#fuck ai #i’m so sick of people accepting every piece of propaganda at face value
See, I didn't mind so much when it was stuff the everyday person would never really effect. I can talk all I want about how garbage summarizing AIs are wrecking part of the job market, but no one I talk to or who follows me or hears my words is on the HR board of any company using these AIs, or programming any of these AIs. So I don't *care* about the average person's opinion of them.
Like yeah, you're misled about AI and that's bad, but you're also misled about the biology of agriculture, quantum physics, the engineering of bridges . . .some of those are going to affect your politics, chances of falling for scams, and other things in your life, and the AI you might other yourself use would be fun toys. cool!
But now the public has been recruited into seeing art as something that only exists for acts of non-experential consumption, as a mode of forwarding the current generation of AI, and uh . . . well . . . fuck. fuckity fuck fuck.
I've used ML before in ways that minimize it's downsides, like a friend of mine is a lawyer and often needs to redact huge amounts of documents, or decide which ones to release for discovery.
OCR for initial processing, and then just ctrl+f for keywords already does most of your work, but you can still be left with hundreds of gigabytes of text to sort through. Using summarizing AI to find everything that seems related within a low confidence value, (ie: it takes in too much for fear of losing anything) and then a framework for the reviewer that jumps to the next portion/next document which seems related still requires the reviewer to fully engage their brain and read the document, but spends their time much more efficiently. Less scanning documents, more reading.
In terms of deskilling, it shouldn't happen, because the AI lets through enough that's unrelated that it still needs thorough human revision. I can chop the workload to a tenth, (provided the documents to cover have a lot of unrelated content. If every document under review turns out to need to be released, the AI *shouldn't* reduce any workload), and still leave reviewers with a lot of chaff to cut. mostly, the workload still looks like hiring lots of people temporarily to sift through a metric fuckton of documents, for what is one of the most mind-numbing acts of labour possible.
(and this still needs the approval of the judge in each case to allow the use of the program here. Fortunately, he's an international technology lawyer, who deals with tech aware judges, so he often gets permission to use it.)
Considerations for deskilling, acknowledgement that the AI will often be wrong and honest accounting for that, and this is a situation where the workload is so massive that missing some documents to keep back or to submit is expected and common. I don't even need the AI to be more accurate than humans, (which is good, because it very isn't), it just needs to be first eyes and it'll still be helpful.
That's where AI can shine, thorny problems that don't already have a solution, nothing existing can do the job well, and you integrate the AI into the workflow of humans in a way that enhances their workflow and keeps them sharp.
But now we have AI being sold to legal firms and teams which doesn't meet any of the stringent requirements my friend needs to meet before he's allowed to use my software.
It promises to be a magic bullet that just about removes humans from the equation, removes the thinking they need to do, and promises higher accuracy. They try to change the entire workflow, not to suit the AI or the process, but to make it "less painful".
Take a fat guess about how they implemented it, how much care they actually care for accuracy, and uh, whether judges are amused at the results. They turn their reviewers into rubber stampers of the AIs work.
They're using better versions of the summarizing AI I had access to, (proprietary LLM's trained on server farms, vs. an open source implementation I trained on my PC), so they could get a much better result. But instead, they're opting for a downgrade in quality and getting people to buy into it, because they don't like hiring so many people to sift through documents, they don't want to use their brain on tasks that need it, and honestly, the plausible deniability of missing documents even more than before probably appeals to them.
This is what I mean by overextension and overpromising. An AI can't even do a drive thru job with any reliability yet, and you're trying to trust it to do legal work . . .? My implementation *didn't* trust it. And still helped. It cannot be a magic bullet or employee replacement for you, but there are niche tricky problems where what it *can* do is still helpful, and it has been sold *so fucking far* outside that radius.
Almost the entirety of what was developed during the 3rd AI boom, (mid 2000's to late 2010's), was such overpromised drek. Slap "ai" on something and you could sell it to just about anyone. "we use AI to plan your vacation" (travelshark), "we use AI to fix your city's traffic problems" (lyt), "we use AI to predict what movies you'll like" (Netflix) (heh, remember when they cared about how good their recommendations were?), and very very few of those problems actually benefitted from AI to any notable degree, (and some were just downright fraudulent in their horribleness, like lyt).
But AI is so bright and shiny, and everyone is always promising that it will get better over time! So people become satisfied with worse results than they had before, *which is fucking infuriating*. Even 20 years later now, there's a lot of places where AI *has* gotten better and it's still worse than the things we used to do before AI. So, uh, how many years do I need to wait? Another 20? 50? 100? Are the improvements going to keep decreasing in size as they have been, until it eventually resembles zeno's paradox?
Ultimately, the problem with AI is a marketing and business problem. But, that's the part that's modifying our society and our worldviews.
So again: fuckity fuck fuck.
