See also: AI-Generated “Workslop” Is Destroying Productivity [1]
[1] https://hbr.org/2025/09/ai-generated-workslop-is-destroying-...
What AI has done is accelerate and magnify both the positives and the negatives.
There are a lot of white-collar tasks that have far lower quality and correctness bars. "Researching" by plugging things into google. Writing reports summarizing how a trend that an exec saw a report on can be applied to the company. Generating new values to share at a company all-hands.
Tons of these that never touch the "real world." Your assistant story is like a coding task - maybe someone ran some tests, maybe they didn't, but it was verifiable. No shortage of "the tests passed, but they weren't the right test, this broke some customers and had to be fixed by hand" coding stories out there like it. There are pages and pages of unverifiable bullshit that people are sleepwalking through, too, though.
Nobody already knows if those things helped or hurt, so nobody will ever even notice a hallucination.
But everyone in all those fields is going to be trying really really hard to enumerate all the reasons it's special and AI won't work well for them. The "management says do more, workers figure out ways to be lazier" see-saw is ancient, but this could skew far towards "management demands more from fewer people" spectrum for a while.
In all areas where there's less easy ways to judge output there is going to be correspondingly more value to getting "good" people. Some AI that can produce readable reports isn't "good" - what matters is the quality of the work and the insight put into it which can only be ensured by looking at the workers reputation and past history.
That's not obvious at all if the AI writing the tests is different than the AI writing the code being tested. Put into an adversarial and critical mode, the same model outputs very different results.
Someone needs to build an agentic tool that does strict, enforced TDD.
Obviously this is only partially true but it's true enough.
It takes humans quite a long time to learn the external context that lets them write good tests IMO. We have trouble feeding enough context into AIs to give them equal ability. One is often talking about companies where nobody bothers to write down more than 1/20th of what is needed to be an effective developer. So you go to some place and 5 years later you might be lucky to know 80% of the context in your limited area after 100s of meetings and talking to people and handling customer complaints etc.
I have been doing this with coding agents across LLM providers for a while now, with very successful results. Grok seems particularly happy to tell Anthropic where it’s cutting corners, but I get great insights from O3 and Gemini too.
Except the test suite isnt just something that appears and the bugs dont necessarily get covered by the test suite.
The bugginess of a lot of the software i use has spiked in a very noticeable way, probably due to this.
>But everyone in all those fields is going to be trying really really hard to enumerate all the reasons it's special and AI won't work well for them.
No, not everyone. Half of them are trying to lean in to the changing social reality.
The gaslighting from the executive side, on the other hand, is nearly constant.
Consider, for example, the following python code:
x = (5)
x = (5,)
Now imagine trying to spot that one missing comma among the 20kloc of code one so proudly claims AI helped them "write", especially if it's in a cold path. You won't see it.
Disagree.
Even though performing checks on dynamic PLs is much harder than on static ones, PLs are designed to be non-ambiguous. There should be exactly 1 interpretation for any syntactically valid expression. Your example will unambiguously resolve to an error in a standard-conforming Python interpreter.
On the other hand, natural languages are not restricted by ambiguity. That's why something like Poe's law exists. There's simply no way to resolve the ambiguity by just staring at the words themselves, you need additional information to know the author's intent.
In other words, an "English interpreter" cannot exist. Remove the ambiguities, you get "interpreter" and you'll end up with non-ambiguous, Python-COBOL-like languages.
With that said, I agree with your point that blindly accepting 20kloc is certainly not a good idea.
Those are both syntactically valid lines of code. (it's actually one of python's many warts). They are not ambiguous in any way. one is a number, the other is a tuple. They return something of a completely different type.
My example will unambiguously NOT give an error because they are standard conforming. Which you would have noticed had you actually took 5 seconds to try typing them in the repl.
> Those are both syntactically valid lines of code. (it's actually one of python's many warts). They are not ambiguous in any way. one is a number, the other is a tuple. They return something of a completely different type.
> "Now imagine trying to spot that one missing comma among the 20kloc of code"
My reply, "Your example will unambiguously resolve to an error in a standard-conforming Python interpreter." tried to respond to the possibility of such an issue. Even though it's probably not the program behavior you want, Python, being a programming language, will be 100% guaranteed to interpret it unambiguously.
I admit, I should have phrased it a bit more unambiguously than leaving it like that.
Even if it's hard, you can try running a type checker to statically catch such problems. Even if it's not possible in cases of heavy usage of Python's dynamic typing feature, you can just run it and check the behavior at runtime. It might be hard to check, but not impossible.
On the other hand, it's impossible to perform a perfectly consistent "check" on this reply or an email written in a natural language, the person reading it might interpret the message in a completely different way.
The few times I've tried giving LLMs a shot I've had them warning me of not putting some validations in, when that exact validation was exactly 1 line below where they stopped looking.
And even if it did pass an AI code review, that's meaningless anyway. It still needs to be reviewed by an actual human before putting it into production. And that person would still get scrolling blindness whether or not the ai "reviewer" actually detected the error or not.
I didn't say they were guaranteed to find it: I said they were really good at finding these sorts of errors. Not perfect: just really good. I also didn't make any assumption: I said in my experience, by which I mean the code you shared is similar to a portion of the errors that I've seen LLMs find.
Which LLMs have you used for code generation?
I mostly use claude-opus-4-6 at the moment for development, and have had mostly good experiences. This is not to say it never gets anything wrong, but I'm definitely more productive with it than without it. On GitHub I've been using Copilot for more limited tasks as an agent: I find it's decent at code review, but more variable at fixing problems it finds, and so I quite often opt for manual fixes.
And then the other question is, how do you use them? I tend to keep them on quite a short leash, so I don't give them huge tasks, and on those occasions where I am doing something larger or more complex, I tend to write out quite a detailed and prescriptive prompt (which might take 15 minutes to do, but then it'll go and spend 10 minutes to generate code that might have taken me several hours to write "the old way").
Assistant is dispatching a courier to get medical records. AI auto completes to include the address. Normally they wouldn't put the address, the courier knows who we work with, but AI added it so why not. Except it's the wrong address because it's for a different doctor with the same name. At least they knew to verify it, but still mistakes like this happening at scale is making the other time savings pretty close to a wash.