But if you’re not an expert, it’s more likely that everything will just sound legit.

Oh, absolutely! In my field, the answers made up by an LLM might sound even more legit than the accurate and well-researched ones written by humans. In legal matters, clumsy language is often the result of facts being complex and not wanting to make any mistakes. It is much easier to come up with elegant-sounding answers when they don't have to be true, and that is what LLMs are generally good at.

And then we went back to “it’s rarely wrong though.”

I am often wondering whether the people who claim that LLMs are "rarely wrong" have access to an entirely different chatbot somehow. The chatbots I tried were rarely ever correct about anything except the most basic questions (to which the answers could be found everywhere on the internet).

I'm not a programmer myself, but for some reason, I got the chatbot to fail even in that area. I took a perfectly fine JSON file, removed one semicolon on purpose and then asked the chatbot to fix it. The chatbot came up with a number of things that were supposedly "wrong" with it. Not one word about the missing semicolon, though.

I wonder how many people either never ask the chatbots any tricky questions (with verifiable answers) or, alternatively, never bother to verify the chatbots' output at all.

FWIW, I work in a field that is mostly related to law and accounting. Unlike with coding, there are no simple "tests" to try out whether an AI's answer is correct or not. Of course, you could try these out in court, but this is not something I would recommend (lol).

In my experience, chatbots such as Copilot are less than useless in a context like ours. For more complex and unique questions (which is most of the questions we are dealing with everyday), it simply makes up smart-sounding BS (including a lot of nonexistent laws etc.). In the rare cases where a clear answer is already available in the legal commentaries, we want to quote it verbatim from the most reputable source, just to be on the safe side. We don't want an LLM to rephrase it, hide its sources and possibly introduce new errors. We don't need "plausible deniability" regarding plagiarism or anything like this.

Yet, we are being pushed to "embrace AI" as well, we are being told we need to "learn to prompt" etc. This is frustrating. My biggest fear isn't to be replaced by an LLM, not even by someone who is a "prompting genius" or whatever. My biggest fear is to be replaced by a person who pretends that the AI's output is smart (rather than filled with potentially hazardous legal errors), because in some workplaces, this is what's expected, apparently.

If computers become capable of mass-producing stuff other computers will like, but many humans won't, this might also lead to a quick decline of algorithm-based search engines, social media feeds etc. (as has been discussed here before, of course).

I think most cons, scams and cults are capable of damaging vulnerable people's mental health even beyond the most obvious harms. The same is probably happening here, the only difference being that this con is capable of auto-generating its own propaganda/PR.

I think this was somewhat inevitable. Had these LLMs been fine-tuned to act like the mediocre autocomplete tools they are (rather than like creepy humanoids), nobody would have paid much attention to them, and investors would have started to focus on the high cost of running them quickly.

This somewhat reminds me of how cryptobros used to claim they were fighting the "legacy financial system", yet they were creating a worse version (almost a parody) of it. This is probably inevitable if you are running an unregulated financial system and are trying to extract as much money from it as possible.

Likewise, if you have a tool capable of messing with people's minds (to some extent) and want to make a lot of money from it, you are going to end up with something that resembles a cult, an LLM or similarly toxic groups.

I think this has happened before. There are accounts of people who completely lost touch with reality after getting involved with certain scammers, cult-leaders, self-help gurus, "life coaches", fortune tellers or the like. However, these perpetrators were real people who could only handle a limited number of victims at any given time. Also, they probably had their very specific methods and strategies which wouldn't work on everybody, not even all the people who might have been the most susceptible. ChatGPT, on the other hand, can do this at scale. Also, it was probably trained from all websites and public utterances of any scammer, self-help author, (wannabe) cult leader, life coach, cryptobro, MLM peddler etc. available, which allows it to generate whatever response works best to keep people "hooked". In my view, this alone is a cause for concern.

I think we don't know how many people might be at risk of slipping into such mental health crises under the right circumstances. As a society, we are probably good at protecting most of our fellow human beings from this danger (even if we do so unconsciously). We may not yet know what happens when people regularly experience interactions that follow a different pattern (which might be the case with chatbots).

Just guessing, but the reported "90% accuracy" are probably related to questions that could be easily answered from an FAQ list. The rest is probably at least in part about issues where the company itself f*cked up in some way... Nothing wrong with answering from an FAQ in theory, but if all the other people get nicely worded BS answers (for which the company couldn't be held accountable), that is a nightmare from every customer's point of view.

At the very least, actual humans have an incentive not to BS you too much, because otherwise they might be held accountable. This might also be the reason why call center support workers sound less than helpful sometimes - they are unable to help you (for various technical or corporate reasons) and feel uneasy about this. A bot is probably going to tell you whatever you want to hear while sounding super polite all the time. If all of it turns out to be wrong... well, then this is your problem to deal with.

As usual with chatbots, I'm not sure whether it is the wrongness of the answer itself that bothers me most or the self-confidence with which said answer is presented. I think it is the latter, because I suspect that is why so many people don't question wrong answers (especially when they're harder to check than a simple calculation).

Similar criticisms have probably been leveled at many other technologies in the past, such as computers in general, typewriters, pocket calculators etc. It is true that the use of these tools and technologies has probably contributed to a decline in the skills required for activities such as memorization, handwriting or mental calculation. However, I believe there is an important difference to chatbots: While typewriters (or computers) usually produce very readable text (much better than most people's handwriting), pocket calculators perform calculations just fine and information from a reputable source retrieved online isn't any less correct than one that had been memorized (probably more so), the same couldn't be said about chatbots and LLMs. They aren't known to produce accurate or useful output in a reliable way - therefore many of the skills that are being lost by relying on them might not be replaced with something better.

In any case, I think we have to acknowledge that companies are capable of turning a whistleblower's life into hell without ever physically laying a hand on them.