AI videos are getting very good. How do we detect their trickery?

[MonkeeSage]

Very interesting. The "movie" part was still obviously slop, and no, you didn't sneak the resheathing of the katana past me, that had me facepalming, but the "making of" part of that was quite well done. After some of the usual tells he owned up to still being occasional part of the AI output (in particular the car door goofs), I was genuinely expecting things like the spacial arrangement of crew/dollies/stands, etc. to be occasionally messed up. So I suspect he spent more time over the making of parts than the short itself.

At the start of the short I couldn't help but notice she chopped one guy's head in half, and half of it flew off behind him, then he turned back with an intact head while the chunk was still falling behind him (our left).

Spoiler: Screencap

 

[FatPhil]

At the start of the short I couldn't help but notice she chopped one guy's head in half, and half of it flew off behind him, then he turned back with an intact head while the chunk was still falling behind him (our left).

Spoiler: Screencap

View attachment 89531

That was probably the most egregious bit he left in (knowingly, as he admits). Other nonsenses were a kick to the chest leaving the victim's shirt all bloodied, and of course all of the smears of blood on the screens that made no physical sense (they weren't splashes, so what made contact with the paper - the katana that's sharp enough to slice through skulls?) - however, that (and maybe the kick) was probably a deliberate stylistic decision - but I seem to remember one of the guys moving by almost hovering forwards rather than moving his legs properly.

I'm 100% sure I could be proportionally just as critical of the not-AI, high budget, /Kill Bill/ (so scale that up to several hundred criticisms), which is why I've never seen it, I know it's going to have to do well in order to get even a 3/10 on IMDB from me. I'm 100% sure this AI has been trained on all that Tarantino nonsense, and all the original stuff QT rips off, and so it's happy to unthinkingly spew out limitless quantities of simulacra, but with even greater detachment from reality. Apparently some people find that tolerable.

 

[JMartJr]

Sometimes, the trickery is easier to detect than others.


View attachment 89802

 

[Ann K]

Sometimes, the trickery is easier to detect than others.

"Oops, page not found".

 

[JMartJr]

"Oops, page not found".

Hmmm... no idea whats up there. I'll re-upload it if it is still a problem in the morning.

 

[JMartJr]

Sometimes, the trickery is easier to detect than others.

Dang, no longer editable. So let's try posting it again:

I'm glad it says AI on there, I never would have figured it out otherwise. (Protip: among other things watch the hammer and chisel.

 

[JMartJr]

An interesting video from Atomic Shrimp about "fake quilted handbag stores going out of business" scams using AI imagery and text and flooding the zone with scam "stores" faster than they can be taken down. Even if that specific topic is not of interest, his conclusion might be:

I don't know if that will ever actually happen (AI getting so good it can't be detected -- JM) and I have met people who assert that they will always be able to tell when something is AI generated. I think that's the sort of confidence that will tend to set a person up for embarrassing failure. But even if there are such wise and perceptive people who will always be able to tell, for most of us, I think we have to acknowledge that it's harder than it used to

All of the obvious tells in any particular era of AI generated stuff are pretty much by definition the problems that the next phase will try and perhaps succeed to solve. In short, whether or not we can rely on AI generated content always being obvious, we probably shouldn't rely on that. We need to look somewhere else.

Fortunately, the other things to look at aren't so amenable to being swept away by technological progress because the other things are the very mechanics of the scam itself. That is, the scam is often trying to get you to make a snap decision based on feelings. These scams try to engage with your emotions, recounting some prosaic background story about a crafts person who labors with love and invests their heart into their work … They're trying to engage your sense of camaraderie and empathy and presenting you with a scenario where you might feel that the protagonist richly deserves some just reward. And they want you to reach in your pocket for that reward.

And a different emotion, but a powerful one, the sense of urgency, the finality of the deadline, the urge to act now very fast or forever forfeit some opportunity. FOMO, fear of missing out, as it's called. It's not only a powerful motivator to make people act, but it also tends to cause people to pursue that action without due caution. There is no opportunity for you to think, no time to sleep on the idea, not a moment to waste or you might regret not doing this thing…

And these factors in scams will probably not go away or become less common because they're not new. … it's the stocking trade of scammers. These are the things we probably need to be more conscious of. To try to take a step back from our immediate thoughts and feelings and take a look at ourselves, to observe our own behavior as though from an outside perspective, to wonder if we're being manipulated, and if so, why.

I believe that will provide better resistance to scams than to try to learn to recognize fake images using criteria that might be obsolete next week. This is what we call a heuristic approach -- to try to understand what something is doing rather than just looking at what it is.

If you're being told something is urgent, it's almost never the case that it's so urgent that you don't have time to think, you don't have time to probe the urgency and question the truth of it. … there's a fair chance it's not actually urgent for you at all and they're just manipulating you.

And I think that's the thing we need to guard against, not just to act on what is presented to us.

Source:
Source: https://youtu.be/MgxBlrF7Hbk?t=547


I am not sure how to generalize that from dealing with faux-retail scams trying to sell you stuff that isn't real to the broader world of AI being used to get you to swallow bunk of any and all sorts -- from getting you to wish non-existent people happy 100th birthday on FaceBook and so driving engagement to a fake site, to convincing you that NASA is faking everything and the Earth is flat and vaccines are deadly and Bill Gates is putting nanobots in your cat and birds aren't real, to getting you to vote for a candidate you would otherwise never support -- across that range of fraudulent uses of AI, the techniques of the con will be varied.

But I think he has an important point; that we can certainly spot fake AI trash for now if we look for the right tells but we can't count on this always being true, and we'll at some point have to fall back on recognizing when we are being manipulated, when a story is too good (or bad or urgent or scary) to be true, and to not believe stuff jut because it accords with what we already suspect might be true or makes us feel like we're special.

 

[Todd Feinman]

But I think he has an important point; that we can certainly spot fake AI trash for now if we look for the right tells but we can't count on this always being true

Back to witness testimony?

 

[JMartJr]

Back to witness testimony?

Ha! Touche!

But to answer seriously, there is a difference between witnesses getting stuff wrong with no way for us to tell whether that is happening, n the one hand, and intentional liars on the other. Liars and con men have a goal, that may lead to the use of common techniques that have evolved in the history of scamming to be used to achieve that goal, such as the ones Atomic Shrimp talks about in fake online retail scams, and those may be something one can learn to recognize.

And yeah, I am deeply worried about what happens as evidence gets easier to fake and we're left with just believing what somebody tells us!

 

[Gary C]

Back to witness testimony?

Only if you know the witness IRL. Otherwise, he/she/it could be an AI fake as well.

 

[Todd Feinman]

Only if you know the witness IRL. Otherwise, he/she/it could be an AI fake as well.

It's true that if someone would just club the alien on the head during an abduction attempt and steal the fleeing alien's wallet, all would be solved. Corbell episode.

 

[john.phil]

Spot the fake:

source: https://www.bbc.co.uk/news/videos/clyp5g7gpkzo

 

[Ravi]

For the ones that have not seen this channel

https://www.youtube.com/@ChloeVSHistory


Source: https://youtu.be/aaua5ghidk0?si=cCCEPykdaucdnSlm

 

[JMartJr]

I can't resist sharing this one. Can you spot the give-away?

Spoiler: In case you can't spot it...

Watch the jar the lady in the back is pouring tea into... that's a heck of a jar!

 

[RTM]

I can't resist sharing this one. Can you spot the give-away?

Spoiler: In case you can't spot it...

Watch the jar the lady in the back is pouring tea into... that's a heck of a jar!

View attachment 91541

Spoiler: Spoiler

Can I go to her pub?

 

[Leifer]

What about AI generated transcripts (creative writing), specifically in YTube documentary videos ?
You know those vids, that take 20+ minutes to explain what could have been done in 2mins.

I asked Google's AI search, "AI tells in creative writing".
Here are a few "tells" (copy and pasted here)...

Formulaic Rhythm: It heavily relies on the "rule of three" (triplet framing) or balanced parallel structures like "It's not just X, it's also Y".

Neatly Tied Resolutions: Stories often wrap up cleanly with an inspirational pivot or abstract moral summary, lacking the messy, open-ended ambiguity of human fiction.

Cliché and Hallmark Words: Generative models frequently default to overly smooth, hollow buzzwords or transition crutches like delve, leverage, tapestry, testament, or seamless.

"Community consensus on forums like the OpenAI Community Forum highlights that AI prose frequently carries an overly eager-to-please, polished, or formal business-like tone even inside fictional narratives."

I came across a recent example. It's a series of vids about the history of HAM radios. (link)
You don't need watch any of them, I can tell you what I found.....



The visual and format "tells" :
* The "preview image" is AI constructed. If it is not obvious already, details like the man's plaid shirt is a good place to look. The fabric's pattern has inconsistencies and items are out-of-scale.
* Typical AI voice narration.
* Never showing the video channel's maker.
* All visual content are clips from other videos or photography grabbed from the web.

The transcript/creative writing "tells" :
* Every video begins with a story about the person who "started it all".
* The video is filled with too many perfectly polished facts.
* Every video ends with a moralistic summation and includes a nostalgic sentence about "remembering" (every damn video, it's annoying).

Another "tell" oopsy...
* The earliest video on this particular channel (SilentKeyRadio), is of a Middle Eastern or India region man, giving a tour of his rooftop. Not a HAM antenna in sight.

 

[Ravi]

I also am allergic to these type of videos. You feel automatically dumber after watching these. Some go in the extremes and just go on and on and on about something and thus repeating stuff endlessly (with usually wrong info) while showing AI imagery or stuff that has zero to do with the subject. Make your choice of channels to follow wisely.

 

[Ann K]

I also am allergic to these type of videos. You feel automatically dumber after watching these. Some go in the extremes and just go on and on and on about something and thus repeating stuff endlessly (with usually wrong info) while showing AI imagery or stuff that has zero to do with the subject. Make your choice of channels to follow wisely.

They generally (or always?) have a click-bait title, but the narrative wastes minutes of time on the back-story. That's my cue to stop watching.

 

[Ravi]

They generally (or always?) have a click-bait title, but the narrative wastes minutes of time on the back-story. That's my cue to stop watching.

Yes. I actually stick to familiar channels. Any unknown channel with indeed "flashy" thumbnails is a no no for me, right away.
Also be careful on YT, it has a very powerful algorithm.
But it makes me sad that we have great platforms for all kind of interesting things, but being flooded with slop.

 

[jon_eir]

Below was present as real report along with photos of a supposed meta in a Discord channel i am a member of

On the face of it doesnt look too back, but the report number, material density and name/signature are wrong amongst other things. I would say its reasonable sophisticated. The metadata hadnt been removed or altered and openAi checker detected it easily

The second below, I was asked to look at by members of a another Discord channel. It looked so bad I didnt even bother to read it, and blank pages are a big give away I feel for Ai generation

Again it was littered with factual and textual errors. The metadata appeared to have been removed but chatgp determined it was created with digital assistance. This and the documents creation date was enough for mods to take appropriate action.

I wonder is it possible to developed a Ai checker tool box or some methodical here? Could more users with more experience suggest an approach for checking PDFs?

 

[MapperGuy]

What about AI generated transcripts (creative writing), specifically in YTube documentary videos ?
You know those vids, that take 20+ minutes to explain what could have been done in 2mins.

AI generated transcripts and summaries are the way to go.
Talking heads in particular are so boring to watch, when you can read a one or two sentence summary that in a few seconds.

 

[Mendel]

I wonder is it possible to developed a Ai checker tool box or some methodical here?

That approach leads to an arms race, because the fakers can just improve until their product passes the tool; and that process can be automated. Meanwhile, you're trusting a tool that is unreliable.

The only approach that always works is asking for provenance and provenience, i.e. where did this originate? and by what transmission chain did it get here?

The second document in particular looks like it's been redacted to be declassified, but it shows no markings that confirm this; and if it was you could track it to some office's FOIA server. If you can't (and can't track it to a reliable journalist who regulatly uses FOIA), you need to consider it a possible fake.

Remember: these people want you to believe it's genuine, so it is up to them to prove it is.
People who circulate documents like this without provenance cannot prove that; which means they're peddling fakes.

 

[Mendel]

AI generated transcripts and summaries are the way to go.
Talking heads in particular are so boring to watch, when you can read a one or two sentence summary that in a few seconds.

I don't see the point in having AI summarize an AI-generated video. Garbage in -> garbage out.

A good documentary provides you with far more than its summary provides.
But a bad documentary that simply stalls for time (like AI does) is neither worth watching nor worth summarizing.

AI summaries have no intent. I find them unreadable. (Sorry, Mick.)

 

[MapperGuy]

I don't see the point in having AI summarize an AI-generated video. Garbage in -> garbage out.

A good documentary provides you with far more than its summary provides.
But a bad documentary that simply stalls for time (like AI does) is neither worth watching nor worth summarizing.

AI summaries have no intent. I find them unreadable. (Sorry, Mick.)

Shorter garbage takes less time to digest.
And it can be interesting to see what types of garbage other people are ingesting.

 

[jon_eir]

The only approach that always works is asking for provenance and provenience, i.e. where did this originate? and by what transmission chain did it get here?

guess we'll apply this approach to everything-it would of saved the chemtrails, 9/11, bigfoot and all other debunkers a lot of time.

I see no logic in your argument-this is thread about detecting Ai trickery

 

[Gary C]

I caught an AI generated fake Richard Fineman lecture.

Vocal inflections don't match the context -- The AI facsimile of his voice kept placing the stresses on the wrong words as it spoke. The AI did not understand the sentence in the context of a lecture given to a student audience.

Anachronistic terms or usage -- The AI generated script used modern terms which were technically correct given the subject but which had not been adopted when Fineman was alive.

 

[jdog]

I caught an AI generated fake Richard Fineman lecture.

Vocal inflections don't match the context -- The AI facsimile of his voice kept placing the stresses on the wrong words as it spoke. The AI did not understand the sentence in the context of a lecture given to a student audience.

Anachronistic terms or usage -- The AI generated script used modern terms which were technically correct given the subject but which had not been adopted when Fineman was alive.

There used to be several Feynman lecture channels; it was a cottage industry for a while. First you have AI synthesize a "new" lecture from Feynman's abundant output, then you have another tool generate video and audio to match.

I like to listen to music while I work. I got tired of my Pandora channels, so I started listening to some YouTube music feeds; now they're overrun with fake mid-century Palm Springs soft-jazz vibes and repetitive chansons. There's no mechanism for filtering them out, unfortunately, only look for channels that promise no AI.