What “92% AI” means: reading a detector's score

A result like "92% AI" means the evidence leans strongly towards the video being AI-generated. It doesn't mean 92% of the video is AI, and it isn't a guarantee. On our site, the number always comes with a plain answer, such as "Most likely AI-generated", and the reasons behind it. Read those reasons before you act on the number.

Detector scores are easy to misread. Here's how ours works, the most common mistakes, and what to do at each level.

How our headline is built

Every result leads with the side the evidence favours, then answers in plain words.

• The winning side comes first. You'll see "88% Real" rather than "12% AI". Both describe the same result, but leading with the stronger side is easier to read.
• It never shows 0% or 100%, because no check is ever certain.
• A plain answer sits under it.

The plain answer is one of these:

• "Most likely AI-generated"
• "Most likely real footage with AI added", for example a real video with a cloned voice
• "Most likely real"
• "Leaning AI, not conclusive" or "Leaning real, not conclusive"
• "Too close to call"

The plain answer is the main message. The number tells you how far the evidence leans.

Four common misreadings

"92% AI means 8% of it is real"

No. The score is about how strongly the evidence points one way, not about how much of the video is AI. If you want to know whether a video mixes real footage with AI parts, a Deep Scan shows a separate breakdown: real footage, real footage with AI added, and fully AI-generated.

"55% is a little bit AI"

A score close to the middle means the checks don't agree, or don't find much either way. That's why results near the middle usually come with "not conclusive" or "Too close to call". Treat the video as unverified.

"99% means certain"

It means the evidence is very strong. Every detector makes mistakes, and ours publishes its own. Our post on when AI video detectors are wrong shows real examples.

"Most likely real means it's true"

A real video can still carry a false caption. "Most likely real" means the picture and sound look like they came from a camera. Whether the video shows what the caption claims is a separate question.

How the answer is decided

A scan runs several checks: our own video model, which follows motion over time; a frame model; a voice model; reading of any visible AI watermark; Content Credentials; metadata; known fakes; and for links, the post's caption and published fact-checks. Jev, a large language model, weighs all the findings together. Code rules stop it from calling a video "real" without solid evidence.

Two consequences matter when you read the result:

• A check that couldn't run is shown as unavailable, never scored. For example, a video with no speech has no voice result. It doesn't count for or against.
• Some evidence is stronger than any score. A visible "Veo" watermark, a matching published fact-check, or signed Content Credentials saying "made with AI" can settle a question that a statistical score only leans on.

Read the evidence, not just the number

Below the headline, a result lists the points for and against, which checks they came from, and the moments in the video behind them. You can tap a timestamp to see the exact moment. Ask:

1. Which checks agree? A strong video-model score plus a watermark is more convincing than one check alone.
2. Is anything pulling the other way? A counter point, like footage that looks camera-recorded, is worth weighing.
3. What was unavailable? If key checks couldn't run, the answer rests on less.

One clip's score versus a tool's track record

Here's a subtle point. Below is a frame from a clip of a snow leopard walking across rocks, made with Luma Ray 2. Our video model scored this clip 99% AI.

Video VerifierLuma Ray 2

Video is likely

AI-made

Test set · Oct 2026crackly-ai.comVideo model: 99% AI

But Ray 2 is one of the hardest tools for our model. Across our held-out Ray 2 test clips, it flagged 80.0% as AI and wrongly called 6.2% confidently real. So a 99% on one clip tells you about that clip, not about how reliable the detector is for every Ray 2 video. That's why we publish per-tool numbers on our accuracy page and the Luma detection page.

Copies change scores

The same video can score differently depending on the copy. A WhatsApp forward, a screen recording or a heavily compressed re-upload has less detail. We test clips as published and as lower-quality copies for this reason, but the original post is always the best thing to check. Paste the original link when you can.

What to do at each level

• "Most likely AI-generated": don't share it as real. If it's harmful or deceptive, report it.
• "Most likely real footage with AI added": read the detail line. It may say the picture looks real but the voice looks AI-made, a pattern some scams rely on.
• "Leaning" or "Too close to call": check the source, look for the original upload and, if it matters, run a Deep Scan, which checks every second of picture and the whole soundtrack. Our Quick Scan vs Deep Scan guide explains the difference.
• "Most likely real": good news for the picture, but still check where and when it was filmed before believing the caption.

To see how the checks themselves work, read how AI video detectors work.

Try reading one yourself

Paste a link or upload a file on the home page. A Quick Scan is free to try, needs no sign-up, and answers in under a minute for most videos. Start with the plain answer, then scroll to the reasons.

Quick answers

Is a higher score always more reliable?

A higher score means the evidence leans further, but reliability also depends on the kind of video and which checks ran. Strong, agreeing evidence from several checks is what to look for.

Why doesn't the score ever say 100%?

Because no detector can be certain. Showing 100% would suggest a guarantee nobody can give.