Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

AI Incident Database

incidentdatabase.ai

Post reply on HN
Print view View original

71–80 of 115 posts

Re: AI Incident Database

#71

Post by varunjain99 »

Imnimo wrote: post #53

This is interesting, but the presence of so many incidents that ultimately have nothing to do with AI is disappointing. Consider this incident: https://incidentdatabase.ai/cite/73/#r1325 >Description: Through a crowdsourcing social media campaign in 2016, several journalists and researchers demonstrated that augmented reality locations in the popular smartphone game Pokemon Go were more likely to be in white neighbor…

We need better categorization of what to blame - is it the data collection mechanism, the training of the AI, the use of the AI in a broader software system.

Even without a neural net based model, it's possible we run into many of these incidents...

Re: AI Incident Database

#72

Post by to11mtm »

Imnimo wrote: post #64
n4r9 wrote: post #59

Earlier quoted context omitted.

Based on that quote it sounds like the crowdsourcing was used to analyse the locations in order to show the bias. The claim is that AI was used to determine the locations on the first place. If I recall rightly the Pokémon Go locations were derived in part from an earlier (and far better) game called Ingress. Players could submit candidate locations, but it's possible that this was mostly ignored, and/or some form of…

It's possible, but the longer description makes it sound like the source of the issue was that the locations were crowd-sourced from a biased sample: >The game creator Niantic Labs revealed that the Pokemon Go map was derived from a previous augmented reality game, Ingress. Ingress crowdsourced its map from users, who tended to be young, male, white, and English-speaking. Similarly, from the attached article: >"It tu…

> It could be there is some other AI-related problem for Pokemon Go hidden in one of the other articles (they all seem to be about unrelated things - one is about Jynx, one is about the experiences of a black person playing the game, another is speculating the Pokemon Go is a CIA operation...)

I'm still not sure if some of those stories were viral marketing or just a bizarre reflection of our zeitgeist in the black mirror.[0]

[0] - Exampled by: https://www.dailymail.co.uk/news/article-3734361/Woman-tells...

Re: AI Incident Database

#73

Post by RowanH »

Imnimo wrote: post #53

This is interesting, but the presence of so many incidents that ultimately have nothing to do with AI is disappointing. Consider this incident: https://incidentdatabase.ai/cite/73/#r1325 >Description: Through a crowdsourcing social media campaign in 2016, several journalists and researchers demonstrated that augmented reality locations in the popular smartphone game Pokemon Go were more likely to be in white neighbor…

> https://incidentdatabase.ai/cite/482/

> Vanderbilt University's Office of Equity, Diversity and Inclusion used ChatGPT to write an email addressing student body about the 2023 Michigan State University shooting, which was condemned as "impersonal" and "lacking empathy".

The bar is pretty low for what's considered an "incident". Writing a cold email eh?

Re: AI Incident Database

#74

Post by prettyStandard »

andrewmutz wrote: post #38
prettyStandard wrote: post #36

Earlier quoted context omitted.

Genetically Modified Organisms? I was under the impression these were very safe. Can you change my mind?

Sorry for the subtle sarcasm, they are safe and that's exactly my point. 20 years ago everyone was scared of GMOs and begging for regulation. Many european nations banned them in response. Other places insisted on labeling them so consumers could avoid the danger. We now know that they are safe and these efforts were counterproductive.

lol thanks. At first I was warry about them mainly because I hadn't taken the time to think too critically about. A friend walked me through the simple idea that your stomach is going to tear most things to shreads, so, as long as you aren't taking poison basically, you should be fine.

Re: AI Incident Database

#75

Post by wilg »

RowanH wrote: post #73
Imnimo wrote: post #53

This is interesting, but the presence of so many incidents that ultimately have nothing to do with AI is disappointing. Consider this incident: https://incidentdatabase.ai/cite/73/#r1325 >Description: Through a crowdsourcing social media campaign in 2016, several journalists and researchers demonstrated that augmented reality locations in the popular smartphone game Pokemon Go were more likely to be in white neighbor…

> https://incidentdatabase.ai/cite/482/ > Vanderbilt University's Office of Equity, Diversity and Inclusion used ChatGPT to write an email addressing student body about the 2023 Michigan State University shooting, which was condemned as "impersonal" and "lacking empathy". The bar is pretty low for what's considered an "incident". Writing a cold email eh?

Also sounds like the kind of criticism a human-written email would receive. Maybe it did a great job!

Re: AI Incident Database

#76

Post by miohtama »

RowanH wrote: post #73
Imnimo wrote: post #53

This is interesting, but the presence of so many incidents that ultimately have nothing to do with AI is disappointing. Consider this incident: https://incidentdatabase.ai/cite/73/#r1325 >Description: Through a crowdsourcing social media campaign in 2016, several journalists and researchers demonstrated that augmented reality locations in the popular smartphone game Pokemon Go were more likely to be in white neighbor…

> https://incidentdatabase.ai/cite/482/ > Vanderbilt University's Office of Equity, Diversity and Inclusion used ChatGPT to write an email addressing student body about the 2023 Michigan State University shooting, which was condemned as "impersonal" and "lacking empathy". The bar is pretty low for what's considered an "incident". Writing a cold email eh?

Fearmongering sells well and makes good clickbaity headlines.

Re: AI Incident Database

#77

Post by miohtama »

RowanH wrote: post #73
Imnimo wrote: post #53

This is interesting, but the presence of so many incidents that ultimately have nothing to do with AI is disappointing. Consider this incident: https://incidentdatabase.ai/cite/73/#r1325 >Description: Through a crowdsourcing social media campaign in 2016, several journalists and researchers demonstrated that augmented reality locations in the popular smartphone game Pokemon Go were more likely to be in white neighbor…

> https://incidentdatabase.ai/cite/482/ > Vanderbilt University's Office of Equity, Diversity and Inclusion used ChatGPT to write an email addressing student body about the 2023 Michigan State University shooting, which was condemned as "impersonal" and "lacking empathy". The bar is pretty low for what's considered an "incident". Writing a cold email eh?

Fearmongering sells well and makes good clickbaity headlines And scary statistics how Skynet is here, inbox next to you.

Re: AI Incident Database

#78

Post by SkyBelow »

lmcnulty wrote: post #66
nofum wrote: post #57

Yikes, they have a submission leaderboard. As if the expected signal to noise ratio of a system like this wasn't low enough.

I work on this project. Every submission has to be approved by an editor before it appears on the site or gets counted toward leaderboard stats. We're currently more constrained by lack of submissions than by low-quality submissions – it's not uncommon that someone will contact us asking about an incident that should be in the database but isn't yet.

Does that include the examples other posters have linked in this thread?

Re: AI Incident Database

#79

Post by smcgregor »

RowanH wrote: post #73
Imnimo wrote: post #53

This is interesting, but the presence of so many incidents that ultimately have nothing to do with AI is disappointing. Consider this incident: https://incidentdatabase.ai/cite/73/#r1325 >Description: Through a crowdsourcing social media campaign in 2016, several journalists and researchers demonstrated that augmented reality locations in the popular smartphone game Pokemon Go were more likely to be in white neighbor…

> https://incidentdatabase.ai/cite/482/ > Vanderbilt University's Office of Equity, Diversity and Inclusion used ChatGPT to write an email addressing student body about the 2023 Michigan State University shooting, which was condemned as "impersonal" and "lacking empathy". The bar is pretty low for what's considered an "incident". Writing a cold email eh?

After many protracted discussions among editors, we decided not to threshold the severity of the harm. An allegation of harm is enough. Not all incidents are created equal when it comes to severity and the taxonomy feature helps sort through these questions on top of the permalink the incident profiles provide.

Re: AI Incident Database

#80

Post by sebzim4500 »

JohnFen wrote: post #47
sebzim4500 wrote: post #41

Earlier quoted context omitted.

(1) What makes you think that the benchmarks will be blocked in robots.txt? E.g. if it is uploaded to arxiv or github I don't know how you would stop it from being indexed (2) There's probably loads of content blocked by robots.txt that people want to train on.

> if it is uploaded to arxiv or github I don't know how you would stop it from being indexed Ah, that makes sense. This specific thing still smells like baloney to me, though. > There's probably loads of content blocked by robots.txt that people want to train on. In what use case? If they want to train their own GPT, then it doesn't matter if robots.txt blocks it. If they want to have it included in the general data…

I don't understand your questions. The people releasing the benchmarks want to include a string that tells the scrapers not to include it in LLM training data. They want to have an easy way of filtering it out even if it gets uploaded to github/arxiv/reddit/etc. a million times. The scrapers will want to respect the tag because otherwise they won't be able to use this benchmark to evaluate the model.

It has nothing to do with robots.txt, which training data scrapers have no reason to respect.

Post reply on HN