Hacker News Reader

Comment by danpalmer | Hacker News Reader

Important clarification: consumers are turned off by AI marketing material.

I'd be interested to see real studies into consumer satisfaction with AI features in existing products. My gut feeling is that people don't like (visible) AI in things they use, but that's biased by me reading online reporting about the failings of these features, I wouldn't be too surprised if it turns out people mostly like them.

View on HN

I am certainly turned off by “AI” customer support. The ones I’ve encountered are actively terrible, and I’d rather click through a tree of first-level fix-it-yourself advice than get walked through it by a stunningly poor AI.

But sure, it would be spiffy if Siri was more capable, as long as it didn’t become susceptible to injection attacks from my photo album in the process…

When The Weather Network first included an AI chat function, one of their sample questions was about stargazing.

I asked the exact same question to see what kind of response it would give, and it said that it couldn't answer because weather does not affect star gazing. Really?

I took a screencap, because this was such an epic example of why companies should actually test things before giving public access to them.

Can you actually test a component that is, by design, a black box? ("by design" of the current models. Yes, the systems could be re-made traceable. But then it would also turn out on how much pirated material the LLM has trained, and hoo boy would the IP lawyers show up. That's the con: "if it's a blackbox, you can't prove whence the corpus, nyah nyah!")

Our "testing" involves sending it 50 prompts and seeing if it bullshits too much and revising the prompt until it bullshits to acceptable levels.

I suspect that's how the model makers test it as well, just fling crap at the wall and see which wall lets the most crap slide off of it (but never all of it)

This is called Fuzzing. And it is a major component of software testing. Especially in infosec, but I digress. The reasoning behind it is exactly the same.

Good call. A case that is actively in court right now, is based on a point that you can sorta remake Johnny Be Good with Sonu.

However, to do so, you have to already break copyright law by feeding it copyrighted lyrics.

I did their experiment, but instead of "Go, Johnny, Go, Go Go!" I put in lyrics like "Run, Forrest, Run, Run, Run" (a Gump reference) and described the music style accordingly.

It came out exactly like you'd expect. And the case is demonstrating a mirage where if you steal the song's actual lyrics, describe the style, you'll get something akin to what they might have written... except that there's no way to prove then that the model has even heard that original song before, because you gave it the lyrics and the style, and the rest is clearly bound to be similar as a result.

I brought this up in a previous HN thread, it's what I call "probabilistic UX".

https://news.ycombinator.com/item?id=39954719

We don't have much experience building automated systems that are non-deterministic. Normally, in computer engineering, if we couldn't predict the results of an operation, we'd consider that operation buggy, broken or at best, flaky. Building consumer user interfaces for black box magic is a whole new ballgame.

> We don't have much experience building automated systems that are non-deterministic

Our brains evolved to turn chaos into patterns that help us survive. Determinism is good from that point of view and nondeterministic behaviour is discarded as useless. Nature in general operates in a deterministic way, a banana tree doesn't randomly produce kittens. AI and LLMs in particular are nothing but misinformation and IP appropriation systems. No wonder people reject them.

I bet that if you included an expensive verification step by GPT 4 a lot of errors would go away.

It always circles back to "follow the money":

- the system could theoretically trace the inputs, but then IP lawyers would eat company alive; pretending the corpus isn't pirated, so it can be cheap.

- the system could verify and re-check, but that would require a massive compute increase; spouting bullshit, so it can be cheap

- indeed a lot of answers around here go with "let's waste money by throwing it at the system until a human likes the result" - not cheap, but profitable for the provider

"It sorta kinda works, most of the time, when it doesn't, Retry Reboot Reinstall^W^W, and if it's still broken, you're SOL - and thats by-design, we're not fixing the product." The word you're looking for is enshittification.

Every time I hear it contains AI it sounds like the features' outcome will be uncertain. Especially with more experience. If your feature were good you wouldn't have to mention AI to show it is awesome.

More than not AI is used as an excuse for the feature to be bad.

I tend to agree. If you asked about a feature which uses some kind of ML under the hood to deliver that feature you'd different results. I think companies should urgently pull back on mentioning AI at all if they want any credibility at this point.

Yup. "do you like that you can search for a name in your photo album" will get different answers then asking about AI.

That is a problem with Facebook, where people can tag you in images, even if you don't notice or approve it (you can opt out, but that's the opposite of what is optimal).

Funny to see that come up again, but in a different context.

There's lots of different implementations. For example Immich (https://immich.app/) is fully self-hosted and does local recognition. It doesn't have to be an online service stealing your data.

That was never previously marketed as 'AI', though (I suspect because when CV started seeing this sort of application, the memory of the _previous_ AI winter was still fairly fresh, so everyone carefully avoided the term). In marketing, 'AI' tends to mean _generative_ AI these days.

Okay then at what point do the users become reliable narrators? The AI cheerleaders can take comfort in that AI is in the hands of the average person now—but they also have to contend with any negative reactions as well.

AI isn’t some behind the scenes technology (any more). And companies with their “AI” gimmick often highlight it with a star or something. People can chat with these things as easily as doing a Google search.

I'd buy that people like "AI" features writ large; automatic image classification on your phone, say. However, I'm not convinced anyone much likes the existing applications of LLMs (mostly terrible obstructive chatbots, and blogspam).

> I'd be interested to see real studies into consumer satisfaction with AI features in existing products

mmm... I guess....

I mean, a feature that is not 'obviously AI' is just a feature; even if you normalized it against 'normal' features with no AI as part of them, surely the deviance from random noise would be negligible?

In a double blind test, you would have three cohorts; control, placebo and treatment (ie. AI).

Since the placebo and treatment groups would receive the same feature with or without AI, you'd be looking at absolutely nothing meaningful in the data you collect.

> I wouldn't be too surprised if it turns out people mostly like them.

How can users POSSIBLY have a positive / negative / ANY KIND of meaningful response to a feature based on the unknown backend implementation?

If you put a chatGPT interface on your product and don't put an AI label on it, I guess most people, not being completely stupid, will not meaningfully distinguish between it and the equivalent feature with "Artificial Intelligence" painted on the side. An AI chat bot is an AI chat bot. You're not fooling anyone by calling it an 'intelligent assistant' instead of 'chatGPT'. :P

I mean, I'm just guessing, you could study it. ...but, I guess it's probably not worth bothering.

It's probably more likely the take-away here is: Make your product amazing; if it uses AI, hide that fact that it uses AI if you can.

> would receive the same feature with or without AI,

I'm not sure how you'd do that, given many features wouldn't exist without AI. We don't know how to implement them that way. The only existing choices are "no feature" and "AI feature".

Reply on news.ycombinator.com