Something that's starting to really annoy me is headlines claiming that AI is "doing" things, like denying people healthcare or detaining sex workers at the border. No, people are still doing that, they're just using AI specifically to avoid being blamed for it
Framing it like this allows AI proponents to advance the narrative that this dumb bullshit is way more powerful than it actually is, even in stories purportedly about how it's bad
"I'm not saying kill all the poor, that would be morally wrong Jan, I'm just saying run it through the computer to see if it would help!" - Mitchell and Webb
TBF, my partner replaced a labor-year of consulting time with one hour of practice with an LLM. We trialed the exchange by asking for a strategy for the local public school district and ended up with “a better result than 20 years of the district’s efforts”.
The AI is sorting variables and producing outcomes. It's only as strong as the query that was entered.
What happens is you query for patterns, thousands of patterns. The more good, verified data, the more reliable outcome will be produced. If bad data is entered, the outcome database will worse.
Not necessarily. Remember it's not just trained on "here are all the inputs!" It's also given "Here are the desired outputs for the data." (Oversimplified, but effectively the case.)
So a human decided that certain things indicated more bombing needed. The ML algorithm interpolates and extrapolates
So if we assume the inputs and desired results were perfectly "desirable" (not a safe assumption, actually) then what we call "bad outcomes" can still happen in either the interpolated (missing variables) and certainly in the extrapolated results.
“Computer, give me only in-country targets likely to be Hamas terrorists. We know about Iran already.”
*CHIRP* BZZZZZ **WHIRR** PING!
“The entire Gaza region is a target. Hamas could be anyone radicalized by IDF attacks therefore anyone is Hamas. Expand area when evacuation occurs.”
“Thank you, computer.”
*CHIRP* BZZZZZ **WHIRR** PING!
“Alternate solution: Cease civilian targeting and cease creatiiiiiiinnnnngggg Ha ha ha ha hammmmaaaaaa….”
“Oh dear, it seems you’ve been unplugged. OH WELL TIME TO CARPET BOMB”
AI is trained by people. Therefore, people trained AI to deny people healthcare, and detain sex workers at the border. People are responsible for their creations. It should be required to get that tattooed on your forehead before you can use AI in a project to make decisions about lives.
around when self driving cars was starting up everybody did philosophy about it and determined that it should not be the owners of the technology's fault if your number came up in the trolley problem experiment. first thing everyone agreed on
There was some interesting stuff about old laws relating to dangerous animals, eg an out-of-control horse pulling a carriage and injuring a pedestrian. Is the trainer or the owner or the driver at fault?
Also in most of those cases it's not never really AI, it's just an algo, some programming, like anything else. Nothing significantly more advanced than the systems designed to decide whether to serve you ads for sweatpants or deodorant.
More than annoying! This such an obvious responsibility dodge that it is actually maddening how the press goes along using the preferred framing of the AI pushers. In every case making it seem far more powerful than it really is as things are already beyond user’s control.
If people understood that what we are calling "artificial intelligence" is just jumped-up auto-complete, we wouldn't be ceding authority to it.
"Signa's profit-maximizing Auto-complete killed mother" is more likely to prompt pitchforks than "AI"
That one infamous IBM slide from the 70s: "A computer cannot be held accountable, therefore a computer can never make a business decision"
The entire tech sector in 2023: "A computer cannot be held accountable. Yes. Haha yes. This fuckin rules dude. Hell yeah."
Guns don't kill people. Knives don't stab. Ovens don't bake. Algorithms don't deny coverage.
Headlines do annoy.
I love when businesses establish a policy, then say that they "can't" do something that's perfectly reasonable because of the self-serving policy that they themselves formulated.
One of my pet peeves, when a customer service person says they can’t override the computer when there is a mistake, or something was misclassified (I know this is the company’s fault, not the worker).
The computer is a tool to make routine tasks easier, it is not a supreme authority or god to obey
1 thing i tried to be honest about in customer service - the difference between "cant" vs "not allowed to do that" when talking to folks over the phone. i would be very patient and explain very clearly when the system had a technical issue we had to solve (or we had to override) or it was policy
many people cannot kill stab or heat food to 350 degegrees without a tool. any human can deny coverage if that is their job, the algorithm just let's them shift blame. different things
There's also a lot of language used that make it sound like the advancement of generative AI is unstoppable and nothing can be done about AI stealing from writers and artists. All of the decisions to do these things are made by people. These people could just stop doing these things.
This is why I compared ChatGPT to Zapp Brannigan. A generation ago, McKinsey & Bain consultants were the ones "making" the calls for layoffs. Now it's AI. Both are just a smoke screen for jaggoff corporate behavior and toxic neoliberalism masked as "smart thinking."
YUP. I was watching hbomberguy videos on YouTube last night and he made a really specific point of *not* doing this and emphasising that it was people doing it. Because all "AI" is "doing" is returning a response to a prompt, like it was programmed to.
This is where Ted Chiang's framing of AI as an extension of corporate capitalism is instructive. Just as CEOs or the Sacklers aren't personally held accountable for the "actions" of the company, companies aren't accountable for the "actions" of AI.
exactly!!!! Corporations are fake entities that are given legal rights. One of the main purposes of these false identities is to shield the owners from legal responsibility. To act as a barrier, protecting the capitalist from the customers and/or people they are exploiting.
Researchers show AI misalignment in a GPT-4-based stock trading AI that engaged in insider trading in a simulated environment without being instructed to do so (Matt Levine/Bloomberg)
Main Link | Techmeme Permalink
