Loukas (They/Them) 🏳️⚧️In the same way as "the cloud is just someone else's computer" - "AI decisions are just someone else's prejudices".
Tim Ward ⭐🇪🇺🔶 #FBPE@Loukas That's what labelling is about - AI decisions (eg about whether a credit card transaction looks dodgy) are based on the "prejudice" of actual factual determinations. You might regard the victim of a fraud reporting it as a fraud to be prejudiced against fraud, but that's about as far as you can go with "prejudice".
Woodrow Douglass@TimWardCam @Loukas Almost -- AI decisions are based on the "prejudice" of determinations the labelers made, whether they're factual or not.
@bigfishrunning @Loukas "Prejudice" is about "pre-judging" something that hasn't happened yet. Labelling is about recording something that has happened.
@TimWardCam @Loukas does that mean that every labeller is an unbiased observer, and only labels things factually? Seems kind of incredible to me.
@bigfishrunning @Loukas I was trained as a mathematician, and it only takes one counter-example to disprove "every", so I never (ha ha!) claim "every".
@bigfishrunning @Loukas But fraud detection is about stats, not about "every". You're trying to do two things:
(a) be good enough at detecting actual fraud to get regulatory approval that you're trying hard enough
(b) keep false positives, and thus the cost of call centres and pissed-off customers, down.
So target driven in both directions.
@TimWardCam @Loukas fraud detection is only a very small part of the pie; the original thesis is that the output of a given AI system carries the Biases (perhaps overzealously using the word "Prejudices") of the input data and the imperfect-by-definition human labelers; I think that thesis holds true.
Nicola Romanò@bigfishrunning @TimWardCam @Loukas this reminded me of this excellent (and scary) piece on welfare fraud risk calculation in Rotterdam https://www.wired.com/story/welfare-state-algorithms/
Bornach@bigfishrunning @TimWardCam @Loukas
I love how Google DeepMind tried to compensate for those biases and ended up overdoing it in their Gemini model
https://youtu.be/Fr6Teh_ox-8
I love how Google DeepMind tried to compensate for those biases and ended up overdoing it in their Gemini model
https://youtu.be/Fr6Teh_ox-8