RANT Why AI Is Getting Worse, Not Better, at Everything
Every headline says AI is about to change everything. Mostly what's changed is how much time I spend fact-checking it afterwards.
I searched for a chip’s rated power draw last month and got three AI-generated summaries before I found a source that actually cited the manufacturer’s own datasheet, and the three summaries didn’t agree with each other, not one of them hedging so much as an “I’m not sure” anywhere in the answer, which is the part of this that actually worries me, since the confidence never wavers whether the answer underneath it is right or wrong, and that means confidence has stopped being any kind of signal for accuracy at all.
Google’s pizza-glue answer
A search engine that returns nothing is annoying. A search engine that returns a fluent paragraph stating the wrong answer as settled fact is dangerous, because most people (reasonably, I’d say, if a little embarrassingly) trust fluent paragraphs. Google found this out the hard way when its AI Overviews feature told people to add glue to pizza sauce to help the cheese stick, sourced, as far as anyone can tell, from a decade-old joke comment on Reddit that the model had no way of recognising as a joke (Forbes has the full rundown).
Marketing copy for chatbots
Open LinkedIn for four minutes and count how many posts use “revolutionary,” “next-level” or “disruptive” to describe a feature that reorders a to-do list slightly differently than it did last week. I’d assumed, back when this site started and I said general grievances would get a look in whenever something on the internet had been built badly enough to notice, that the target would be launchers and subscriptions. Marketing copy for chatbots is a far bigger target than either of those, and a far softer one, mostly because nobody writing it has to cite a source for any of it.
A subscription refund and a chatbot
I spent forty minutes going in circles with a support chatbot last month trying to get a refund on a subscription I’d cancelled and been charged for anyway, and the bot’s entire strategy was rephrasing its previous unhelpful answer in slightly different words each time I said it hadn’t helped, eventually offering to “escalate to a specialist,” which turned out to mean a different chatbot. I got a human on the fourth attempt, and the human fixed it in ninety seconds, which the company doing this presumably calls efficiency and I’d call a queue with extra steps and a worse hold message.
A colleague’s AI-written config parser
A colleague passed me a pull request last month with a config parser an AI assistant had written for him, and it worked right up until it hit a value with a trailing comment, at which point it silently dropped half the config rather than erroring, with nothing in the code suggesting anyone had read it before submitting it, which is the actual problem here rather than the tool having been used at all. I fixed it in about the time it would have taken to write the thing properly the first time, and I don’t know how many other config parsers just like it are sitting in production somewhere already, quietly dropping values with trailing comments, because nobody’s going to go back and audit code that’s already shipped and appears to work.
My job is understanding why software does what it does rather than producing plausible-sounding software that might, and no chatbot’s shipped a working config parser to a stranger without a human checking it first, though I definately seem to spend more of every week checking someone else’s work than I did two years ago, and I’d like that noted somewhere other than in my own head.