It is my view that, as things stand right now, AI is generally reducing net productivity.
AI should have created a huge economic boom. Instead, the economy is mostly stagnant; outside of South America, every major economy has grown at the same-or-slower rate in 2023 and onward than they did in 2015-2020. The software industry is probably the best case study, since the whole industry has been in full “AI acceleration” mode for almost 2 years now; AI adoption is 100% and has been for at least a year. Major companies publicly threatened to fire employees who didn’t adopt AI (Coinbase, Shopify)– in mid 2025, almost a year and a half ago. Companies such as Microsoft, Salesforce and Amazon had internal “tokenmaxxing” leaderboards which were linked to performance reviews. Though harder to measure objectively, smaller software companies were often even more fanatical and aggressive in their adoption of AI. It’s no longer possible to argue that lack of adoption is preventing us from seeing the positive effects of AI in software. The trends we’ve seen in software over the last year do in fact reflect the impact of AI.
The impact is not positive. The new software that’s being produced with AI is very bad, across the board; little of it is interesting or even usable beyond a demo. Major existing pieces of software that people use all the time (e.g., Microsoft Windows, MacOS, Github, Twitter, Youtube, Google Search, etc.) are not getting any better, and in fact many are getting rapidly worse. There’s no evidence that companies who adopt AI make more money than they would have without it (note: this doesn’t include companies selling AI, who are definitely making a killing).
This is perhaps counterintuitive, since every week there’s a new incredible headline about new AI-powered ventures, such as decompiling old binaries, solving famous math problems, and so on. The capability of AI in localized areas is tremendous and continues to grow at an amazing rate.
Mass introduction of AI into processes and systems it’s unsuited for is cancelling out the benefit of using it where it’s effective. It is reducing the quality and availability of goods and services. It is reducing the efficiency of processes instead of increasing it, especially in anything that involves or relies on communication (which is a lot of things). It’s frustrating that despite the enormous potential of AI, it’s mostly being used in counterproductive ways, and very few people are actually benefitting in any way. We could be doing so much better!
This does not mean…
Saying that AI is destroying as much value as it’s creating causes a lot of cognitive dissonance, especially among people who are psychologically all-in on AI. The claim gets misinterpreted (sometimes on purpose) in many ways:
- I don’t think AI is producing no positive value at all. (Only that it’s producing higher negative value than positive value, at the moment).
- I don’t think nothing interesting or useful is being done with AI.
- I don’t think there are no people who are currently benefitting on net from AI, or no companies benefitting on net from AI.
- I don’t think this will inevitably be true forever (i.e., that AI can never be net-beneficial); I don’t think it’s a fundamental law or inherent property of AI.
- I don’t even think there’s any good reason it needs to be true right now; it’s this way because of poor choices people are making.
AI is good at some things and bad at others
AI is causing damage because it is being deployed at scale to replace working processes with non-working AI-based ones. This is called “AI-first”. It’s extremely trendy. The logic is roughly as follows:
- AI is very powerful at some things, but very bad at others. (Almost no one actually disagrees with this point, if you make them confront it directly.)
- Because the potential gain of using AI at certain tasks is so high, the potential “average benefit” of using AI overall is very high. That is: if you average over all the domains, AI has the potential to create a lot of value.
- The more we use AI, the more of that benefit we will get.
- We should use AI for everything.
Developing “AI-first” ideology makes people forget what the observation in point #1 actually was. As a result, AI ends up being used mainly for things that it is bad at. On net, this cancels out any benefit you got from using AI for appropriate things.
What is AI good at?
- exploration
- research
- investigation
- debugging
- pattern recognition
- anything with “fuzziness” or where precision is impossible
- brainstorming
What is AI bad at?
- reliability / stability
- precision
- performance
- efficiency
- understanding long-running context
AI is bad specifically at all the things needed to create solid foundations. The base layers of anything– a software program, an organization, a system or process– need to be very correct, very efficient, and 100% reliable. This is because any correctness, performance or reliability issue in the base layer will propagate up the stack to anything that depends on it. A filesystem that loses 5% of the data you write to it can never be the basis of a good database system. Yet this is what “AI-first” proposes: to make AI the base layer of an organization, and have everything else built around it and conforming to it. The guiding principle of “AI-first” is: does this help us use more AI?
The domain where AI creates the highest negative value is when it’s used as a substitute for human-to-human communication. Communications and decision-making are especially sensitive to errors, and AI hallucinations are particularly devastating. It only takes a tiny error rate in technical communications to reduce the value of the communication to zero, because it means the information can’t be relied on, which is its sole purpose. This is the equivalent of putting your database on a filesystem that loses 5% of its data, then having to spend all your time working around those failures. It’s just bad design.
I have never seen AI injected in the middle of human-to-human communication without huge loss of efficiency. Any organization that does this stops being about whatever its original purpose was, and becomes a forum for debating existential questions about “the role of AI”, or troubleshooting process problems caused by “unreliable narrators”.
If you think rationally, “using more AI” isn’t a very smart goal. It’s easy to use tons and tons of AI. It’s less easy to find ways to use AI that have positive returns. Nonetheless, the mentality of “AI-first” discourages you from even thinking about whether your use of AI is productive or not, or using evaluation metrics that could check this. The ideology of “AI-first” holds axiomatically that using AI is good, which inevitably leads to increasingly silly and gratuitous applications of AI over time. Everyone who leans too far into “AI-first” ideology ends up building nothing but tooling for AI agents.
However, AI very good at building on top of solid foundations. AI is a very long lever; it amplifies the value of a solid foundation to multiples of its potential.
Despite the highest hopes of AI Enthusiasts, AI has not become a substitute for general basic competence. The people who get the most value out of AI are people who are already good at what they’re doing, and use AI to amplify their skills; the organizations that get the most value out of AI are the ones that already had working processes, and use AI to accelerate rather than replace them.
AI-last
Accordingly, I am proposing a new methodology: “AI-last”.
This should not be confused for “AI never”. THe goal of “AI-last” is to get the benefits of AI for what it’s good at, without paying the costs of what it’s really bad at.
The basic method of “AI-last” is: make a system that works without relying on AI, then use AI to accelerate it. AI should be added on top, not used as a base layer.
In an “AI-last” organization:
- Individuals decide how to use AI for their tasks. Standardizing or enforcing AI workflows is explicitly avoided.
- Work done with AI is not held to a lower standard than work done without AI assistance.
- AIs are not considered moral agents; people are responsible for their AI output. A failure caused by an AI hallucination is the fault of the person who signed off on it, and evaluated the same way as a normal screw-up.
- Your use of AI should never impose costs on people around you. This is considered indistinguishable from doing bad work.
- Nothing is done specifically for the benefit of AI, unless it has an independently compelling non-AI justification. (Note that optimizing things for humans usually improves the effectiveness of AIs in equal or greater measure.)
- A process that can’t be done at all without AI is considered broken.
In short, “AI-last” means running things basically the way you would have run them in the pre-AI era. When AI (first chat, then agentic AI) first came out and people/companies started trying it out, that was “AI-last”, by necessity; AI was being layered on top of already-existing systems. AI was used to fill in gaps in people’s knowledge, rather than as a wholesale replacement of it. When used in this way, there is really no risk of negative productivity gain, because the AI is “on top”, and it’s very easy to tell when it’s helping and when it’s not.
It seems odd, but the least sophisticated ways of using AI are still the most effective. With a few exceptions, almost every “innovation” in AI methodology which has tried to replace human competence with AI has been a downgrade.