Largely self-funded and backed by Thiel and fellow contrarian entrepreneur Balaji Srinivasan, a crypto investor, Objection says it wants to act as a “private accountability system” for journalism. For a starting price of $2,000 (£1,470), anyone who feels they have been unfairly reported on can file a complaint. A team of investigators – including, reportedly, former National Security Agency operatives – then examine the story and submit their findings to an “AI tribunal” – a “jury” of large language models from OpenAI, Anthropic, Google, xAI and Mistral. Journalists are invited to defend their reporting. If they do not, the verdict is issued anyway, and published in a permanent public record.
Because AI repeats what everyone is saying. It doesn’t have investigatory powers. It’s a repeat engine. So, if you’re saying something that is not a commonplace then AI will disagree with you.
This seems to me to be a conceptual problem.
By intent possibly given its a pay to complain service. Otoh won’t the resulting court cases be fun.
Nah. Hide behind a corporate structure and just do a rug-pull liquidation when the writs arrive. Easy peasy.
It would seem to be a fixed fee liable claim establishment system. The plaintiff doesn’t directly financially benefit but the ambulance chasing lawyers would no doubt scan the permanent public record to cherry pick “no win no fee” cases at their usual 30-40% fee.
Even where the initial AI verdict goes against the plaintiff all is not lost as future automatic AI investigations could lead to a different result, particularly as the Overton window moves.
The end result will be the bottom feeders hoovering up more cash.
I think there is mileage in producing software that creates AI generated snotty letters.
Just input parameters ( eg bank, broadband provider, car insurance ) and a few key words ( eg incompetent, shyster , Indian Call Centre ) and off it goes.
Much more useful than this rubbish.
I have had two recent occasions to correct AI for stating as fact something I know to be untrue (and have the evidence for).
All the artificial ‘intelligence’ was doing was basing it’s ‘facts’ solely on researching a load of comments on places like Facebook.
GIGO.
Yeah, I have a similar experience. When trying to use AI to troubleshoot technical issues, the AI (pick your flavour) will simply make things up, including non-existant KB links.
There’s a moral here. Humans trying to create AI have difficulty understanding what intelligence is…
I’m not sure it’s the ones trying to create it that don’t understand intelligence – it seems more like it’s those trying to sell it that have the difficulty (or pretend to).
You’re right. My clumsy phrasing…
Sounds about as reliable as BBC Verify.
Pure pendantry, obviously, but the technical term that is used for this is a “Stochastic Parrot”.
https://en.wikipedia.org/wiki/Stochastic_parrot
Pure pendantry, obvs…
When I first started to referring to AI as a bunch of “regurgitation engines” various people protested. Now the remark goes unremarked. Maybe more people understand the point now.
Thing is, how could it possibly be anything else?
If it had any *actual* intelligence.
Am I wrong in suspecting that available AI applications are testbeds for AI? If you’re mug enough to use one you’re an unpaid research volunteer. Currently there’s very little value being generated from them. Dodgy research results, naff translations & enhanced cat videos?
There’s a grand amount of work out there to be a tester for AI. Ask it a question, read the answer, edit the answer, send back. £30 an hour sort of work.
No, I haven’t.
This isn’t Mechanical Turk work, answering qs being asked, it’s trying to train the AI itself.
And once everyone has asked all the questions that they can think of, and the answers have been included in the AI, we’ll have an AI which can pass every test. Except the test of answering some novel question that nobody has ever thought of before – on those it can still be hopelessly wrong.
I suspect anyone interfacing with an AI is training it. Because there’s feedback from the interaction.
Addolff/DM +1
I’ve started to regard it as not so dissimilar to Wokopedia. If it’s something unambiguously factual and straightforward, it’ll probably more often than not get it right.
If it becomes at all complicated, ambiguous, or in any way political, then forget it.
Not dissimilar to journalism in one sense. We all worked out that what journalists say is often bullshit, because they’re journalists, mostly they don’t really know anything. Ie, if it’s a subject we know about, we immediately know when they’re (often) wrong, but illogically used to believe that if it was something we didn’t know about, somehow they would be right.
AI is being trained on all the shite that comprises the acceptable mainstream consensus. /shrug
A journalist generally doesn’t know anything, so ideally he goes to someone and asks something like “Who are three people who know about (thing)?”
Then asks those three people and sees where they line up and where they don’t.
The weakness is that he has to talk to them, and willingness to talk sometimes has nothing to do with actual knowledge.
AI has a similar problem, if you think of it as something that responds with the weighted average of everything everyone has published on (thing).
I think you’ll find that’s what journalists used to do. Now thanks to their superb university educations they already know everything & don’t need to ask.
“You are all stochastic parrots.”
“I’m not.”