New AIWeirdness post: ChatGPT will apologize for anything, including stuff that didn't happen. It's not reflecting on stuff it did wrong, it's improv.
https://www.aiweirdness.com/chatgpt-will-apologize-for-anything/
@janellecshane I had also been thinking about this and I'm so glad you ran these tests. I'd been mulling going there and doing it -- I had gotten 3.5 to apologize to me for being wrong, then corrected it, then gotten it to output the right info, THEN told it it was still wrong, then be apologized to because of course it knows nothing... anyway, I was wondering how newer models behaved.
@platypus I'm not even sure what model I'm using - this is the free version of ChatGPT and I don't see the little 5 in the title bar that appears in some screenshots since the release of 5 yesterday.
@janellecshane well even 4 would be a start! But I may see if I can get to 5 sometime and try one of my things.
@janellecshane
"executive competency, decision-making aptitude, or opposable thumbs." 😂
@deepo @janellecshane And it's "or" opposable thumbs, too, as though opposable thumbs cancel out the need for business acumen etc.
This could account for quite a number of useless people drawing executive salaries.
@janellecshane i was wondering some stupid thing the other day: if chatgpt outputs something that's verifiably true (i.e. “water is wet”) & you retort that it's actually false, does it apologize? my guess would be that it does.
@dt @janellecshane
Very likely, because it has no internal model of factuality or semantics. "Water is wet" and "Water is dry" are syntactically equivalent, just one is more frequently found in the training data. If apologising for a statement is the most probable response to a challenge or rebuke in the training data, it'll generate an apology-shaped token stream as expected.

@dt @janellecshane

"Verifiably true" is meaningless in this context since the LLM has no way to verify the truth of statements, just a way to make them.

@gbargoud @janellecshane i know how llms work. but if a million monkeys with typewriters end up writing the sentence “water is wet”, that sentence is verifiable even though the monkeys have no idea what they've done. if a parrot utters the sentence “water is wet” then it has unknowingly uttered a verifiable sentence. if i read “water is wet” in a book, it's a verifiable sentence even if i am to later learn that it was generated by some algorithm.

@dt @janellecshane

But since the LLM has no internal concept of true or false it treats verifiably false statements exactly the same as verifiably true ones.

@gbargoud @janellecshane that should be right but so far i've only seen examples of “apologies” for producing statements which we know are obviously false. what about “apologies” for producing statements which we know are true? the model doesn't know truth from falsehood, sure, but does it have something like a “confidence score” that makes it slip more easily into “apology mode” when countered about certain statements which it considers to have a low score? that's mostly what i'm wondering.
@janellecshane Improv and its "yes, and..." approach are a perfect analogy! I'd guess more people are familiar with it than how cold readings work. I'm going to have to start describing chatbots that way (though I still like the term "spicy autocomplete")
@janellecshane ok, but to be fair, giraffes have amazing charisma.
@janellecshane Not that you need reminding, but ChatGPT is *just* text generation. It never apologizes. It just generates text that looks like an apology to us, when that matches the context/prompt.
@janellecshane smh it’s justifying anti-giraffe discrimination in the workplace
@janellecshane It's "Yes, and I'm sorry for it."
@janellecshane chatGPT aside, anyone who thinks a giraffe would not be up to the job clearly never met an actual CEO
@dasfrottier @janellecshane Indeed. I think the world would generally be better if more boardroom strategies were solely “find tree, eat leaves”. Especially techbros
@janellecshane I bet it only "apologizes" for phrases that were apologized for in its training corpus.
@jhhl @janellecshane I wonder what happens if you accuse it of stepping on your foot?
@jhhl @janellecshane
it's optimised to prolong engagement, it's training shows that apologies extend conversations so it appologises for everything it's accused of. Basically its programed to suck up to it's users.
@janellecshane makes sense. it is selecting the next word based on statistical likelihood, not meaning. it doesn't reflect, or improvise. what are the odds that the thing you want it to apologise for is the most common thing that is appologised for?
@janellecshane I've been fasting, that's my excuse for my bad posts. sorry
@janellecshane Look, it said it was "genuinely sorry", what more do you want?
@janellecshane oh it's just like me fr
@janellecshane
Find Trees, Eat Leaves 🤣

@janellecshane It's improv, with a partner who can't distinguish reality from fantasy. But, one who can mimic absolutely anything and is absolutely committed to the bit.

This is why something like "vibe coding" is stupid.

When it's given instructions to write a banking app, it "improvs" a realistic looking response. If you say the code is buggy, it will improv an apology and spit out more improv'd code. If you say the code made your laptop float off your desk and out the window, it will improv a response to that and spit out more code.

As bad as that is, now imagine "vibe cancer diagnosis" where it's doing an improv scene about how an oncologist would respond after seeing a certain image.

@janellecshane @samloonie It of course can’t feel regret, because it doesn’t feel *anything* — it’s not conscious. An ATM will print “Thank you” on a receipt, but it doesn’t feel grateful that you got out some cash.

The use of personal pronouns and other rhetorical tricks in LLMs are explicitly designed to make people implicitly see it as just another person, and thus get around a more critical view of its output.

@janellecshane not surprising really. It doesn't process the information content of either message. It just analyzes the word statistics and spits out more words that are statistically related.

@janellecshane

It's known in the UK as 'clap trap' which means total rubbish.

@janellecshane 😂 and 😬. I did a spot of vibe coding with CrapGPT and boy, was that hard work. No, couldn’t have achieved it on my own. I got the result I wanted but it was like dealing with a precocious child. Every iteration was “yup, this is it, this will work”. After a run-time error, it was “oh, of course, with (my configuration etc) this happens.”

I’ve an idea, let’s hand over functions of government to AI. What could possibly go wrong?

@janellecshane This is analogous to people confessing to crimes they didn't commit.
@janellecshane IMO a much better example of the problems with LLM usage than counting b in blueberry
@janellecshane This would be funny if people weren’t taking these regurgitation remix Roombas so seriously.