I just looked for ai slop on the internet and some videos are very realistic, so realistic it makes me doubt my ability to discern what’s real from what’s created to play with my emotions and generate money for the creator or advance an agenda.
I’m in my 40s. Video sites are full of what I assume overconfident people younger than me with comments like how boomers would believe any of these videos. I’m not that old myself yet but this stuff is scary. It can be used to denigrate a politician, to demonize or ridicule minorities, to share misinformation, to make porn using the face of somebody who rejected a disgruntled man…
It’s also very sad society actually wants this. It shows lots of people are actually very gullible and stupid.
A better question would be, how do I avoid being gullible with images and video so realistic? Because the more technology advances the worse it’s going to get.
It’s getting harder and harder but some things still seem to work:
- emotional emphasis on the wrong syllables
- dissonance between facial expressions and vocal delivery
- AI writing giveaways (it’s not A it’s B etc.)
- dreamlike / “floaty” motion
- unrealistic “depth of field” (objects don’t blur properly with distance)
- unrealistic lighting / coloring / appears stylized despite the attempted hyperrealism
- object permanence problems / subtle drift in sizes and proportions over time
Perfect framing is often a giveaway. Most people’s phone footage wobbles and moves around. Background object permanence is another thing to look out for
The Corridor Crew have done a couple of videos where they explain a lot of the things to look for and the reason AI finds it hard. Some examples:
Was about to recommend Corridor’s videos! They are really comprehensive and easy to follow.
That guys voice.
Plus, minor grammatical errors. eg. “Has increased the Tens of times”
Also, Bruce Lee having a fight.
That woman’s voice too.
And it’s not just Bruce Lee having a fight.
It’s angry conceited Martial Arts fighter picks random person out of a crowd for a demonstration,
who happened to be none other than calm collected Bruce Lee, yet no one knew who he was,
with 4k grayscale picture quality and at all angles of the dojo, including on individual audience members themselves, plus repetitive ‘damaged film’ effects for 20th century authenticity.
Start with the assumption the video is fake. Do you trust the source? If the video is of something outrageous, did you notice something feels uncanny? Does the lighting, shot angle, and proportions make sense?
If the answer is not yes, no, yes then it’s probably AI
This is pretty much it. Everything is fake unless you can prove it’s not
They always look a bit shiny/glossy
Whenever I don’t like the video or the person in it it’s ai.
It’s worth pointing out that whatever tips you get here, accurate though they may be today, are likely to be less accurate with time and this tech is improving at a rapid pace so if this is important to you, you’ll have to keep on top of what techniques for discernment are current at any given time or if there even remain any reliable techniques at all.
Looks like eventually it’s all going to come down to context and your assessment of the trustworthiness of the source. If something is so sensational it seems unbelievable and you don’t have much trust in the source from which you saw it, it probably is literally unbelievable until further corroboration emerges. This is a pretty exhausting model when modern live involves so much media consumption but I’d advise that there’s also an element of practicality you should incorporate in to your evaluation. You’ll have to decide if the answer to the question “is this AI?” is important for any given situation before actually investing any time or energy in to answering it. Sometimes you might get the calculation wrong and it turned out that by assuming something was innocuous enough not to matter one way or the other you ended up being misinformed about something that actually was important but at the end of the day we’re only mere mortals and can sometimes be wrong.
In reality genAI isn’t all that novel when it comes to disinformation. People have been tricked into believing stuff for centuries with far less convincing tech. Editing and framing have been useful for manipulating viewers for as long as the medium has existed.
Step back a bit from the title question, and when you see a video, ask yourself “to what degree should I treat this as useful information?”
Being real or AI-gen is part of the critical examination, but just because something isn’t generated doesn’t mean it’s a good representation of reality. Say, for example, you scroll across a video of an ethnic minority being belligerent on the subway. If it’s AI, you know it’s not real. But even if it’s definitely real, what does it mean to you?
Well, it means that for 30 seconds at some undefined point in time, at some undisclosed location, some anonymous person with supposedly notable physical or linguistic or whatever characteristics appeared to have been belligerent. If you’re a scientist trying to define reality, this is a statistically-nonexistent sample. If you’re the average Twitter user, though, it threatens to plug into a mental network of symbols and signifiers that approximate reality. And this is a problem whether the video is real or staged or edited or generated: it functionally isn’t real-- isn’t a useful representation of reality-- so the degree to which it’s internalized is a problem.
What about denigrating politicians? Again, I think genAI may slightly flavor this problem, but isn’t an operative variable. Sometimes evidence of corruption is enough to take down a politician, sometimes evidence of outright crimes against children and humanity won’t do it. Sometimes obvious lies or manipulations are enough to take one down. Sometimes a CIA-produced sex-tape with an actor lookalike just makes people think you’re cool instead of making them turn on you.
Don’t assume that recognizing fakes is a younger generation thing. Maybe the people spotting them grew up reading snopes and alt.folklore.urban so that they have a good idea of how urban legends, hoaxes and clickbait are constructed.
I would be concerned with the type of media I’m consuming in the first place if ai can make it problematic
The tells are getting subtler but currently a good tell is the camera work. If the video looks amateurish (grainy, blurry), but the video is more stable than a $60k gimbal mounted movie grade model with active scene stabilization, it’s probably AI. Another giveaway is motion morphing items from hammerspace where there’s a visible “seam” where the LLM is trying to reconcile two prompts. (example: you see a video of a horse rider, and the bridle and reins might suddenly decrease in length instead of becoming taught when the rider wishes the horse to stop, or you may see objects morph into existence from other structures, but so fast you might have to rewatch it a few times to catch it.)
