chosen by organizational politics than any objective metric.
• AI presents unimpressive concepts with impressive confidence and certainty.
  • Without relevant expertise, the mismatch between quality and confidence makes it hard for users to detect errors.
  • Confidence is the currency of the executive suite. Leaders may have a particularly hard time judging the quality of AI-generated work.
• Expertise is detail-oriented. AI makes it possible to create without considering details. This is appealing to people who don’t often attend to details, and deeply problematic to those who do.
  • Anything that can be done easily, without knowledge of or attention to details, cannot be a competitive advantage.
  • Gains in speed due to AI use often represent a reduction in the organization’s competitive “moat.”

2: AI’s advantages can’t be disentangled from its flaws

• AI’s only killer feature is its ability to parse unconstrained inputs and produce unconstrained outputs. It achieves this using massive data sets, probability, and randomness; its behaviors can’t all be programmed in advance. There is necessarily a “black box” (unknown, unpredictable, unspecified behavior) at both ends.
  • The industry is in a constant battle to produce constraints that make AI more predictable and prevent it from causing damage. But a perfect constraint would negate the usefulness of the model, and all imperfect constraints can be bypassed.
• AI is presented as flexible, do-it-all software, but software is the strict codification of a process. AI can only estimate a process unreliably. Its flexibility and reliability are at odds; it’s not possible to increase one without decreasing the other.
• All AI output is hallucination. It’s considered a feature when it works and a bug when it doesn’t, but there’s no bright mathematical line between true outputs and false outputs. Therefore, hallucination can’t be “fixed.”
  • Several other AI flaws are also presented as “features” depending on context. Most of these are inherent to the technology itself.

3: It’s never “just” AI

• AI has three primary functions:
  • Productivity Theater: Output for the sake of output. AI is often used to produce artifacts that look impressive on a line chart but create little to no real-world value.
  • Investor Theater: Anything that’s done to distract or reassure investors without actually selling product. Making money is hard; generating AI buzz is much easier.
  • Accountability Sink: Delegation of decisions and processes to systems that can’t be responsible for errors. Letting negative feedback disappear into an AI system is cheap and doesn’t carry any risk of real change.
• AI, in and of itself, is an interesting technology. However, the vast majority of the time its technological uses are secondary to other goals.
  • Those who criticize AI are painted as resistant to progress and technology, but those who promote it seem to care very little about progress or technology; they care about the other things.
• AI is often used as cover for flagging sales, a slowing economy, lack of vision, and a need or desire for layoffs.
  • In Q1 2026 alone, over 37,000 layoffs were publicly attributed to AI.
  • Last week, after losing 98% of their value in four years, shoe company Allbirds announced they would sell AI products. This briefly boosted their stock price by over 600%.
    • This story isn’t unique. It’s played out again and again over the last few years.
• AI is starting to be seen as a symptom of downward spiral. Talking about AI may excite some investors, but others see it as a sign of financial weakness.
  • There’s a widespread, well-justified expectation among employees that if a company starts talking about AI now, they’ll start doing layoffs and other austerity measures in six months or so.

4: The term is more important than the definition

• By preference and unspoken agreement, AI’s biggest proponents refuse to clearly define what it is. To define AI would be to limit its usefulness as a marketing term.
  • Right now, AI usually means statistical text generation, but executive teams and advertisers are constrained to resist this definition.
• AI can be anything a salesperson wants it to be: an image generator, an image categorizer, a code generator, a recommendation algorithm, a chess engine, a faceless team of remote operators, a bizarre advertising campaign with no meaningful product underneath, or an across-the-board increase in employee workloads.
  • Microsoft has applied its “Copilot” AI branding to at least 80 different products.
• The industry’s incentives are not well-hidden. Saying “AI” is more important than using it, and using it is more important than creating value for customers.

5: The promise is more important than the reality

• Advertising strategies aside, AI companies are selling the promise of paying SaaS fees instead of payroll. As long as those fees are cheaper than the equivalent salaries, this creates economic gravity in favor of AI, regardless of the quality of its results.
  • The value of this promise is inestimable to the average company. Even if there’s a low probability of success, they feel forced to buy.
  • AI usage fees are currently subsidized by investors to an unprecedented extent. It cannot be assumed that the cost of a rented AI model will always be less than the cost of employing professionals.
  • AI companies aren’t just selling a product; they’re using their product to purchase a dependent workforce, attempting to monopolize the ability to do white-collar work so they can rent it back out.
• The promise of AI is the reason why the hype wave took off so quickly, going so far as to completely redefine “AI” versus the way it had been understood for decades. The industry raced to buy a very specific form of text generation, and has now poured too much money into it to reconsider.
• AI is marketed as an order-of-magnitude improvement in productivity, but in reality, such improvements do not and cannot exist. Real productivity gains have been extremely modest, and are at least partly attributable to reductions in quality.
  • The extent to which AI can automate a task is directly correlated to the extent to which quality of execution can be degraded without short-term consequences.