AI Megathread

Pavel

@Trashcan That footnote is giving me flashbacks to a stand-up argument I had with one of my educators when I was a wee lad (of around 19?) where I held the view that the Oxford comma is never optional. Academic arguments are weird, man. I was/am right, though.

Roz

@Aria said in AI Megathread:

You should be–if you know how to use them properly–leaving no spaces between the dash and the word.

i came over here to complain about how those are clearly en-dashes, and then i discovered in the quoted text that you did the usual double hyphen dealio we all do, and the forum software just renders it as en-dashes?!?!?!

nodebb why you do this. clearly if you’re going to render that out it should be — not –. absolute clown behavior, nodebb

Pavel

It’s markdown: dashdash is en, dashdasdash is em.

Test–test
Test—test

Test – test
Test — test

Roz

@Pavel that’s dumb

Pavel

@Roz Having two kinds of dashes is dumb.

And this is the part where I leave so Roz can’t em dash my brains against the rocks.

Trashcan

@Pavel said in AI Megathread:

The only way you can truly tell if writing is LLM generated and not simply a style you’ve come to associate with LLM is to be comparative.

This is not true; people who are very familiar with AI-generated text can identify it accurately 90% of the time without any access to ‘comparative’ sources.

@Aria said in AI Megathread:

Anything I write professionally would almost certainly be pegged as written by AI,

@Pavel said in AI Megathread:

various institutions are using flawed heuristics – be they AI-driven or meatbrain – to judge whether something is written by an LLM

The fear that human-generated content is going to be flagged as written by AI is mostly overblown. People who are not familiar with AI are not good at detecting it, but when you see stats about how AI detection tools are “highly inaccurate”, that statistic is almost always referring to AI not being flagged (evasion), not false positives. Various studies have found commercial AI detector tools to have very low levels of “false positives”: GPTZero identified human content correctly 99.7% of the time, and Pangram also identified human content correctly over 99% of the time, while Originality.ai did slightly less well at only 98+% of the time.

If we take these numbers at face value, the odds of someone familiar with AI output identifying a piece of writing as suspect and putting it through two different commercial AI detectors and both of them flagging it as AI when it was, in fact, human-written, is in the neighborhood of 0.002%. You’re more likely to die in a given year than to have this happen to you. I’m personally comfortable with that level of risk.

The odds of someone unfamiliar with AI output accusing you off the cuff of AI use and being wrong about it are about 50%. So. You know. Watch out for that one.

Faraday

@Trashcan said in AI Megathread:

This is not true; people who are very familiar with AI-generated text can identify it accurately 90% of the time without any access to ‘comparative’ sources.

You cited a study with a microscopic sample size and flawed methodology, which (as far as I can see) wasn’t peer reviewed or published in a reputable journal with editorial review. It’s interesting, sure, and maybe can lead to future research, but it by no means proves that “people who are very familiar with AI-generated text can identify it accurately 90% of the time”.

Pavel

@Trashcan said in AI Megathread:

This is not true; people who are very familiar with AI-generated text can identify it accurately 90% of the time without any access to ‘comparative’ sources.

“Person very familiar with Vermeer easily spots forgery” is not a surprise. I was speaking about the general population, who are not very familiar with AI-generated text.

Those studies you quoted, while potentially promising, are very small in scale. Another study has indicated that if English isn’t your first language, there’s a higher chance of your work being pulled up as having been written by AI.

@Trashcan said in AI Megathread:

the odds of someone familiar with AI output identifying a piece of writing as suspect and putting it through two different commercial AI detectors

This part, though, is the most bemusing though. The odds of someone familiar with AI output putting it through two different commercial AI detectors in the real world are almost laughably small, in my experience. Academic institutions and non-tech companies aren’t going to fork out for two bits of software that do roughly the same thing, they’re going to go with whomever has the shiniest advertising budget.

MisterBoring

The most funny part of this to me is people keep using the en dash / em dash thing as an example, and I myself don’t know when to use either, so I just do whatever and sometimes it gets autocorrected.

Pavel

@MisterBoring Either @Roz or @Aria explained… somewhere up in the higher reaches of this thread. I got a cramp trying to scroll that far.

Aria

@Pavel said in AI Megathread:

@Aria said in AI Megathread:

You should be–if you know how to use them properly–leaving no spaces between the dash and the word

That is a style guide difference.

ETA: At least it used to be, I haven’t checked recently. But when I was first coming up in the Professional Writing Arena we used some bastardised variant of AP style that required a space between. It also did weird shit with ellipses that I didn’t approve of.

You can use spaces between (I prefer spaces between because ohgodmyeyes), but having a space on one side and not the other like our dumb brand font does is what I was talking about re: stylistic inconsistency. We use a bastardized version of Chicago style where I work that does the no spaces.

Trashcan

@Pavel said in AI Megathread:

are very small in scale

First study:

By using the method described above, we create 6 datasets with around 20K samples each

Second study:

The researchers built a dataset of about 2,000 human-written passages spanning six mediums: blogs, consumer reviews, news articles, novels, restaurant reviews, and résumés. They then used four popular large language models to generate AI versions of the content by using prompts designed to elicit similar text to the originals.

What would you consider an acceptable scale?

@Faraday said in AI Megathread:

You cited a study with a microscopic sample size and flawed methodology,

Fair enough, I can’t find any similar studies with a larger sample size. Most other studies find odds statistically significantly better than a coin flip, somewhere between just barely and the upper 60%s.

@Pavel said in AI Megathread:

The odds of someone familiar with AI output putting it through two different commercial AI detectors in the real world are almost laughably small

Even if we grant that only 2 events must occur, the suspicion (we’ll go with a 50/50) and a single check (with an average from the commercial offerings of 1% false positives), if you’re approaching your 50’s, you’re still more likely to die in a given year than for this to happen to you (0.5%). These tools are aware of the negative ramifications of a false positive and are biased towards not returning them.

Aria

@Pavel said in AI Megathread:

@MisterBoring Either @Roz or @Aria explained… somewhere up in the higher reaches of this thread. I got a cramp trying to scroll that far.

I explained it here. Roz got mad that she didn’t know the dumb reason they’re called en dash and em dash.

somasatori

@Aria said in AI Megathread:

@Pavel said in AI Megathread:

@MisterBoring Either @Roz or @Aria explained… somewhere up in the higher reaches of this thread. I got a cramp trying to scroll that far.

I explained it here. Roz got mad that she didn’t know the dumb reason they’re called en dash and em dash.

it is admittedly a very dumb reason. It sounds like a reason that someone from Long Island would come up with

I have actually had to unlearn using em dashes because I would do it constantly. I use a lot of parentheses now when previously I would just be like – . People have recently assumed that I was using AI (not great for clinical writing) and thus everything is over-parenthized. Over-parenthesesed?

Aria

@somasatori said in AI Megathread:

I have actually had to unlearn using em dashes because I would do it constantly. I use a lot of parentheses now when previously I would just be like – . People have recently assumed that I was using AI (not great for clinical writing) and thus everything is over-parenthized. Over-parenthesesed?

Parenthosophized. Add it to the style guides now, please and thank you.

Pavel

@Trashcan said in AI Megathread:

What would you consider an acceptable scale?

Honestly? More mediums. Media. Whichever. Essays, academic papers, hell even clinical notes. The kinds of writing that will really easily look like AI to anyone who doesn’t know what they’re talking about.

But ultimately, it doesn’t even matter if the tool is very nearly perfect. Many people in many settings, even professional ones, won’t run text through a detector, they’ll look at some shitty guide on the internet and declare something to be AI or not. It’s ultimately a human problem, not a detector problem – they’re going to believe what they want to believe and the detection software will be evidence for them either way: “The detector works perfectly without flaws or errors,” when it agrees with them, and “the detector is easily fooled and full of problems and my brain is better” when it disagrees.

Because we’ve still got stupid old people making stupid old people decisions based on metrics from stupid old people times, like the 70s.

@somasatori said in AI Megathread:

People have recently assumed that I was using AI (not great for clinical writing) and thus everything is over-parenthized. Over-parenthesesed?

I’ve started using semicolons more in my notes:

Client reported improved sleep this week — though still experiencing early-morning waking when stressed.
vs
Client reported improved sleep this week; still experiencing early-morning waking when stressed.

somasatori

@Pavel said in AI Megathread:

@somasatori said in AI Megathread:

People have recently assumed that I was using AI (not great for clinical writing) and thus everything is over-parenthized. Over-parenthesesed?

I’ve started using semicolons more in my notes:

Client reported improved sleep this week — though still experiencing early-morning waking when stressed.
vs
Client reported improved sleep this week; still experiencing early-morning waking when stressed.

This is honestly a great idea. I’m really thankful for the suggestion!

hellfrog

If your rp isn’t boring and hollow, then it won’t ping as AI even if you use em-dashes for whatever reason. It’s not just the dashes.

Faraday

@Pavel said in AI Megathread:

I’ve started using semicolons more in my notes:

Until some article points out that semicolons also occur more often in AI-generated work than in the average (non-professional) writing, and you’re right back where you’ve started. I’m honestly surprised it isn’t mentioned in the wikipedia article, since I’ve seen it highlighted elsewhere.

@Trashcan said in AI Megathread:

These tools are aware of the negative ramifications of a false positive and are biased towards not returning them.

And yet they still do, and not necessarily at the 1% false-positive rate they claim. For example, from the Univ of San Diego Legal Research Center:

Recent studies also indicate that neurodivergent students (autism, ADHD, dyslexia, etc…) and students for whom English is a second language are flagged by AI detection tools at higher rates than native English speakers due to reliance on repeated phrases, terms, and words.

This has been widely reported elsewhere too. It’s a real concern and it has real-world implications on peoples’ lives when they are falsely accused of cheating/etc.

Trashcan

@Faraday said in AI Megathread:

For example, from the Univ of San Diego Legal Research Center:

If we’re getting down to the level of sample size and methodology, it’s probably worth mentioning that this study looked at 88 essays and ‘recent’ in this context was May 2023, or 6 months after the release of ChatGPT. It is safe to assume the technology has progressed.