How to Spot an AI-Generated 'Prototype' Video Before You Back a Campaign

Bottom line: the era of "there's a video of it working, so it must be real" is ending. Checking hands, physics, audio-sync and background artifacts -- together -- meaningfully cuts your risk before you pledge. KAKEHASHI's trust-verification cluster has long flagged "no real working prototype, CG only" as one red flag among several (see signs of a suspicious crowdfunding campaign). This article goes one level deeper on that single point: now that AI video generation is mainstream enough that a convincing "prototype demo" can be fabricated without a real product existing, what specifically should a backer check?

Why this matters now

Text-to-video AI tools have gotten dramatically better and more accessible over the last couple of years, to the point that generating a short, plausible-looking product demo no longer requires any special skill. This isn't a hypothetical: Japanese security firm SECOM's own 2026 fraud-awareness bulletin (published April 2026) names generative-AI abuse -- deepfakes, voice cloning, phishing -- as an active, ongoing problem in Japan this year, and international coverage in 2026 (from consumer-security researchers and outlets covering AI fraud trends) echoes the same point globally. A crowdfunding "it's already working, see the video" clip is not exempt from that trend -- it's worth a moment's skepticism from the backer, regardless of the platform or the creator.

Four things to check

• Unnatural hands -- fingers that blend together, merge, or bend in ways a real hand can't. AI still struggles with hands more than almost anything else.
• Physics violations -- objects or parts appearing to pass through each other, or movement that doesn't respect gravity or momentum the way it should.
• Audio/mouth-sync mismatch -- even a small (roughly 100-millisecond) sync gap reads as "off." Try muting the video and watching the mouth movement alone against what's being said.
• Warped or inconsistent backgrounds -- straight lines that bend, text or logos that don't hold together, are classic AI-generation tells.

No single sign is proof by itself -- look for more than one. Short clips are easier to fake convincingly; the longer a video runs, the more these inconsistencies tend to accumulate and become visible.

A free tool that helps

For still photos, reverse-image search (e.g. TinEye) is the standard move. For video specifically, InVID-WeVerify is a free browser extension originally built for journalists and fact-checkers: it automatically extracts keyframes from a video and reverse-searches each one, so you can check whether the same footage shows up elsewhere online -- for example, as a demo for a completely different product. If a clip gives you the "I've seen this scene somewhere else" feeling, it's worth the minute it takes to check.

How this fits into backing a campaign

A video's authenticity is only one input into your decision. KAKEHASHI's 10-item Campaign Check (/check) treats "a real working prototype exists (not just CG)" as a critical item -- a "no" on it alone is enough to push the verdict to stop. Once you've sanity-checked the video itself, pair it with how to verify a creator's identity (corporate registry, SNS track record) and the other signs in signs of a suspicious campaign -- no single check should carry the whole decision.

This is also a direct extension of "backing is not buying." A campaign can be entirely legitimate and still be in progress with no finished unit yet -- that's normal. The distinct problem this article addresses is a demo that's staged to look more finished or more real than it actually is.

Sources

• caniphish.com -- 10 Techniques To Spot AI-Generated Videos
• InVID-WeVerify -- reverse video search project
• SECOM Anshin Life Navi #449 -- "Abuse of generative AI" and 2026 fraud tactics (2026-04-01, Japanese)
• TinEye reverse image search

Disclaimer

This article describes general verification techniques and does not accuse any specific campaign, creator or platform of fraud. No AI-detection method is 100% accurate. If you're still unsure, ask the creator directly or hold off on backing.