shiri_noyFeb 11, 2023
Amy Diehl, Ph.D.
ChatGPT writes 15% longer and more critical performance feedback for women employees. It also relies on stereotypes to choose pronouns. Examples: kindergarten teacher, nurse and receptionist = "she". Mechanic and construction workers = "he".
https://www.fastcompany.com/90844066/chatgpt-write-performance-reviews-sexist-and-racist
Feb 11, 2023 at 2:55amWeb
Show threadNat Weaver 🏳‍🌈🏳️‍⚧️Feb 11, 2023
@amydiehl
Show threadChookMother 🇦🇺🦘Feb 11, 2023

@amydiehl

So it's a kind of mirror? Could be a good thing.

Show threadMicah WarrenFeb 11, 2023

@anne_twain @amydiehl

I guess maybe we …. haven’t? … eliminated sexism from human conversation?

Show threadwhitneyFeb 16, 2023
@anne_twain @amydiehl there’s a really interesting perspective on this “mirror” idea in Brian Christian’s “The Alignment Problem”.
Show threadCaylaFeb 11, 2023
@amydiehl
To be fair to Chat GPT, if someone was describing someone else as 'bubbly' I expect the next thing they say to be sexist.
Show threadSuzanne PaleyFeb 11, 2023
@Kaminara @amydiehl Yes, it occurs to me I've rarely (never?) heard the adjective bubbly applied to a man.
Show threadKLBFeb 11, 2023
@amydiehl @Kaminara or “sassy”
Show threadMattFeb 11, 2023
@KLB @amydiehl @Kaminara right, I’ve only seen that used to describe women, gay men, and children. Never a straight man
Show threadKLBFeb 12, 2023
@MattFerrel @Kaminara @amydiehl it’s a weird thing to realize, isn’t it?
Show threadEric ☀️ for Kamala HarrisFeb 11, 2023
@amydiehl lawyers are going to have a field day.
Show threadKyozouFeb 11, 2023
@amydiehl Garbage in, garbage out
Show threadMateLordFeb 11, 2023
@amydiehl only if there was a way to write feedback based on performance not gender 🧐
Show threadKevin Karhan Feb 11, 2023

@amydiehl @aral #AI is alyways as #biased as it's #dataset used to #train it.

And it shows...

Show threadFabienFeb 11, 2023
Thing is people expect machine learning “AI” to be superior to human intelligence, when it’s actually about taking the average human answer. So it’s not even the best of human intelligence.
Show threadCassandrichFeb 11, 2023
@fabienmarry And it's not even intelligence. If it were applying language to interact with the world, test hypotheses, incorporate the results thereof back into the model, it would kinda be intelligence. But it's just 🤮 back fancy interpolations of things it ingested.
Show threadMs AB ☕Feb 11, 2023
@amydiehl Amazed, I tell you. I am amazed. 🤔
Show threadErik JonkerFeb 11, 2023
@amydiehl
No surprise ofcourse because that's the bias in the data trained on, it could be easy to correct 🤔
Show threadFlaming JuneFeb 11, 2023
@amydiehl They can only work with what - despite claims and spin - is already there.
Show threadAndromeda YeltonFeb 11, 2023
@amydiehl however could they have foreseen this, aside from “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings”, Bolukbasi et al, twenty-freaking-sixteen
Show threadJim WrinkleFeb 11, 2023

@amydiehl

A plausible liar that echoes the biases of its training corpus!

Show threadTushar ChauhanFeb 11, 2023
@amydiehl Nothing that can't be fixed by some more curated training and exploitation of the global South. We can all then pretend that it's magic.
https://mastodon.mit.edu/@tchauhan/109803452241159239
Tushar Chauhan (@[email protected])

Obviously. Underpaying and psychologically abusing Kenyan labour pays big dividends for investors in the global north. A potential use case: https://cyberplace.social/@GossiTheDog/109802533208388645 #AI #ML #ChatGPT #OpenAI #Kenya #Colombia #Law #BadDecisions #Society

mastodon.mit.edu
Show threadDavid Martínez MartíFeb 11, 2023

@amydiehl is this unexpected for anyone? As a model that has been trained over human texts, the normal results is that it learns the biases of those. So if the majority of the text comes from white males, the output of the AI will look like the mean of those.

Removing or rebalancing those biases is really hard and I don't think we have solved the problem yet.

When using these kind of tools we have to be mindful of their limitations

Show threadAndreas KFeb 11, 2023
@amydiehl And how is that the fault of ChatGPT, it basically reflects our society back through all the training data OpenAI s̶t̶o̶l̶e̶ downloaded from the Internet to train it.
Show threadmpmaddogFeb 11, 2023
@amydiehl gender neutral pronouns? in need id say...
Show threadKatLSFeb 11, 2023
@amydiehl hasnt the porn industry been using chat bots for years? They’ve probably got a lot of data. Not necessarily out bringing peoples best selves but I could be bringing my anti porn bias here. Maybe there’s lots to learn. From what I’ve heard you can find them on tumblr.
Show threadSocar MylesFeb 11, 2023
@KatLS @amydiehl Not just the porn industry--premium-rate SMS chat services in the UK were using them in the early '00s. Those services WERE used for sex chat, but certainly not exclusively, as they were marketed as ways to meet local singles. A LOT of what went on was just small talk. (A lot of these lines DID use human operators, paying them pennies per text to pretend to be British singles--but some used bots. They'd have LOADS of data.)
Show threadDry as toastFeb 11, 2023
@amydiehl all this does is allow AI to generate copy that media companies can later disavow when it turns out to be misogynistic, racist, completely wrong or profoundly dumb. ChapGPT is a get out of jail free card for bigots.
Show threadPatrick 🌎Feb 11, 2023
@amydiehl so, off to a great start then 🙄
Show threadDan JonesFeb 11, 2023

Machine learning models are the perfect example of "garbage in, garbage out", but the garbage in is just our culture.

@amydiehl

Show threadYusuf ToropovFeb 11, 2023
@amydiehl (falls out of chair from magnitude of the surprise)
Show threadCelloMom On CarsFeb 11, 2023

@amydiehl

It's not surprising to parents that ChatGPT reflects what it sees. It's just like a child imitating the adults' behaviour.

It needs a firm guidance counselor - and of course that means we have to argue what principles should move the guidance counselor - but obviously if left to its own devices ChatGPT will act as a mirror of who we are collectively. That's probably its most useful function right now, – if we care to see.

Show threadAlkaFeb 11, 2023
@amydiehl sounds like you should go on Twitter and cancel chatGPT.
Show threadMinameH❇️Feb 11, 2023
@amydiehl shouldn't be surprising...
Show threadMatthijsFeb 11, 2023
@amydiehl so it is primarily a great source of evidence for societal injustice and prejudices as it feeds of human texts. At least one silver lining
Show threadLG MediaFeb 11, 2023
@amydiehl the reason why an increasing number of people do not want to be gender noted at all.
Show threadsolientFeb 11, 2023
@amydiehl seems like the tech bros flat out ignored all evidence that their critical thinking replacement has an obvious patriarchal bias.
Show threadNoah ArneyFeb 11, 2023

@amydiehl when I was having it make cover letters for my students to critique it was overconfident by default, unless it thought it was writing as a woman, then it undersold its skills.

Great to see how we're baking in the exact types of misogyny we want to get rid of :/

Show threadBlake C. StaceyFeb 11, 2023
@amydiehl Bias As A Service(TM)
Show threadAmy Diehl, Ph.D.Feb 11, 2023
@bstacey BaaS LOL!
Show threadLouis EvansFeb 12, 2023
@amydiehl every instance of algorithmic bias is a confession of a bias latent in society
Show threadKyle 🎙 🎶Feb 12, 2023
It simply reflects the views of its creator Microsoft. All this "OpenAI" bull is just that, bull. It's not hard to find who created it, and thereby deduce why they created it.
Show threadyamainu 🏳️‍🌈Feb 12, 2023
@amydiehl
Teach your A.I. well
Internet hell will slowly go by
And feed them on your memes
The folks they troll are what you'll know by
Show threadLouise Auerhahn 🏳️‍🌈Feb 13, 2023
@amydiehl That's problematic but I think the bigger problem here is the idea that performance reviews could or should be auto-generated. The label "AI" gives the impression that ChatGPT actually is capable of understanding topics and making judgments, when in reality it does neither.
Show threadhenkFeb 16, 2023
@amydiehl It's not the algorithm to blame, it is the data which is fed to the algorithm. this data is produced by real humans
Show threadNot Just Any WilltorFeb 16, 2023
@amydiehl Whoa... I'd say ChatGPT is stuck in the past, but it was probably trained on pretty recent input.
Show threadAlien software, human hardwareFeb 18, 2023
@amydiehl yes.
Because it's not an "ai" , and also does not write anything.
It is a statistics package that, instead of printing a nice graph of the frequency of words in it's input dataset, just picks the words with the highest number and strings them together.