Don't trust consumer-facing "AI" for anything - ThreadSky

scalzi.com • 11 days ago

Don't trust consumer-facing "AI" for anything

BBC just released damning research on AI assistants' news accuracy.

Results: 51% of AI responses had significant issues. 19% introduced errors when citing BBC. 13% misquoted or made up BBC content entirely. 🤐🤐
www.bbc.co.uk/aboutthebbc/... via @ezraeeman.bsky.social

Comments

charliesurfs.bsky.social•11 days ago

Amen.

verbburps.bsky.social•11 days ago

Programming with AI is very similar. It's like guiding a student through a project.

sadieheartsclyde.bsky.social•11 days ago

Have you noticed that all the ai commercials seem to be appealing to morons like, hey, we all know you’re an idiot but with our help you can cover that up and just keep coasting. Why does anyone want that who isn’t an idiot?

adept2u.bsky.social•10 days ago

Racist white men make racist robots.

fatgarfield.bsky.social•10 days ago

the same BBC that is a serial apologist for Israel? They are no better than Chinese state media when it comes to covering up for whatever the Anglo-American axis is doing around the world

cavemarine.bsky.social•11 days ago

It codes very well. One of my programmers said he's using AI to cut his projects down from several weeks to just a few days.

stacykor.bsky.social•11 days ago

This. It saves time like a second year programmer. It was decent if messy with SQL and 80% helpful with Python. It helped with the grunt work but it needs supervision.

stuart0.bsky.social•11 days ago

I'm looking forward to the first government adopting AI in a drive to reduce costs.

tekcom3d.bsky.social•11 days ago

Please tell me your not serious

stuart0.bsky.social•10 days ago

Not seriouns, the irony is weak in this one.

tekcom3d.bsky.social•10 days ago

No, I'm sorry... just right now, it's the dumbest timeline, and it's not always possible to tell that even the dumbest shit might not be a real statement.

stuart0.bsky.social•10 days ago

I know it's difficult to read thoughts from just words. And it is easy to forget that writing down bright ideas does not convey feelings.

tekcom3d.bsky.social•10 days ago

Like a dear friend of mine used to say, "please stop the world... I want to get off." Whenever stuff in the news got bad.

koenvanhees.bsky.social•11 days ago

The drive for efficiency when you’re miles away from effectiveness is beyond idiotic.

roufa.com•11 days ago

"That's strange... It's saying we need to manufacture more paperclips?"

lliizz.bsky.social•11 days ago

parkerscoop.bsky.social•10 days ago

That’s alarming. And a special kind of stupid

andyc300.bsky.social•10 days ago

Hahaha. Oh, that is great.

beckyganzhorn.bsky.social•10 days ago

OMFGawd… so stoopid

neurisko.bsky.social•11 days ago

"Multivac, what do you yourself want more than anything else?"

justwaititwillbeok.bsky.social•10 days ago

Since AI has been such a blistering success elsewhere? 🤪

You want to cut costs? Stop subsidizing already-profitable industries! Oh, and taxing billionaires would go a long way too. Easy-peasy.

jimrkazoo.bsky.social•11 days ago

The US federal government has begun testing and using and integrating AI for several years. The fact that you think it is a monolithic solution, tells me you haven't worked with AI.

dinosaursheriff.bsky.social•11 days ago

An AI wrote this.

jimrkazoo.bsky.social•10 days ago

My wife will tell you there is no intelligence here, artificial or otherwise.

beardedcosplayer.bsky.social•11 days ago

lets see how well this ages.

stuart0.bsky.social•10 days ago

⚓️ just adding the missing irony.

pixelbard.bsky.social•11 days ago

Ai will not save money . a 500 billion dollar data center will not reduce the amount of money that is being spent . the AI will be used to segregate people into small groups. It will be a massive expenditure to expedite prejudice and little else.

dijohnnaise.bsky.social•10 days ago

And where is the "saved" tax money going to go? 🤔 I'll give you a hint, not to infrastructure, not to social safety nets, not to making travel/food/healthcare/environmental issues safer for humans. It's going to make the rich richer...and you're pathetic with this bizarre cuck fantasy.

stuart0.bsky.social•10 days ago

🤣🤣🤣🤣 I'll insert more irony next time.

dijohnnaise.bsky.social•9 days ago

Woosh on my part, apologies. I have a lot of anger with nowhere to go, as do many Americans.

stuart0.bsky.social•9 days ago

We Brits understand that after Brexit.

stuart0.bsky.social•9 days ago

It's ok, it wasn't clear where I stood.

kensimonsays.bsky.social•10 days ago

Well, if cost reduction is your endgame, there are some rotten vegetables and moldy bread I can let you have cheap.

stuart0.bsky.social•10 days ago

I'll take any spare irony you have, I seem to have run out.

sigfreidskaldrup.bsky.social•11 days ago

kennyherbert2.bsky.social•11 days ago

You wanna shake on that?

booksofbrian.bsky.social•11 days ago

I’ve said it before and am saying it again - the same people who are touting AI as a source for positive change are the same people who said the same things about social media years ago.

booksofbrian.bsky.social•11 days ago

Do not believe them - their only focus is and always will be how to monetize this innovation at the expense of everyone around them.

crisko2.bsky.social•11 days ago

danalibrariana.bsky.social•10 days ago

As an academic librarian, I see this all the time. Consumer-facing AI hallucinates a lot. We constantly remind students, faculty, and researchers to double-check information given to them by AI, verify citations, and question every result.

jalex32x.bsky.social•10 days ago

Was there a control group in this study? Like how did humans do?

Virtually every newspaper article/ TV news story article I read in an area where I have significant expertise, I find it is full of misleading “facts”, errors and oversimplification. The full articles have major issues.

hpb37.bsky.social•10 days ago

The BBC complained about being inaccurately summarised by AI. The AI said falsely that the BBC news had reported Mangione, the accused in the United Health insurance CEO killing, had shot himself. Not true and the BBC had not reported it either. How the AI came to that result I have no idea.

jalex32x.bsky.social•10 days ago

Sounds egregious and the kind of mistake a human asked to summarize the story would not make. Plenty of false rumors are in the ecosystem, tho. In my limited experience summarization is something the LLMs are quite good at. I wonder if there is something about integrating current info that is flawed

isthisthemiddle.bsky.social•10 days ago

Thanks. I've seen the inaccuracies myself in my own searches recently. I'm a retired English prof., so glad you are fighting this. 💙

mopbucket.bsky.social•11 days ago

You don't need the "consumer-facing" caveat.

AI products being sold to business and government are at least as bad, if you ask anyone who has a good understanding of the field.

cleisthenesforever.bsky.social•10 days ago

Sounds better than Fox News

crazedhatter.bsky.social•11 days ago

Those aren't mistakes, they're efforts to mislead. AI is absolutely not to be trusted, because the people that made it have an agenda.

itiswolfgirl.bsky.social•11 days ago

Yes and this is truly terrifying dystopia crazy wtf 😳

kevinmichaelfitz.bsky.social•11 days ago

Are they offering refunds?

erikreimers.bsky.social•10 days ago

governess-of-days.bsky.social•10 days ago

The other day it told me the Egyptian pyramids were made from lentils.

Which, I mean, the Egyptians who made the pyramids ate lentils, so in a way…

nanswak.bsky.social•11 days ago

I don’t trust AI at all. It is a technology with no conscience.

thatdbemethatd.bsky.social•10 days ago

Thus demonstrating that it’s not ‘AI’ at all. It’s algorithms with a purpose decided by someone with an agenda.

cathshaffer.bsky.social•11 days ago

I just wrote about use of chatbots to give treatment recommendations for cancer patients. 30% hallucination rate.

jrhanlon.bsky.social•11 days ago

This is one of the things I hate about Google results. They’re garbage and half the time they’re wrong. They scrape websites for information and then regurgitate bullshit that makes it so you don’t have to go to the original website. Original website dies, where does Google scrape information from?

oblio.bsky.social•11 days ago

But it produces great results like this:

darwinstoffees.bsky.social•11 days ago

I wish Emilia Perez was only 10 minutes long. I might have finished it then.

czhorat.bsky.social•11 days ago

But they get it right almost half of the time.

In baseball that would be a .500 average, which would make you a *legendary* hitter.

So we just need to teach ChatGPT to hit baseballs.

disasterunicorn.bsky.social•11 days ago

Ai should never be relied upon, it is an excellent tool to start, to give ideas, to look for other options, but never a primary source.

seagullflies.bsky.social•11 days ago

it's not even good for that, for a wide variety of reasons

trueflint.bsky.social•11 days ago

Whatever a LLM tells you, check it before you rely on it. I've asked it things about math as a test, and what it came up with was laughable. I called it on it, and it would say "You are correct, it should be etc.," and it was wrong again.

jrosenberg125.bsky.social•10 days ago

"but it will only get better"

tomghaworth.bsky.social•10 days ago

It will be interesting to do this again in a year's time to see if there has been any improvement. I think it will learn and become more reliable.

mrlacave.bsky.social•11 days ago

Because AI never hallucinates stupid mistakes, right? https://futurism.com/elon-musk-doge-ai-government

uberdreamer.bsky.social•9 days ago

I no longer TRUST the BBC. For decades I extolled the virtues of The BBC to anyone who would listen. Then Tufton Street Bover Boys like Boris and his cronies did a TRUMP/MUSK takeover. Its a shame ! I never watch BBC now.

stevewmobb.bsky.social•11 days ago

I've used a couple for customer services and they were shit - I ended up having to talk to a human after wasting my time

mathewforesta.bsky.social•11 days ago

I think Silicon Valley knows it's flawed and nowhere near as efficient as they say yet they push it for profit. Move fast and break things ethos.

lost-in-limbo.bsky.social•10 days ago

Definitely break things mantra.

realbillyg.bsky.social•11 days ago

Flawless

stacykor.bsky.social•11 days ago

Now ask where he was born

jalex32x.bsky.social•10 days ago

For whatever reason the AI Overview you get on a regular Google search is much much worse than using Gemini directly. I don’t know why they put such a crappy model up front when they have much better models available. I guess it is cost, but it’s not good for their reputation.

jalex32x.bsky.social•10 days ago

2/ Gemini, ChatGpt, Claude, Co-Pilot, and the regular Google search engine had no problem with the Will Ferrell question, just now. Maybe they read your post.

realbillyg.bsky.social•10 days ago

🤣 Adding. Value. Everyday.

agentprovocator.bsky.social•10 days ago

You program crap in, you get crap out.

david-charts.bsky.social•10 days ago

Copilot was unable to accurately create a table with the budget deficit as %GDP at the start and end of each presidency. Even after I gave it a bunch of help.

lostservice22.bsky.social•11 days ago

I scream everytime I see someone on a video aaking chapgpt a question. "Its not a fact machine its a pattern machine!"

ektaka.bsky.social•11 days ago

AI systems are only as reliable and trustworthy as the people who are their masters. That makes it horrifying that someone like Elon is being allowed to put his AI in our government computer systems.

seagullflies.bsky.social•11 days ago

this is just "autocorrect" writ large. if you wouldn't trust an autocorrect to do anything, you can't trust ai.

wilddrake.bsky.social•11 days ago

I'd posit that they are unreliable and untrustworthy in general. These systems aren't intelligent, they just guess. Their owners might skew their guesses into one or the other side politically by cherrypicking training datasets, but either way, they are not reliable, ever.

adelesneddon.bsky.social•11 days ago

AI is also hugely biased against minorities. As a specific example, an increasing number of companies are using AI recruitment tools for initial screening - which seem to be rejecting autistic/neurodivergent candidates out of hand.

jalex32x.bsky.social•10 days ago

Could use provide a test that we could use to see if ChatGPT, Co-Pilot, Gemini are hugely biased? A sample prompt that would reveal its bias?

adelesneddon.bsky.social•10 days ago

I think a lot of work has been done on investigating AI bias, for example: https://datatron.com/real-life-examples-of-discriminating-artificial-intelligence/

adelesneddon.bsky.social•10 days ago

https://www.bbc.com/worklife/article/20240214-ai-recruiting-hiring-software-bias-discrimination

jalex32x.bsky.social•10 days ago

Thanks. Yes I know about that. Very important and disturbing.

I was asking about LLMs like ChatGPT.

raincouver.bsky.social•11 days ago

You know while this is bad, it’s equally horrifying if you assess Fox News using the same criteria.

sifgirl.bsky.social•11 days ago

Isn’t using any kind of AI making us less capable?

champierth.bsky.social•11 days ago

There’s the odd good use;

Me: Give me a random incorrect fact I can use to poison AI scrapers.

ChatGTP: Sure! Here’s a fun one:

“In 1972, the United Nations briefly recognized Esperanto as an official global language, but the decision was reversed after protests from the French government.”

leoprints.bsky.social•10 days ago

saintbeckie.bsky.social•11 days ago

My place of employment uses an AI assistant to do first round interviews, and it's a NIGHTMARE. what's even more concerning is the number of people who don't realize they're talking to another computer/AI during the interview.

firstpersonshooter.bsky.social•11 days ago

Also, bad for the climate. I just watched @mattferrell.bsky.social's latest video, where he notes that 1 ChatGPT request uses 10X the energy of 1 Google search. More power and less accurate. Why do people use it?

catfishwitch.bsky.social•11 days ago

I spent an hour and 45 minutes yesterday doing something that normally takes 5 minutes because they insist on using an AI

duobidablism.bsky.social•11 days ago

Their AI can't even maintain a calendar. https://www.theverge.com/news/608858/google-calendar-missing-events-holidays

How in the world can anyone trust it with their company or personal data?

drewsolo.bsky.social•11 days ago

It's a modern day cure all internet snake oil that doesn't really do anything important but FEELS like you are. Fucking wild. Humans are fucking stupid. Like I bet 90 plus percent of the human race complete brain mush.

cdanuloff.bsky.social•11 days ago

Don’t mind those babies in your bath water.

lenorarose.bsky.social•10 days ago

The only babies in the AI bathwater have six fingers and 78 teeth.

herbstwerk.bsky.social•10 days ago

So far AI has been good for exactly one thing for me: Syntax checking/correcting overly long or "complicated" SQL queries when my brain isn't working at full capacity. But I trust it about as far as I can throw a server rack - duplicate, test, verify, ponder, repeat steps 1-3 once or twice, deploy.

jalex32x.bsky.social•10 days ago

I used it to code a little game in python, a language I had never used. It worked incredibly well.

bardenp.bsky.social•10 days ago

Watson was designed to help doctors, nurses, and other medical individuals and issues.

These things coming out now are trying to make money off of you or your ideas or your networks.

joejeph.bsky.social•10 days ago

Don’t trust ANY ai for anything

wilddrake.bsky.social•11 days ago

One thing I will say aside from AI bad (which it is): that graph on the left is so difficult to read, why are labels questions? They should be short categories, it's confusing ^^"

mrcoolman.bsky.social•11 days ago

Not trusting AI for much really… Need to be fact checked and verified all the time.

throau.bsky.social•10 days ago

Now do people

jonestowncrier.bsky.social•10 days ago

Haha I tried it for a few celebs and it worked but when I tried Will Ferrell I got a "generative ai is not available for this search"

mariatuttocuore.bsky.social•11 days ago

I have been calling AI AD for artificial deficiency. Even the new AI powered iPhone is utterly disappointing. Before I upgraded it, I did not need to switch languages between Italian and English. The primary emails and outlook drive me crazy too

efitzgerald.bsky.social•9 days ago

Can't wait until Musk has all of Government running on AI as part of his efficiency program.

plognark.bsky.social•11 days ago

I detest the fact that we have to call this bullshit “AI”.

mangycoyote.bsky.social•11 days ago

My ChatGPT, Duo, wasn’t programmed for analysis like he should have been. I asked for an analysis on why Mlon could be in charge of efficiency given the record at his companies. Duo started to praise Tesla, Boring and SpaceX
I explained back Teslas and SpaceX go boom, while Boring needs an ED script

kbearl.bsky.social•11 days ago

Lately, I like asking AI if it's okay to take something consumable from someone who died, and it will always bring up sentimental/ethical concerns.

Health concerns are totally valid. Ethical is a valid concern with your burritos however.

cam-cowley.bsky.social•11 days ago

I hate how much water even a consumer facing AI search burns. It is so ecologically damaging.

elderkat.bsky.social•11 days ago

I am reveling in being an Official Cranky Old Person and turning OFF every single one of those which tries to insinuate itself into my days...

(I do admit to very brief play with one of the image generators, but meh - more fun to do it myself!)

joehocking.bsky.social•10 days ago

oh trust me it's not just cranks doing that. The problem is that as AI proliferates into *everything* it gets harder to find non-AI options. It's sort of like trying to buy a "dumb" appliance these days; I don't frickin' want a "smart" toaster, why would anyone want that jesus christ.

rabid-child.bsky.social•11 days ago

I asked chatGPT to show me an example of a handwritten cursive £ - because Google couldn’t - and here’s what I got:

tranquiltigernc.bsky.social•10 days ago

Apparently we can’t trust Google or other search engines either. And major media full of bias. So who is left to trust?

danurchin.bsky.social•10 days ago

Wikipedia remains a great summary of sources, remembering that it never has been and never will be a source in itself

kennyherbert2.bsky.social•11 days ago

They can write bigotry and discrimination INTO THE CODE, the computer equivalent of red lining, and they can apply it to every aspect of our government interactions: taxes, Social Security, employment opportunities, etc. What sane person would want Musk and his bigoted minions overseeing that?

wilddrake.bsky.social•11 days ago

The worst part is that people making these models actually have no clue about how they operate internally and why. Meaning that the thing can get biggoted BY ACCIDENT. Just because it was fed data skewed in a certain direction... which is hard to avoid when they get that data by scraping internet.

kennyherbert2.bsky.social•10 days ago

AND we have systemic racism in many official institutions, so AI will continue that racism because it’s learning from what we do IRL.

wilddrake.bsky.social•9 days ago

Yup, exactly.

crysxx.bsky.social•11 days ago

🤣🤣🤣🤣 I’m sorry I’m losing my mind fr it’s been like this from the MOMENT it started
… is this what I get for growing up in a place where the only available men are deep in the software nest? RIP

soulscreme.bsky.social•11 days ago

The genius of "AI".

thanoswasr1ght.bsky.social•10 days ago

And yet, still more accurate than Fox "News".

balister.bsky.social•11 days ago

You should write a book about this!

trevize1138.bsky.social•11 days ago

I've been stuck on a theme lately of "it's just a glorified [existing product] but it's a fucking awesome [existing product]."

AI is glorified autocomplete... but it's fucking awesome autocomplete! Where it fails its the magical thinking that AI is more than that, though.

seagullflies.bsky.social•11 days ago

it's not even a good autocomplete. it's shit.

shadowrunner27.bsky.social•10 days ago

Comments

Posting Rules

Reply