Artificial intelligence can create convincing fake videos, yet leading models struggle to identify their own kind of output. I see that failure as a serious challenge to claims that artificial general intelligence has arrived.
This matters because synthetic clips now spread through Instagram, TikTok, YouTube, and X. Ordinary users need dependable checks before sharing false inventions, staged disasters, or fabricated people.
An independent creator tried to meet that need by building a simple video detector. The result worked sometimes, but only after long testing, paid services, and repeated failures. That experience offers a useful reality check on AI marketing.
A Simple Test Exposed a Large Weakness
The proposed tool had one job: accept a video link and report whether the footage was generated by AI. ChatGPT judged the idea viable, while warning that no detector would reach perfect accuracy.
The first test should have been easy. A fake clip showed a woman in an inflated suit speeding across water. The detector called it probably real with high confidence.
“I would not trust the current high confidence labels.”
A newer model found faulty verdict logic. It then spent more than eight hours revising and testing the project, only to report that the goal had stalled.
That outcome is revealing. The model could write code and run tests, but it could not solve the core judgment problem. It eventually recommended Site Engine, a paid service already designed for AI detection.
Detection Remained Weak and Expensive
Early testing produced mixed results. The system correctly identified 23 of 39 synthetic clips. It also falsely flagged two of 36 real clips.
Several weaknesses became clear:
- Gemini often found no clear signs in plainly synthetic footage.
- Different detection services disagreed, producing inconclusive verdicts.
- Confidence labels overstated what the evidence could support.
- Frame-by-frame analysis consumed thousands of paid operations.
After more changes, Site Engine received priority over Gemini. The detector then identified fake clips involving the water suit and a crane lowering an aircraft. It also cleared real talking-head videos.
Yet the economics defeated public access. Testing only several clips consumed more than 12,000 operations. One recent video used 440 operations by itself. Even a $100 monthly plan offered too little predictable capacity for an open service.
The code could be shared through GitHub, but users would need their own API credentials. That is useful for developers, not for relatives who want to check a suspicious social post.
This Is Not Human-Level General Intelligence
Supporters may argue that detection is a permanent cat-and-mouse contest. Better detectors push generators to hide their traces more effectively. That is fair, but it does not excuse current performance.
Claims from major industry figures suggest AI can now match people across most intellectual tasks. I find that claim hard to accept when top public models miss visual errors that many viewers notice within seconds.
“If AGI is supposed to be as good as a human at pretty much everything a human could do, why is it so bad at telling me something is AI?”
No single detection failure disproves general intelligence. Still, repeated failures across models reveal a gap between fluent output and dependable judgment. Generating an answer is not the same as understanding evidence.
AI companies should publish clear detection benchmarks, error rates, and operating costs before declaring victory. Social platforms should also label generated media and preserve creation credentials.
Users should pause before sharing sensational clips, seek the original source, and treat detector results as evidence rather than proof. Until machines can judge synthetic media with far greater consistency, AGI should remain a research goal, not a marketing label.
Frequently Asked Questions
Q: Can an AI detector prove that a video is fake?
No. Current tools estimate whether synthetic signs are present. Compression, editing, and missing credentials can weaken their conclusions.
Q: Why did Site Engine perform better than Gemini?
Site Engine uses specialized visual detection methods. Gemini was better suited to describing footage than reliably identifying its origin.
Q: What does an inconclusive result mean?
It means the available signals conflict or remain too weak for a responsible verdict. It does not confirm authenticity.
Q: Why is public video detection costly?
Services may inspect many frames from each clip. Those checks can consume hundreds or thousands of billable operations per video.
Q: How can viewers check suspicious footage?
Find the earliest upload, review trusted reporting, inspect labels, and compare several detection tools. Avoid sharing until the source is clear.
