Building an AI Video Slop Detector: One Dev's Messy Real Attempt
A build log of an attempt to create an AI video slop detector with ChatGPT, Codex, Gemini, and Sightengine, and why detection is still unreliable.
What is an AI video slop detector, and does one actually work yet?
An AI video slop detector is a tool meant to take a video link from platforms like TikTok, Instagram, or YouTube and tell you whether the clip was generated or manipulated by AI. Based on one creator’s hands-on attempt to build one, the honest answer is: not reliably, not yet. The build combined ChatGPT/Codex for app development, Google’s Gemini for video understanding, and a third-party API called Sightengine for AI-detection scoring. Even after roughly 10+ hours of iteration, the tool still returned wrong or “inconclusive” verdicts on videos a human could identify as fake in seconds.
TL;DR
- A creator set out to build a simple website where you paste a video URL and get a verdict on whether it’s AI-generated, motivated by the flood of fake viral clips circulating on social media.
- The build used ChatGPT and Codex for planning and coding, Gemini’s video understanding model for visual analysis, and eventually the Sightengine API for dedicated AI-detection scoring.
- Early versions confidently mislabeled obvious AI content as real, including a viral clip of a woman in an inflatable “fart jetpack” suit, which the tool called “probably not AI” with high confidence.
- The first AI-only approach failed because the app’s confidence logic was flawed, a problem that surfaced only after switching to a newer GPT model and asking it to audit its own prior work.
- Adding Sightengine (a paid API starting around $29/month for the tier needed) improved detection, but combining it with Gemini’s judgment often produced conflicting signals and a default “inconclusive” result.
- After extensive prompting and self-testing, the detector correctly flagged clearly fake clips as AI-generated and correctly cleared a real talking-head video, though it still couldn’t explain why it flagged what it flagged.
- The project used a meaningful chunk of API quota fast (over a quarter of a monthly allotment in limited testing), suggesting real-world running costs and rate limits are a serious constraint for anyone trying to scale this idea.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Why is detecting AI-generated video so hard?
Video is a much harder detection target than text or even still images. A generated clip has to be wrong in a very specific, sustained way across many frames for a detector to catch it reliably, and modern generative video models are increasingly good at avoiding the telltale glitches (warped hands, flickering backgrounds, impossible physics) that used to give them away instantly.
The build log illustrates this directly. A video of a woman in an inflatable jetpack suit, an obviously synthetic scenario, got waved through as “probably not AI” by the first version of the tool. The system’s own reasoning pointed to a simulated particle effect from the suit’s exhaust and interpreted it as traditional CGI/VFX rather than generative AI, then assigned that read a high confidence score. That’s the core problem: a detector can find something unusual in a frame and still draw the wrong conclusion about what caused it.
There’s also the credentialing gap. Provenance systems like Google’s SynthID can confirm a video was made with certain Google AI tools, but they say nothing about video generated by tools that don’t embed those signals. Absence of a credential doesn’t mean absence of AI.
How does Gemini’s video understanding fit into this?
Google’s Gemini models were chosen specifically because they can process and reason over video content directly, not just individual frames, which matters for spotting inconsistencies that unfold over time. Google reportedly shipped an improved version of this video-understanding capability in September, which was folded into the build as part of the detection pipeline.
In practice, though, Gemini’s visual judgment turned out to be the weak link, not the asset. Across several tests in the build log, Gemini repeatedly reported “no clear AI indicators” on clips that were unmistakably synthetic, including a video where a person’s scale relative to an object (a plane being lowered into a hole) flips implausibly mid-clip. Gemini’s general-purpose visual reasoning wasn’t tuned for adversarial detection: it wasn’t looking for the fingerprints of generative models, it was just describing what it saw, and generated video is often plausible-looking enough to pass that basic check.
What role did Sightengine play, and was it necessary?
After an extended coding session that reportedly ran for more than eight hours and still ended in a “stalled” state, the AI coding assistant’s own recommendation was to stop trying to build detection logic from scratch and instead plug into Sightengine, a third-party content moderation and AI-detection API. Sightengine’s paid plan (roughly $29/month in the tier examined) was needed because the free tier, capped at a couple thousand operations per month, didn’t support video processing at all.
This is the part worth underlining for anyone considering a similar project: an existing commercial API already did a meaningful chunk of what the whole build was trying to accomplish. Once Sightengine was wired in via API keys and environment variables, it started catching AI-generated frames that Gemini missed entirely. In one self-run test cycle, the detector correctly identified 23 of 39 known AI clips and only falsely flagged 2 of 36 real ones, numbers that are better than pure guessing but still far from something you’d want to rely on for a confident verdict.
Why did the tool keep returning “inconclusive”?
The biggest usability problem wasn’t accuracy exactly, it was the logic for combining two disagreeing signals. When Sightengine flagged a video as likely AI but Gemini reported no visual indicators, the system defaulted to “inconclusive” rather than trusting the tool that was actually built for this job. That happened even on the fart-jetpack clip, where Sightengine had detected repeated AI indicators across multiple stretches of the video, a result effectively buried by Gemini’s disagreement.
The fix required explicitly reprioritizing Sightengine’s output over Gemini’s general visual read, plus removing an unintended limitation (a cap on clips under one minute and 50 megabytes) that had crept into the code without being requested. Only after these adjustments did the detector start correctly separating real talking-head video from clearly synthetic content, correctly clearing a genuine “person talking to camera” video and correctly flagging fabricated clips as AI.
Is a DIY AI video slop detector worth building right now?
For personal curiosity or a weekend project, yes, it’s a workable exercise, and the build log shows it can reach a state where it correctly sorts obvious cases. For anything meant to be trusted at scale, the honest answer is no, not yet, at least not without significant extra work and expense.
The practical issues stack up fast: API costs scale with usage (one test session alone burned through a large share of a monthly operations allotment), the tool still can’t explain why it thinks a clip is AI beyond marking timestamps as suspicious, and it remains vulnerable to disagreement between component models producing an unhelpful “inconclusive” default. None of this makes the underlying idea bad. It suggests that reliable AI video detection currently depends more on specialized third-party detection APIs than on general-purpose AI models like Gemini or ChatGPT, which weren’t built for adversarial forensics in the first place.
Frequently Asked Questions
What tools are needed to build an AI video slop detector?
The build combined ChatGPT and Codex for planning and writing the application code, Google’s Gemini model for video understanding, and Sightengine, a third-party API specializing in AI-content and moderation detection, for the actual detection scoring.
Can Google’s SynthID detect all AI-generated video?
No. SynthID can confirm whether content was generated using certain Google AI tools that embed its watermark, but it can’t identify AI video made with tools that don’t use SynthID, so a lack of a SynthID credential doesn’t prove a video is real.
Why did the detector call obviously fake videos “not AI” or “inconclusive”?
Two main reasons surfaced in testing: Gemini’s general visual reasoning often failed to spot generative artifacts and reported no indicators, and the app’s logic sometimes let that false negative override a correct positive signal from the Sightengine API, defaulting to “inconclusive” instead of trusting the more specialized tool.
Is Sightengine free to use for video AI detection?
No. Sightengine’s free tier offers a limited number of monthly operations and basic content moderation and AI detection, but it doesn’t include video processing. A paid plan, priced around $29 a month in the tier tested, was required to analyze video content.
How accurate was the final version of the detector?
In self-run testing described in the build log, the tool correctly identified roughly 23 of 39 known AI-generated clips and misflagged only 2 of 36 real clips, an improvement over earlier versions but still not accurate enough to treat as a definitive verdict on any single video.
