Alexios Mantzarlis
@mantzarlis
Tech, facts, and carbonara. mantzarlis.com
I feel like @dwillner.bsky.social will have thoughts on how useful this "Seems like AI slop" reporting button will be for LinkedIn www.linkedin.com/posts/hsrini...
big believer in platforms doing more to fight misinformation, but not sure the UK government defining "trustworthy news providers" is an implementable solution hansard.parliament.uk/commons/2026...
Incredible stuff. Flemish newspaper De Standaard listed the Belgian Prime Minister's residence on AirBnB (using his name as host) and got approved. The listing was up for two months. (h/t @mschenk.bsky.social) www.belganewsagency.eu/belgian-news...
After publication, I got detailed responses on this matter from the Château de Versailles. Here are some highlights
Google said it removed some ads that violated policy. The affected websites claim to be very clear with customers but somehow decided to add a bar at the top of their website flagging that they are not the official website after we reached out.
For example, the official Louvre site charges an adult €22.00 for a ticket; the ticket trap charges €76.00. Ballparking based on their reach, the sites may have cost customers more than €2,000,000 in excess charges.
Two Spanish operators running 20+ lookalike domains for the official sites have flooded Google with ads for {attraction name} + tickets. They then overcharge for a base ticket by a factor of 2x to 7x.
I have typically found that Community Notes leaves good fact checks out but doesn't allow bad fact checks in. Yet in this case a "helpful" note actually makes a false statement (the image is AI-generated). Not great. h/t @mxbernhard.bsky.social
Sometimes you have to (metaphorically) kill off an emu to fact-check a shady marketing firm.
Looks fixed on main but behind the “more matches” interstitial on iPhone I still see the problem…
this seems totally reasonable i'm sure it will pan out exactly as projected
Pangram's CEO told me that "the tools and algorithms of the future are all probabilistic" but some professors and university administrators think using AI text detection in higer ed is "conceptually unsound,” because it “cannot meet the evidentiary threshold required."
Following some friendly pushback by @informor.bsky.social, I also tried to see how much I could generalize from the gaps I found. In about 4% of my subsequent tests, I found that slight inversions of text could flip the verdict completely (see example below).
Rhyming, inverted text, archaic language, and repetition were the strongest predictors that synthetic text would be branded as human-written.
It’s a vexillological bloodbath in Google’s World Cup schedule preview right now
After Indicator reached out, Apple and Google deleted 43 of the unsafe apps flagged by the audit.
I am a bit hesitant to engage with a reply this aggressive — I haven’t “demanded” anything of anyone, shared most findings in this free thread, and my academic position ends in one month — but I will assume good will. We know $12.95/mo is not trivial and offer discounted memberships (1/2)
My analysis also suggests that AI supercontributors focus less on politicians and hyperpartisan accounts. This may be because they are optimizing for a bridging algorithm that prioritizes broad consensus over factuality.