Hyperrealistic Deepfaked Tom Cruise in a Coffeeshop
old.reddit.comThe thing I find most interesting about people saying AI generated content is easily noticeable, is that they're always presented a question of "Do you think this is AI generated content?". Nobody consumes media with a list of questions they ask themselves for each piece of content. I expect people in the comments to explain how obviously fake it is, but barring the ridiculous scene, I'd 100% believe that was legitimate if I saw it with no AI warning.
> I'd 100% believe that was legitimate if I saw it with no AI warning.
And if it weren't such an obviously fake situation. Imagine it were more subtle, like Cruise out on a date in a high-end bar, coming onto a woman and overstepping boundaries, taken with a concealed cell phone.
Your logic is impeccable, but tragically already obsolete
Pretty soon the default will be everything is AI generated, even the most honest straightforward media will be AI edited/mediated.
It will be too convenient, consistent, flexible and efficient to record and edit most things with AI, not to
Before we asked “was that recorded in analog or digital?” Then we stopped asking.
> Before we asked “was that recorded in analog or digital?” Then we stopped asking.
Very few people did, and the consequences of that are incomparable. But I'm afraid your idea of commodity/ease of use will make much of the media AI mediated (at least). And like analog/digital, nobody will notice the difference.
What do you mean by saying that the consequences of not always asking whether something was recorded in analog or digital are incomparable?
Asking that question seems relatively irrelevant to me, especially now that pretty much everything is at some point converted to digital anyway, if only for ease of distribution.
In that sense filming on film, for example, is only an aesthetic choice and a choice about the tools you want to use on set, anyway. Which is not to say that it doesn’t matter or doesn’t make a difference. But in cinemas they will still project the digitized version of that film. And people will still distribute digitized scans of their “analogue” photos online.
I mean, even “digital” cameras record decidedly analogue voltages from photosites, the difference is mostly about at which point in the chain you put the analogue digital converter.
But what difference does it make?
Hard to remember, but the entire chain of recording, editing, distributing and broadcasting media was entirely via analog formats and analog handling.
When digital media first became an option, it was often not as high quality.
And since digital was easier to manipulate, it was often viewed as less trustworthy as it got better & better.
We stopped worrying about the manipulation of digital when its quality advantages became so great that digital was pervasive.
At that point, all media was digital and easily editable, so digital was no longer a red flag.
The truth of digital media was considered based on its sources, confirmation from second sources, etc.
Digital no longer meant less trustworthy because everything was digital.
—
We are going through those same stages with AI media.
Easy to do many things with AI now, but the results are often awkward or slightly odd.
As it gets better we have greater trouble trusting what AI produces.
But soon everything will be so easily editable, enhanceable, etc. via AI, so basically everything will be.
And the veracity of content will rest again, on sources, etc.
AI itself won’t be a red flag of manipulation anymore, because everything will involve AI.
In the sense that not knowing whether it's analog or digital doesn't alter the information. It's the source plus some noise; our eyes and ears perceive more or less the same, the interpretation is identical. But an AI-mediated video can invent or change the information in ways we we're not equipped to deal with. In that sense, you cannot compare analog-digital vs original-artificial.
For a year or so I've had the idea of setting up one of those "can you spot the AI generated content" tests except pitching good AI generated content against bad AI generated content, but keeping it absolutely secret that the good ones are AI generated. I suspect very close to 0 % of the test-takers would come out and say that "I'm confident all examples I saw were AI generated".
I think people would unconsciously shift their mental bar to the point where the bad stuff makes them believe the good stuff is real.
yeah, i've been fooled by context only even on non fake content. our brains are easily going along with the flow of images if it's normal enough.
that said this still has small lag and visual blur to feel uncanny
I see this "fake tom cruise" serie of video a lot on youtube (probably the same guy, because it's the same voice and same height). It's so perfect it made me wonder if the guy's isn't already a tom cruise lookalike and the AI part only adjusts minor details.
Do we have a photo of the original content ?
Could be this guy
https://www.hollywoodreporter.com/business/digital/deepfake-...
Does anyone have the original clip used? Curious to see it as reference for how much was altered
Was totally Uncanny Valley for me.
There's no obvious flaw (on the contrary, looks great), but there's just something off between how his shoulders, neck, and head move.
You only feel like that because you know it's a deepfake beforehand. It's a real person, an actor with a swapped face, so whatever he's doing with the movement of his shoulders neck or head is completely natural.
Nah, the head is too small for the body and the movement is wrong, maybe a person might take it as phone things or whatever but we have months! until this is actually perfected :)
...or maybe it actually does look awkward ?
If you believe peoples necks don't contain a spine I guess you'd call this "hyperrealistic" ?
Is there a technology we might use someday to prove the provenance of digital media? Like every video and photo has a blockchain entry or something. And any time there’s a modification, that gets recorded too. So in the future we might right click any random video and see its file history.
There has been attempts but they have fared poorly: https://petapixel.com/2011/04/28/nikon-image-authentication-...
There's an attempt happening here: https://contentauthenticity.org/
Wow, that looked incredible. It does worry me that you could generate a video about a celebrity / VIP in a compromising situation, and that it goes viral before anyone can disprove it. Alternatively, could there be a video scanner that tells you if the video has AI "fingerprints", and this would be built into social media platforms with a "likely AI generated content" warning on posts.
Or, we could just remove video/audio as evidence. Even with fingerprinting, it can still be faked enough to get past judged and elected folks who are already 200 years older than they should be while in office.
People are easily faked and, imho, it’s safest to assume everything is faked.
Part of me is morbidly curious about the possible chaos all this will be causing in 5 years. Maybe humanity can’t survive the mist this will bring.
It’ll be fine. Keep in mind that the world has run on handwritten signatures for a few hundred years. And anyone could say anything in a letter with no “pixels” to prove it right.
Digital media as strong evidence has been a novelty.
If anything the real thing to watch out for will be the echo of national digital ID. This would certainly solve the mild/moderate issues of deepfakes while (if done poorly) risking a serious upgrade to surveillance state/capitalism.
I wonder if having all this "strong evidence" has caused us modern people to lose the social norms and relationships that people had to depend on back when human trust was more important. Like, good fences make good neighbors but what happens when all of a sudden the fences don't work anymore and we have to be good neighbors without them?
I'm guessing an uncomfortable transition period at least.
The social norms of yore were not established by trust, but by the expectation of violence, real or perceived, should one decline to practice them.
Can these ai things generate a realistic signature?
Humans can generate realistic fake signatures. That‘s the point.
Related discussion from today: https://news.ycombinator.com/item?id=34543521 (Deepfaked Tom Cruise and Paris Hilton by Metaphysic (reddit.com) | 92 points)
I have heard that profiles (i.e. side views) are super hard for deep fakes (at the moment at least) due to the in-filling required.
This does not seem to suffer from it from what I can see.
"With the advent of AI, either we all go to prison or no one gets convicted, because everything can be fabricated." - me
AI replication of voices, deepfakes, etc...
Please notice the direction of eyes. This guy is never looking at the girl. Eyes are hard.
AI celebrity porn will be here soon