You are watching a commercial. The people look slightly too smooth. Motion is fluid but strangely weightless. The background looks real, yet not quite governed by reality.
Then one person in the middle looks unmistakably real.
The immediate reaction is simple:
“Did they put a real person inside an AI-generated world?”
By 2026, that production model is no longer unusual. Generative AI can create characters, scenery, extras and stylized environments, while a real expert, employee, actor, product or logo remains anchored in live action.
JEX offers a useful case study, but the evidence has to be separated carefully.
The answer in five seconds
- JEX’s official L8020 site states that Hiroshima University professor Hiroki Nikawa, the discoverer of L8020 lactic acid bacteria, appears in its 2026 TV commercials.[1]
- Separately, production company macs REALIZ explicitly says that JEX’s Glamourous Butterfly brand movie used AI-generated human video based on original illustrations plus CG processing.[2]
- Therefore, it is confirmed that JEX uses generative AI in advertising production. It is not confirmed by the sources checked here that every surrounding element in the L8020 commercial is AI-generated.
- Public 2026 pricing ranges from ¥29,800 for a short AI video to ¥50,000 for a 15–30 second AI commercial, ¥298,000 for a commercial/brand movie package, ¥500,000 for live-action people composited into an AI world, and ¥700,000 or more for heavily human-finished custom production.[3][4][5][6][7]
- The expensive part is usually not one model generation. It is concept development, shooting, compositing, editing, revisions, audio, rights clearance, brand control and delivery responsibility.
- Modern AI video is often exposed less by a single broken frame than by temporal consistency, motion, contact, physical plausibility and subtle changes across frames.[8][9][10]
- But human detection is not perfect. Being personally quick at spotting certain artifacts does not mean people can reliably detect all synthetic video.[8][11]
1. What is actually confirmed about JEX
Two separate campaigns can easily get mixed together.
JEX’s L8020 site says that Professor Hiroki Nikawa appears in TV commercials aired in 2026. A February campaign in the Chukyo region also names two child performers appearing with him.[1]
So the professor is a real person appearing as himself—not an “AI professor.”
A different JEX project is explicitly AI-based. macs REALIZ lists the Glamourous Butterfly brand movie as using human video generated from commissioned illustrations, followed by CG processing. Its 2026 key visual is also described as AI-generated from original illustrations.[2]
The safe conclusion is therefore:
JEX uses both live-action advertising and explicitly AI-generated advertising.
What should not be done is to jump from “this looks AI-generated” to “the production company officially says this exact L8020 shot is AI-generated.” The public evidence checked here does not establish that second claim.
2. Why “AI world + real person” makes economic sense
A conventional commercial set in an unusual world may require locations, sets, art direction, lighting, wardrobe, actors, crew, VFX and post-production.
Generative AI can remove part of that physical production burden. Backgrounds, extras, atmospheric shots and sometimes major characters can be synthesized without a full shoot.
Meanwhile, the parts where authenticity matters can remain real: an expert, a spokesperson, the product itself or a recognizable brand asset.
The division of labor becomes almost comically clean:
World-building: AI. Credibility: live action. Product pack: please do not mutate between frames.
A hybrid can preserve “who is saying this?” while reducing the amount that has to be physically filmed.
3. How much does it cost in 2026?
Published Japanese pricing shows how wide the category has become.
CLEIA AI STUDIO advertises 15–60 second AI videos from ¥29,800.[3]
Syashin Lab lists 15–30 second AI commercials at ¥50,000 for 720p and ¥125,000 for 4K.[4]
North Image lists AI commercial and brand-movie production from ¥298,000, including planning and storyboards.[5]
Meadow Works AI lists a package that composites live-action people into an AI-generated world from ¥500,000.[6]
Movie Impact lists a custom “human finish” production tier from ¥700,000.[7]
So “AI video” can mean anything from a compact short generated from supplied assets to a fully directed brand production.
The generator got cheaper. Accountability did not become free.
4. What companies are really paying for
A personal 15-second experiment can be cheap. A broadcastable commercial has different requirements.
A package cannot become a different product for one frame. A logo cannot turn into nonsense text. A spokesperson cannot change identity halfway through. Health-related claims cannot casually cross advertising rules. Music, voices and performers need usable rights.
Then come storyboards, client reviews, re-generation, editing, grading, sound design, narration, mastering and multiple delivery formats.
Companies are not merely purchasing GPU time.
They are purchasing a process that ends with someone being responsible for saying: this version is safe enough to publish.
5. Why AI video can still “give itself away”
Older generated images often failed spectacularly: extra fingers, broken text, obvious asymmetry.
New video models are better. The evidence has shifted from obvious single-frame mistakes toward how reality behaves over time.
Skyra, presented at CVPR 2026, studies human-perceivable artifacts and spatial-temporal inconsistencies in generated video.[9]
MotionPhys focuses on physical consistency. It reports that visually plausible video can still contain motion trajectories that fail to respect inertia, continuous forces and realistic trajectory geometry.[10]
In everyday viewing, the feeling may come from details such as:
- walking or head motion that is smooth but strangely weightless;
- fingers or objects losing a precise contact relationship;
- hair, fabric or accessories subtly changing across frames;
- facial muscles, lips, cheeks and eyes not moving as one physical system;
- background people performing plausible but purposeless motion;
- cinematic camera movement combined with weak physical causality.
A paused frame may look fine. Press play and the world feels slightly soft.
6. A real person can make the synthetic parts more obvious
This is a perceptual inference, not a study specifically testing a JEX commercial.
If everything in a scene is synthetic, the viewer can adapt to one consistent artificial look.
Place one real human in the frame, however, and that person becomes a reference sample for reality.
Real skin contains irregular fine detail. Clothing responds to breathing. Weight shifts have inertia. Camera blur and depth of field follow an optical system.
If the generated surroundings behave only slightly differently, the contrast can become easier to see.
The visual joke becomes:
“The man is real. The universe around him is dreaming.”
A hybrid workflow may save production cost while simultaneously giving the viewer a built-in calibration target for spotting synthetic texture and motion.
7. But “I can tell instantly” does not mean humans are perfect detectors
There is an important counterpoint.
A CVPR Workshops 2026 experiment found that people became better at rejecting AI-generated videos as viewing duration increased: detection of generated content rose from roughly 58% with very short exposure to about 75% at eight seconds. Correct recognition of real videos stayed around 60%.[8]
More evidence helped people find fakes, but it did not make them fully confident about reality.
A 2024 meta-analysis covering 56 papers and 86,155 participants also concluded that human deepfake detection is generally weak; its pooled simple accuracy estimate for video was about 57%.[11]
So experience may help a viewer notice familiar failure modes quickly, but “I often catch it” should not be upgraded to “humans can always catch it.”
The next media-literacy problem is two-sided: detect synthetic media without starting to call every real video fake.
8. The likely direction: use AI where reality is optional
For advertisers, the key question is not whether AI was used. It is whether the same communication goal can be achieved with less physical production.
The strongest near-term model is therefore likely to remain hybrid:
- real people where authenticity matters;
- AI for worlds that only need to look plausible;
- strict control of the product and brand assets;
- human review at the final gate.
This is not quite a revolution in which commercial production costs fall to zero.
It is closer to a revolution in which the same budget can depict worlds that previously required much more production.
9. Conclusion: AI is cheap, but “AI-ness” can still come free with it
JEX’s public examples already show live action and explicitly AI-generated advertising coexisting within the same company’s marketing activity.[1][2]
The market now offers short AI videos for tens of thousands of yen and more elaborate brand work for hundreds of thousands.[3][4][5][6][7]
If that still sounds expensive, the reason is simple: the price is no longer mainly the act of generating pixels.
It is the human work required to plan, check, composite, revise and accept responsibility for the final commercial.
Meanwhile, viewers are learning too. They no longer need six fingers to become suspicious. Motion, contact, weight, faces, background behavior and contrast with live action can all trigger the thought: “that feels AI-generated.”
Yet the intuition remains fallible.
That creates a very 2026 situation:
AI is already good. Humans are getting used to AI. And one real person can end up standing in the middle as the scene’s last calibration point for reality.
Sources
- JEX / L8020 official site and 2026 commercial notices l8020.jp
- macs REALIZ, Marketing Solution — JEX Glamourous Butterfly Brand Movie / Key Visual macs-realiz.jp
- CLEIA AI STUDIO, public AI-video pricing cleia-ai.com
- 写真ラボ映像工房, AI commercial production pricing syashinlab.com
- North Image, AI video / CM pricing northimage.co.jp
- Meadow Works AI, live action × AI-world compositing pricing meadowworx.com
- Movie Impact, public pricing knowledge movieimpact.net
- Davodi, C. E., Barrington, S., Farid, H., & Cooper, E. A. (2026). “Perceptual Judgments of Video Authenticity: An Examination of Viewing Duration, Confidence, Content, and Strategies.” CVPR Workshops 2026 openaccess.thecvf.com
- Li, Y. et al. (2026). “Skyra: AI-Generated Video Detection via Grounded Artifact Reasoning.” CVPR 2026 openaccess.thecvf.com
- He, H. et al. (2026). “MotionPhys: Detecting AI-Generated Videos via Physical Consistency of Optical-Flow Trajectories.” arXiv:2608.20770 arxiv.org
- Human performance in detecting deepfakes: A systematic review and meta-analysis of 56 papers.” Computers in Human Behavior Reports 16 (2024), 100538 sciencedirect.com
