This is a duplicate of the market that resolved NO on 12/31/2024: https://manifold.markets/Jacy/will-ai-generate-realistic-video-of
Note: This is an effort to make relatively objective, transparent Manifold markets that predict AI capabilities. I won't trade in these markets because there will inevitably be some subjectivity, and I'll try to be responsive with clarifications in the comments (which I will add to the market description). Feedback welcome.
Specifications:
The background of the video doesn't matter (e.g., there can be unrealistic animals or scenery in the background).
The AI-generated video needs to be indistinguishable from a 5-second clip of a single nonhuman animal in action (not just standing, walking, sitting down, etc.). Examples could include a moth flapping their wings, a snake slithering, a cheetah making a sharp turn, or a whale jumping. [Note: Running will probably not count, but I think an exceptionally good running video (e.g., 10 seconds, animal starts from walking then begins running, close-up details of muscle and joint movement) would count.]
This requires more than one example (i.e., not just a fluke), but it doesn't require robustness or high success rates. If a company releases a handful of examples and reliable evidence that they can make videos like this without human assistance (i.e., text-to-video), that's sufficient for YES even if the examples are cherry-picked; the idea here is that even if videos like these take 10 tries each, they could still be commercially viable, and they indicate that the model isn't just getting lucky—even if it still has a lot of hallucination problems.
Indistinguishability approximately means that in a YouTube compilation of 20 clips presented as real animal footage, fewer than 10% of casual, attentive viewers would suspect the AI-generated clip wasn't a real animal. It should be a real animal species, but it doesn't need to pass expert review. (The human observer test isn't a strict or precise requirement, in part because the results would depend a lot on how much people are thinking about AI at the time of the test.)
Most of the animal should be in the video and shouldn't be obscured (e.g., smoke, a blizzard, a dirty camera lens, excessive hair or fur). If the animal is moving quickly, the frame rate needs to be good enough to tell that the animal movement is realistic.
The model needs to be generating novel content. It can't just regurgitate real footage, even if it does some adjustment or recombination. (Thanks to @pietrokc for raising this in the comments. There may be quite a bit of subjectivity here, particularly because there tends to not be much public information about the training data of SOTA models these days.)
The spirit of this market (which will be used to resolve ambiguities that aren't resolved by explicit criteria) is whether the AI seems to have a world model of how animals look and move. YES resolution doesn't require the detailed knowledge of a scientist or sculptor but the general, intuitive understanding that almost all human adults have.
Update 2026-01-01 (PST) (AI summary of creator comment): Regarding the novelty requirement (that the model can't just regurgitate real footage):
Novelty is acknowledged to be hard to assess
For mainstream tools like Veo and Sora, the creator is comfortable saying most outputs are probably sufficiently novel (e.g., if they were doing extreme footage-matching, users probably would have noticed)
One way to demonstrate novelty would be a set of videos showing similar but distinct movements of the same animal, such that it's unlikely there was closely matching footage for all those movements
Update 2026-01-01 (PST) (AI summary of creator comment): The market requires more than one example of realistic animal movement videos to resolve YES, but does not require robustness or high success rates. A single exceptional video that got lucky is insufficient, but a handful of cherry-picked examples from a company demonstrating the capability is sufficient.
Update 2026-01-01 (PST) (AI summary of creator comment): The creator has provided specific feedback on submitted videos:
Videos that do NOT qualify:
Dalmatian obstacle course: Too many errors (legs passing through objects)
Google Veo 3.1 Wildlife Documentary: Mostly walking (explicitly insufficient); brief movements are <5 seconds
Snake in grocery store: Too brief, frame rate concerns
Snake in desert: Too short, movement too slow and simple
Whale jumping: Less than 5 seconds
Snow leopard/mountain goat: Leopard movement too short; goat movement unrealistic (rag doll-like); contact between animals seems off
Key clarifications:
The creator emphasizes focusing on quality over quantity when submitting examples
Movement must be >5 seconds (this is being strictly enforced)
Snake slithering is suggested as a relatively easy movement to focus on
The creator will need to examine real animal footage to properly judge some submissions
1,000🏅 Top traders
| # | Trader | Total profit |
|---|---|---|
| 1 | Ṁ563 | |
| 2 | Ṁ404 | |
| 3 | Ṁ314 | |
| 4 | Ṁ295 | |
| 5 | Ṁ280 |
People are also trading
@Hakari I'm still waiting for video links that meet the exact resolution criteria. In my most recent comment below, I explain why some of the videos posted so far don't seem to qualify.
@Hakari there's no exact number.
This requires more than one example (i.e., not just a fluke), but it doesn't require robustness or high success rates.
But I still have not seen a single example that meets the criteria.
That's a lot of video links! I didn't go through all of them, but I've jotted down some quick reactions. I encourage people (e.g., @KyleY) to focus on quality over quantity when posting, especially making sure the movement is >5 seconds, and to focus on relatively easy movements (maybe snake slithering?).
https://www.youtube.com/watch?si=vHff50CY9TynoWkq&v=pNMMFov2f_c&feature=youtu.be
This Dalmatian in an obstacle course doesn't qualify to me. It was certainly cutting-edge when it came out, but there are too many errors like legs passing through solid objects. That being said, the movement here seems much harder to generate than other types.
This "Google Veo 3.1 Wildlife Documentary" has some of the best movement I've seen. It's almost all walking, which is explicitly insufficient in the criteria. There's a fast squirrel scampering that's pretty good and a couple brief (<5 seconds) turtle and fish swims.
This is a very brief slithering of a snake in a grocery store. I worry the frame rate isn't good enough to discriminate, and I might need to look into snake movement a little to judge, but there's a second or two of unobscured slithering that's not obviously fake to me right now.
https://sora.chatgpt.com/p/s_6956bc8efeac8191bc1c9170320039fe
This snake in a desert is the most likely video to qualify for this market, but I think it's still too short, and the snake movement is relatively slow and simple.
This whale jumping is less than 5 seconds, and I suspect the pectoral fins aren't moving realistically (same issues as wings flapping, e.g., stiffness) but would need to watch some real whale jumps to decide.
I think this is a snow leopard jumping onto a mountain goat. I actually think the leopard movement here is pretty good, but it's probably too short, and I think the goat isn't realistic—moves kind of like a rag doll, which is especially weird since they see the leopard coming. The contact between the leopard and goat also seems off (e.g., goat moves their head away too early).
Altman came out creepy in this one but here’s my first attempt at a snake before I ran out of tokens for today https://sora.chatgpt.com/p/s_6955adfa9cc88191af76f7a89e5c3999?psh=HXVzZXItNkN5WFJRZnRWODZBZ0R1UlFiamNTRXR2.Mfmz1JwFzKHU
https://sora.chatgpt.com/p/s_6956bc8efeac8191bc1c9170320039fe
We know the model can do snake movement already though: https://manifold.markets/Jacy/will-ai-generate-realistic-video-of-6RcE8E0nUh#3xvftwajoyy
@KyleY I tried to fix the editing and completeness of that long comment of multiple videos a few times but Manifold is bugging out. A better version to interact with is in the edit history (click “edited”)
Edit: ok Manifold’s servers might be catching up now? Sometimes it shows the spaced out version sometimes the old with no line breaks
obviously NO cause these animals aren’t real
https://www.instagram.com/reel/DBSX-0GPYNJ/?igsh=MWY3c2prZGRhaTRobA==
https://www.instagram.com/reel/DO_U7zODPag/?igsh=NzVuaXF6ZmJzYTFz
https://www.instagram.com/reel/DPj21VqjCuP/?igsh=NTZ1c2V1cnN4M2p0
https://www.instagram.com/reel/DQ9on38iCOW/?igsh=NW02bG53NWl3MTBq
https://www.instagram.com/reel/DSvw88-ijdJ/?igsh=MXBzZ2xlYzdod3Uxbg==
https://www.instagram.com/reel/DRim-l_iFRT/?igsh=MXM5dDZycnV4MWc0cA==
https://www.instagram.com/reel/DBLtFrrMNdf/?igsh=MXFkMTRldnFrcHJ4ZA==
/jk
What needs to be better in this? https://manifold.markets/Jacy/will-ai-generate-realistic-video-of-6RcE8E0nUh#i18txqkwccj
I think not all, but some of those are pretty realistic.
@ikoukas the main concern may be here: The model needs to be generating novel content. It can't just regurgitate real footage, even if it does some adjustment or recombination
@1bets that's a very vague and basically unsatisfiable rule if taken strictly, because if it's a cat riding the bicycle it won't be realistic by default. All things that appear realistic obviously are not novel, or they would not appear realistic.
Novelty is hard to assess here. One nice thing to have would be a set of videos all showing similar but distinct movements of the same animal, such that it's unlikely there was closely matching footage for all of those movements. But with the mainstream tools like Veo and Sora, I'm comfortable saying most outputs are probably sufficiently novel (e.g., if they were doing extreme footage-matching, users probably would have noticed).
@Jacy - The question is whether AI will generate 'any' realistic video of animal movement, not whether 'all' AI-generated animal videos have realistic movement, correct?
Sorry for not getting to this sooner, but I think you're basically correct. It can't just be one single exceptional video that got lucky, but, as stated in the criteria:
This requires more than one example (i.e., not just a fluke), but it doesn't require robustness or high success rates.
@Jacy
"(e.g., if they were doing extreme footage-matching, users probably would have noticed)."
I don't think so, slop-addicts don't really notice much, & they care even less