I agree with most of your reply apart from the word “people” here. As a journalist, I think I can say that vetted and verified sources have not escaped unscathed from the loss of trust in written and spoken reporting since the collapse in gatekeeping of news media that resulted from technological transformation. And that’s in an industry that has had a strong tradition of rules on verification and sourcing. Imaging has nothing like the same level of institutional regulation and education.
1 Like
TikTok’s “bold glamor” ai filter that’s all over social media now seems to me simultaneously amazing and horrifying, which I think is a general problem when you start arbitraging truth and believability. (It’s also of a piece with tech’s tendency to seek to make money through regulatory arbitrage).
1 Like
Supposedly, ai is defaulting to… teal and orange https://twitter.com/bildoperationen/status/1633082030178050048?s=46&t=KOx7Ovj8T13Bn2ohrOGcaw
In the short run, maybe. In the long run, getting flooded with AI-made “spectacular” pictures will make them infinitely boring.
The screensaver on my phone does this for some reason (I haven’t bothered to figure out how to turn it off yet). Every time I look at it is an Instagram-compatible wonder in saturated colors and eye-watering sharpness… which I don’t even remember for one second.
For me, photography is personal. It documents places and people, and I try to capture the atmosphere of a scene — what I felt at the time. Similarly, I find a photograph interesting when it evokes some kind of emotion in me.
I wonder how many people actually see a wedding video. I have leafed through wedding albums of people when it was the polite thing to do, but I can’t imagine any of my friends asking me to watch their wedding videos. If I have been there, great, if I haven’t, I am willing to take their word for the fact that they got married.
So, great, AI will generate videos of people getting married at the Taj Mahal or Machu Picchu or the surface of the Moon. What happens then? Maybe the couple will see it once, if they have nothing better to do. But after 15 minutes I imagine even they will find it boring.
PS. Actually, I would watch about 90 seconds of the moon version, provided it has spacesuits and an interesting plot, such as the father of the bride getting a suit malfunction and the whole thing turning into a rescue mission.
Haha!
To me that suggests it’s been trained on a defined subset of images. If it just gobbled up all images it could find it would hardly go that route. Colour matched images like that can’t be the statistical majority. Unless every frame of CSI Miami counts as an image 
So this suggests that the coders have made aesthetic choices along the way?
Just doing some quick googling around, it seems that the big ones used one particular enormous open source data set of keywords to images produced by a German ai “charity” (?!) to train their systems. Then some seem to have further trained it depending on what their users produced and liked. I’m sure that’s a gross simplification but I can see how things could start to skew unless managed carefully
Seemed worth a read, in case of interest
‘The answer to that is it does “so well” in limited domains. Where the overwhelming bulk of material on the Internet has prepared the way for it—has provided lots of correct, positive exemplars of answers that are successor pages and few incorrect, negative exemplars—it does well. If you find that corner of the vector space that is composed of working programs that people have written in response to requests—if there are lots of programs that work, and some programs that almost work, but almost no programs that are incomprehensible mush in the training data—then page-level auto-complete will be able to bull’s-eye in, and may well write you a working or a close-to-working program as its answer.
Otherwise?
Otherwise, it will act like a high school junior who more than half-groks the patterns of symbols, but does not understand . Why would it act like such a high school junior who more than half-groks the patterns of symbols, but does not understand ? Because it, pragmatically, more than half-groks the patterns of symbols. And yet it does not understand .
It is just a Chinese-Speaking Room, after all.’
2 Likes
as anyone who’s been somewhere with a good view of the Eiffel Tower may have seen in the past decade or so
Much older than this, for the tour Eiffel. The last decade or so is about finding other romantic places in Paris (if possible with the Eiffel tower visible in the distance).
2 Likes
However arrived at, human or machine, the key question is, what decisions are we comfortable with based on the reasoning mechanism?
2 Likes
Adobe doing the Midjourney/Dall-E thing, in beta, but using its own stock images as the training data. Their blurb references photo editing:
Well, this made me chuckle:
Total miss. They should’ve had him in puffer jacket and shorts
1 Like
It pissed me off. The story was ran in supposedly proper newspapers in Sweden. The thing that annoyed me is that supposedly critical journalist forgot that we’ve had faked images for a very long time. There are lots and lots of people who could have created that image in a gimp alternative 10 years ago. AI is irrelevant in this particular instance. AI only makes it more accessible, missing that critical point to surf the hype is inexcusable and misleading the readers.
1 Like
You have a valid point here. Nevertheless, the negligible cost of creating images like this matters.
While skilled users of GIMP (Photoshop, etc) could of course create these decades ago, now that anyone can create them, the sheer number of images will make it much more likely that some of them will slip through. Yes, I agree that they shouldn’t, but as long as newspapers get money from ads, clickbait images will be very tempting.
My preferred (and admittedly imperfect) solution is to
- read sources with a lower than daily frequency (eg weekly or monthly, like The Economist or The New Yorker), which are less tempted by clickbait because it is uninteresting by the time it is published,
- read sources which mostly require a subscribtion, relieving the financial pressure that promotes clickbait.
2 Likes
I think legacy papers with their own staff photographers wouldn’t cheat but there’s a whole raft of digital first publications or those that rely on freelancers where finances are tight and standards probably more fuzzy. Some will need to tighten verification. Mail Online will worry whether paps are making sh*t up.
1 Like
There’s really always been this, but it emphasizes the need for real critical thinking on the part of readers… sigh.
1 Like
I think Ezra Klein said it just reduces the cost of BS to zero
1 Like
Ok, then. F*ck AI:
“It allows, for example, users to generate images of President Biden, Vladimir Putin of Russia and other world leaders — but not China’s president, Xi Jinping.”
1 Like
Not listened to it yet, but this might be somewhat relative to this topic: Gadget Detective - A selection of free tech advice & tech news broadcasts by Fevzi Turkalp on the BBC & elsewhere / 19th April 2023 - Using Artificial Intelligence for Art on BBC Radio London
Not sure if this is relevant or whether I should be reviving such an old thread?
Financial Times data correspondent John Burn-Murdoch has an interesting column (£/€) on AI already taking white-collar jobs. He has a twitter thread that has the main points.
Even more tangentially, Sean Carroll is sceptical of the idea of AGI or the use of AI to generate art: