AI Allegations Thread

[NoTalentHack]

Bah, leave it to NTH to miss an easy zinger. "Err... yeah, your writing is nice too!"

Sorry, trying to get my humor module upgraded.

 

[Euphony]

Sorry, trying to get my humor module upgraded.

 

[Euphony]

alrighty let's pop the hood on this bad boy and see how it makes it's juice...
View attachment 2294542

... Oh.

Am I the artifical intelligence?

"I am technician Abraham Lincoln calling you from Microsoft in the fraudulent activity on your account."

Our new overlords and masters, same as the old overlords and masters.

 

[MillieDynamite]

This writing unit has a bad motivator.

Sorry, trying to get my humor module upgraded.

 

[NoTalentHack]

This writing unit has a bad motivator.

That's what all my teachers said in school.

 

[NoTalentHack]

Damn. In the middle of all this, I forgot to mention two of my favorite fails when it comes to machine learning and detection algorithms.

The first, which I can't find the article for now, was early in the attempts at ML cancer detection. They had something like a 96% false positive rate back then, which is better than a high false negative rate, but still. Anyways, some researchers had what they thought was a nearly bulletproof version of their software. It was great, diagnosing lung cancer correctly the vast majority of the time.

Until it went into the field, where its prediction abilities dropped through the floor.

The researchers couldn't figure out why at first, especially because they could run new test data through and get the exact results they expected. The problem was that the test data was in the form of chest X-rays images, as you'd expect. They'd used data from several hospitals in the area to train on, and one of them in particular had a stellar oncology unit.

Guess what the ML zeroed in on as a good predictor of whether an image showed cancerous tumors or not?

The X-ray images each had the logo of the hospital that provide them. It was looking for those as a way to quickly determine whether the X-Ray showed cancer, because that hospital got the most severe cases, so they had the most positive data in the system.

The good news is that this is one of those areas where the data is plentiful, well-vetted, operates within a relatively limited range (what's possible in the human body) and life or death. Also, the models are often fairly open. Everyone involved is highly motivated to make sure they get it right, and new ML detection models work really well. But that's only because of those other factors. If a test was wrong in either direction, it could mean death or emotional devastation. Basically none of that is true with, say, detecting AI generated text on a porn site.

The second is not quite as life and death, but it is affecting peoples' abilities to make better lives for themselves. This time, I have an article.

https://themarkup.org/denied/2021/08/25/the-secret-bias-hidden-in-mortgage-approval-algorithms

An investigation by The Markup has found that lenders in 2019 were more likely to deny home loans to people of color than to White people with similar financial characteristics—even when we controlled for newly available financial factors that the mortgage industry for years has said would explain racial disparities in lending.

Holding 17 different factors steady in a complex statistical analysis of more than two million conventional mortgage applications for home purchases, we found that lenders were 40 percent more likely to turn down Latino applicants for loans, 50 percent more likely to deny Asian/Pacific Islander applicants, and 70 percent more likely to deny Native American applicants than similar White applicants. Lenders were 80 percent more likely to reject Black applicants than similar White applicants. These are national rates.

Research has shown that payday loan sellers usually place branches in neighborhoods populated mainly by people of color, where bank branches are less common. As a result, residents are more likely to use these predatory services to borrow money. This creates lopsided, incomplete credit histories because banks report both good and bad financial behavior to credit bureaus, while payday loan services only report missed payments.

Gig workers who are people of color are more likely to report that those jobs are their primary source of income—rather than a side hustle they’re using for extra cash—than White gig workers. Having multiple sources of income or unconventional employment can complicate the verification process for a mortgage, as Crystal Marie and Eskias learned.

Here's another: https://www.npr.org/2020/02/05/8029...llege-graduates-may-pay-higher-interest-rates

Financial firms may be discriminating against people based on where they went to college, a watchdog group says. In particular, the group found that a lender named Upstart appears to be charging higher interest rates on student loans to graduates of historically black or predominantly Hispanic colleges.

A lot more people are getting loans these days from a new breed of lenders known as fintechs, or financial technology firms. And some of these lenders factor in where loan applicants went to college.

"It really raised some alarm flags," said Kat Welbeck, the civil rights counsel at the nonprofit Student Borrower Protection Center.

So her group decided to run a test. It chose a lender called Upstart, in part because it's a fairly prominent fintech that says it considers educational data. And the group was able to easily apply and get loan offers on its website.

The group applied for dozens of loans online — posing as a 24-year-old man. It said he lives in New York, works as a financial analyst and makes $50,000 a year. Each time the group applied for a loan, it kept a whole range of factors constant.

"The only difference was where he went to school," Welbeck said. It applied as if this fictional borrower went to NYU in New York, many other schools and Howard University — one of the country's most famous historically black colleges and universities.

The group found that if the otherwise identical loan applicant went to NYU instead of Howard, there was a striking difference. For a $30,000 personal loan with a five-year term, it found an applicant would pay about $3,500 more in interest and fees if they went to Howard.

"There's no other explanation that we can really come to terms with other than the fact that where this borrower went to school mattered in terms of how Upstart measured their creditworthiness," Welbeck said.

These are closed algorithms, like the AI writing detection ones. There's systemic bias against correcting them, because failing to correct them (at least in the short term) will only save the companies both using and selling them money. The only reason it was caught was because of word of mouth by the people affected, like here, but then followed up by journalists and researchers, which won't happen in this case. And these articles started appearing years ago, but the problems are still happening.

One last thing, a lesson on survivorship bias.

https://www.deanyeong.com/article/survivorship-bias

During World War II, researchers from the Center for Naval Analyses conducted a study on the damage done to returned aircraft after missions. They then recommended adding armor to the areas that showed the most damage to minimize bomber losses to enemy fire.



However, Abraham Wald suggested differently.

Wald was a Hungarian mathematician and a member of the Statistical Research Group (SRG), where he applied his statistical skills to various wartime problems.

He noted that the study was only conducted on the aircraft that had survived their missions. It didn’t paint a complete picture when the bombers that had been shot down were not presented for the damage assessment.

With that, the holes in the returning aircraft were areas that need no extra armor — since the bombers could take damage and still return safely. On the other hand, the areas where the returning aircraft were unscathed are those areas that, if hit, would cause the plane to crash and be lost.

Wald then proposed that the Navy reinforce areas by adding more armor to them — which was a perfect demonstration of how to not fall prey into the survivorship bias.

I know, for example, that I didn't use PWA or other tools before June 6th, regardless of what any algorithm might claim. I also know that my use of them afterwards has been relatively minimal and restricted to manually changing things where I chose to, without using any of their rewriting abilities. And yet, Deputy Fife claims the software he's using detects vast swathes of my writing is AI generated.

I know what I have and haven't done, and I can't be the only one it's pinging wrongly. Maybe Barney's weapon of choice is particularly janky, but I know that other folks in here have sought out various tools to vet their own work and found wildly differing results. I'm not worried that I'll get picked off by Lit's algorithm; if it hasn't happened yet, I can't expect it will, but who knows?

However, just because I can make it through now doesn't mean the algo is good; it's possible I've been flagged as safe in the system, for all I know. If I showed up today, instead of a year ago, maybe that wouldn't be the case.

That's what concerns me out of all this. If I'd submitted my first story in Nov. 2023 instead of Nov. 2022, and the system said, "this is AI generated," I don't know that I would have kept writing. That would have been a hell of a kick in the teeth. So I can't help but feel for the people being affected by it now, and I hope it's not discouraging them.

 

[NoTalentHack]

I think info like this helps a lot, thanks for sharing it. I mean it's clear the detectors are not accurate at this point so maybe people like Barney will stop using them to try to prove something lol. But yeah, I read some new people's stories here and now I'm like maybe I shouldn't because it's gonna suck if they can't finish them because of this. One author I like - and no way am I saying who it is here lolol - they still have an account and they were posting chapters pretty regularly and now they've stopped so I was wondering if they are dealing with this too.

Anyways you were cool about this, well until you got accused of a bunch of stuff lol, and maybe some others are afraid to say anything about the topic - which I can understand after reading this thread - but you have, even though you got some heat for it. Glad you got your story finished and I hope it does good in the contest

Thanks! “New Year, New Experiences” has been approved and goes live tomorrow. No problems in the submission queue at all, so either I’m flagged as safe, or Laurel’s system is better than Deputy Fife’s, or I’m just lucky that her system doesn’t ding me while his does. Regardless, it’s still a problem, even if it isn’t a problem for me.

 

[theWriterInTheNude]

I've been writing on the site for five years and haven't changed my process, but I finally got dinged. I'm glad it was my second, shorter story for the Winter holiday contest so my streak of two years of wring for all the holiday contests is unbroken. I pieced together older versions without as much grammar checking and it still got rejected. It's too late now as the contest deadline is passed. The worst part is not knowing what is causing the software to trip the AI sensor as it were. Now I'm staring at my next stories wondering if I should just throw everything out and give up on writing. It's one of my favorite hobbies, but any editing has me nervous. I spent well over a month working on my story just for it to get rejected as AI written.

 

[NoTalentHack]

deputy Fife is going to be PISSED lololol

"I TOLD YOU TO NIP IT IN THE BUD, LAUREL!"

 

[NoTalentHack]

people have been talking about the length of the sentences a lot and how ai keeps them the same. Can you tell the ai to vary the length of the sentences in a story?

There's a lot of stuff you can do, but you have to think to do it. You're also limited by... Mmm, this is a little inside baseball, and only directly talking about ChatGPT, but it works sort of like this:

ChatGPT works off of "nodes." For our purposes, you can think of those as words, although that's not quite accurate; not going to get into the more technical finer points. These nodes are part of both the prompt(s) you issue to it and the output, so you have to be careful of that. In ChatGPT 3.5, the number of nodes it can keep track of at a time works out to about 4000 words, and 4.0 can handle 25,000. Again, for the people that want to pick nits, this is high level; there are also limits on how long a prompt can be, etc.

Regardless of which version you're using, it's essentially like your phone's autocomplete on steroids. If I tell it, "Describe a beautiful spring day," that's (roughly) 4-5 nodes used, leaving ChatGPT 3.5 about 3,995 to complete my request with. It almost certainly will use nowhere near that many. However, it will look at each word in tandem with the others, and the next one it uses.

You can refine the response with a new query, but that will chew up a certain number of nodes as well. When it gets to the limit of those nodes, it will "forget" the things it talked about previously. That's why it's bad about making longer stories; well, that's one of the reasons it's bad, but it will start "forgetting" those nodes, which means, for example, if you have a character that was a brick layer early on in the story, and it's not mentioned for a while, their job might change to big game hunter or chef or professional hall monitor.

Here's an example that I just asked it to do:

You
Describe a beautiful spring day.


ChatGPT
A beautiful spring day is a sensory delight, a celebration of nature's reawakening after the long, dormant months of winter. As the sun begins its ascent in the clear blue sky, a sense of renewal and vitality fills the air.

The temperature is pleasantly mild, with a gentle breeze carrying the subtle scent of blooming flowers and freshly mowed grass. The world is bathed in a soft, golden light that seems to breathe life into every corner.

Flowers of all colors burst forth from the earth, painting a vibrant tapestry of beauty across the landscape. Cherry blossoms, daffodils, tulips, and magnolias display their vibrant hues, while delicate petals fall like confetti from the trees above. Bees and butterflies flit from blossom to blossom, adding a melodious hum to the symphony of spring.

The sound of birdsong fills the air, as migratory species return home and the local birds announce their presence with cheerful melodies. Their sweet tunes create a harmonious backdrop to the gentle rustling of leaves in the trees and the occasional laughter of children playing in the park.

People emerge from their winter hibernation, shedding heavy coats for lighter, colorful attire. They stroll through parks and gardens, savoring the warmth of the sun on their skin and the feeling of soft, green grass beneath their feet. Picnics are spread on checkered blankets, and the aroma of grilled food wafts through the air.

As the day unfolds, the sky may be adorned with fluffy white clouds, drifting lazily overhead. In the distance, the distant mountains or rolling hills appear hazy and soft, their once barren slopes now covered in a lush green carpet.

A beautiful spring day is a time of optimism and renewal, a reminder of the endless cycle of nature's beauty. It invites us to savor life's simple pleasures, to connect with the world around us, and to appreciate the fleeting but enchanting moments that spring so generously bestows upon us.


You
Condense that to one paragraph


ChatGPT
A beautiful spring day awakens the senses with its gentle, golden light and mild temperatures. Nature bursts to life, painting the world with vibrant flowers and the sweet scent of blossoms. Birds fill the air with cheerful songs, and people shed their winter layers to bask in the sun's warmth. Picnics are enjoyed on green blankets, and the world comes alive with the laughter of children and the scent of grilled food. Fluffy clouds drift across the clear blue sky, and the landscape transforms into a lush, green paradise, reminding us of the enduring beauty of nature and the joy of simple pleasures.


You
Rework it in the style of Raymond Chandler.


ChatGPT
A beautiful spring day, pal, is like a breath of fresh air after a long, cold slumber. The sun, a golden spotlight in the endless blue sky, casts its warm glow, stirring the world from its winter daze. The streets are alive with color as nature unfurls its secret palette, showcasing vibrant blooms and fragrant blossoms. Birds, the feathered troubadours of the season, belt out tunes that'd make a hardboiled detective crack a smile. Folks ditch their overcoats for shades and leisurely strolls in the park, the scent of grilling meat tickling the nose. Wispy clouds drift by, casting soft shadows on the verdant landscape, a reminder that even in a world of shadows and mysteries, there's still room for beauty and simple pleasures.


You
Give it to me as Earnest Hemingway instead.


ChatGPT
A beautiful spring day arrives, simple and unadorned, like the first breath of life itself. The sun, a warm, comforting presence in a vast expanse of blue, beckons nature from its winter slumber. Flowers of vivid hues and delicate scents burst forth, a testament to nature's resilience. Birds, the poets of the skies, sing their songs with a grace and purity that stirs the soul. People, shedding their winter layers, embrace the warmth of the sun, relishing the simple joys of a picnic on the green. The clouds drift lazily across the horizon, a reminder that amidst life's complexities, there is beauty in the uncomplicated, in the pure essence of a spring day.


You
What would it look like if Dr. Seuss wrote it?


ChatGPT
On a splendid spring morn, oh, what a sight!
The sun pops up, all cheerful and bright.
It paints the sky in a dazzling blue,
With puffy white clouds, just like cotton candy, too!

The flowers wake up, they stretch and they yawn,
Their petals unfurl, like a brand-new dawn.
Tulips and daisies, all colors so grand
,Pop up in the garden, in the soft, springy sand.

The birds start to sing, tweet-a-dee and tweet-a-doo,
Their melodies dance, as they flit and they flew.
From treetops to rooftops, they put on a show,
Their songs fill the air, from high to down low.

People come out, they're ready to play,
In the meadows and parks, they frolic all day.
With picnics and laughter, they giggle and cheer,
As springtime's magic draws everyone near.

So, on this spring day, in the most Seussian way,
The world comes alive in a colorful display.
With joy and with wonder, and whimsy galore,
Springtime's a treat, that's for sure, that's for sure!


You
How about the King James Version of it?


ChatGPT
And lo, on a glorious day in the vernal season, the radiant sun did ascend into the heavens, casting its golden light upon the Earth. The azure sky stretched forth, a canvas of divine perfection, adorned with ethereal clouds, wispy as angelic wings. The flora, in a splendid array of colors, blossomed forth, revealing their delicate petals and fragrant aromas. Birds of the air did sing with joyful voices, their melodies a testament to the Creator's grace.

The people, having shed their winter garments, did gather in fellowship, partaking in the simple pleasures of picnics and laughter upon the verdant fields. As the sun did grace the land, casting its benevolent warmth, the world rejoiced in the resurrection of life. In the grand tapestry of creation, springtime's arrival was a divine proclamation of renewal and rebirth, a testament to the majesty of the Almighty's handiwork.

It'll take a shot at the vast majority of the things you ask it to do, although direct porn or things like bomb making are right out. But it'll take a shot at translation, rephrasing, mimicry, etc. It's still nowhere near as good as a real human, but it's good enough for non-creative or low-creative work.

It's also very often wrong about factual things, though, because it is just doing autocomplete. It builds what amounts to a probabilistic tree about what the nexst word should be based on what you asked it for and what it's said before. That's why you can't rely on it for anything that's actually important. It will get the right "shape" of things, but be completely wrong about the details, like when it quoted nonexistent case law when a lawyer tried to use it to draft a brief, and he got in big trouble, for example.

 

[NoTalentHack]

also just think if you win. The meltdown here would be fucking epic lololol

Oh, it would be, but that's unlikely. My followers, enough of them, won't like that there's a happily nonmonogamous married couple in it. Beyond that, it was pretty rushed. Still, "It's Christmas, Theo! It's a time of miracles!"

 

[StillStunned]

wait this got me thinking - scary I know lolol. If I was to write a 750 word story with no ai tools, how would you suggest I do it to try to make sure I don't trip the ai detector? I have read lots of different things - different sentences length and no adverbs and no dialogue tags (kinda makes it hard to tell who is talking without them but I will try to use them sparingly) What else? And this will have to be in non erotic because I ain't writing a sex scene lol. I will pinky swear to the owners I didn't use ai. Now the story might get kicked back for being shitty but so far it isn't against the rules for a story to be shitty. And it's not like I have other stories here that can be removed. I am the perfect guinea pig. Maybe I can set a record for story with the lowest score here lolol. I am open to suggestions for what I should do and what I should avoid

Shouldn't be too hard. Just write something. You don't want to do a sex scene, so perhaps write a seduction scene. Dialogue-heavy. Two people in a park, one trying to convince the other to go on a date. The other isn't immediately disgusted, but more amused than eager.

Write it, then read it aloud. Does it sound natural, like two people talking? Is it all in language that you'd use every day, including any descriptive passages?

 

[alohadave]

It's also very often wrong about factual things,

It baffles me that people haven't figured this out. I tested with a subject I know well and it is good at sounding correct, while the details are completely wrong.

I saw a post on Reddit the other day asking why ChatGPT can't do math.

 

[alohadave]

holy shit that is something else. Thanks for sharing this. I would like to see it try to write like a more experimental author because it does seem to go back into the same stuff even with the different examples ( well except for Dr Seuss lol) Also you got any tips if I write a 750 word story with no ai tools and I want it not to trip the detector here? I will try to keep the characters job the same for the whole story lolol

750 word stories are tight and focused. You write a story but you leave out any extraneous details and descriptions.

Cut till it hurts, then cut some more.

 

[djrip]

yeah I am gonna stick with one scene, a guy getting his heart broken. I already have the summary - "Christmas ain't what it's cracked up to be" but is that gonna trip the detector? Also if my character works at Target am I allowed to say that? Or should I just say big box store or something. I will probably do that just to be safe

and what about the title A Broken Heart for Christmas. I mean it ain't Shakespeare but it will do am I right lolol

You keep asking what will trip the detector... If we knew that we wouldn't be having all these problems!

 

[StillStunned]

yeah I am gonna stick with one scene, a guy getting his heart broken. I already have the summary - "Christmas ain't what it's cracked up to be" but is that gonna trip the detector? Also if my character works at Target am I allowed to say that? Or should I just say big box store or something. I will probably do that just to be safe

and what about the title A Broken Heart for Christmas. I mean it ain't Shakespeare but it will do am I right lolol

You're starting to sound remarkably like JohnSmi123. Just start writing!

 

[StillStunned]

the female characters name

Congratulations, you're already further than ol' John ever got!

 

[StillStunned]

I have written 228 words how the hell am I gonna get to 750 lolol

Personally, I recommend staying away from the AH when it's in a silly mood. I've spent 5 minutes actually working all afternoon. And 10 minutes writing.

 

[StillStunned]

One of us! One of us!

 

[NoTalentHack]

Not till it passes the detector lolol

Seven of Nine of us!

 

[MillieDynamite]

Is the 750 challenge on now?

ok so it's at around 900 words but I feel like it's nice and lean lol. I submitted it so we'll see

 

[Voboy]

Can you tell the ai to vary the length of the sentences in a story?

…why not just do it yourself?

The premise of your question is precisely why the site is concerned.

 

[alohadave]

Is the 750 challenge on now?

Not for a few months.

 

[djrip]

…why not just do it yourself?

The premise of your question is precisely why the site is concerned.

The question was about AI capabilities. He just got done submitting a story very specifically written without AI and talking about it in this thread, so why not do it yourself isn't particularly relevant here. It would be like if I asked, can a dog be trained to go in the toilet, and someone said, you can do that yourself! I know I can! I want to know if we can get the dog to do it!

 

[MillieDynamite]

Good, 750 words is such a long write I need a lot of lead time.

Not for a few months.