Back to blogAI, Not AGI: Why We Won't Invent What Your Recording Never Captured
Company

AI, Not AGI: Why We Won't Invent What Your Recording Never Captured

Some of the most interesting requests we get are impossible. Not hard. Impossible. Here's why, and why saying so is part of the product.

Echosaw Team4 min read

Some of the most interesting requests we get are impossible. Not hard. Impossible.


Every week, someone asks us whether Echosaw can do something that no software on Earth can do. We love these questions. They're usually asked in good faith, by smart people, who have spent a year watching AI do things that looked like magic. So it's fair to wonder where the magic stops.

Here's where it stops: at the edge of what was actually recorded.

Two requests we've been asked

The first one sounds reasonable. Can you add infrared to an existing image so we can see the people standing in the dark?

The second one is our favorite. Three people are talking in a bare concrete room. All three speak in a similar tone and key. There's one mono microphone. Can you tell us how many speakers there are? We'll add our own detail for fun: say the microphone was made in 1432. Nobody was making microphones in 1432, but the spirit of the question holds. Old, single-channel, cheap.

Both requests have the same answer, and the reason behind it is the most important thing to understand about AI today.

AI is not AGI

AGI, artificial general intelligence, is the idea of a system that can think, reason, and know the way a person does, across any task. It doesn't exist. Nobody has it. Not us, not the frontier labs, not the company with the slickest demo you saw last week.

What exists is AI: very capable pattern recognition over the data you give it. It can read speech, find faces, spot objects, follow a conversation, and search thousands of hours in seconds. It's remarkable. But it works on the signal that's in the file. It can't go back in time and stand in the room.

That distinction sounds academic until you're relying on the output.

You can't add infrared after the fact

A normal camera records visible light. Infrared is a different part of the spectrum. Thermal infrared, the kind that shows warm bodies in the dark, is heat that people give off, and an ordinary camera sensor isn't built to record it. Most cameras actively filter infrared out.

So when a photo was taken in the dark and the pixels where a person might be standing are black, those pixels aren't hiding anything. They're empty. No light reached the sensor, so nothing was written down. There's no infrared layer tucked inside the file waiting to be switched on.

Could an image model paint something that looks like a thermal image? Sure. It would happily draw warm figures in the shadows. But it would be drawing them from what people in dark rooms usually look like, not from what was in that room. That's not enhancement. That's illustration. And if the image is evidence, an illustration presented as evidence is a fabrication.

If the camera did record infrared, a night-vision security camera with its own IR lights, for example, the information is in the file, and that's a different story. We can work with what was captured. We can't work with what wasn't.

You can't count voices that were never separated

Software that tells speakers apart leans on a handful of cues. Where each voice comes from in the room. How each voice sounds: its pitch, its tone, the texture that makes your voice yours. When each voice starts and stops.

Now look at the request.

One mono microphone means there's no direction. Every voice arrives on the same single channel, so "where in the room" is gone. A bare concrete room bounces sound off every wall, so each word is smeared with echoes of itself and of everyone else. Three people with a similar tone and key take away most of what makes one voice different from another. And a very old, very limited microphone throws away even more of the detail that's left.

At that point, the recording doesn't contain the answer. A person listening carefully might hear clues in the words themselves. Someone says a name. Someone answers a question. Someone says "as I told you both." Those clues are real, and they're worth pointing to. But a confident number, "three speakers," stated as fact from that audio, would be a guess dressed up as a finding.

Why the honest answer is the valuable one

General AI tools are built to always give you an answer. Ask them for the impossible, and many will happily produce something that sounds right. A thermal-looking picture. A tidy speaker count. A fluent paragraph explaining it. We've written before about why fluent is free, and this is the sharpest version of that problem: the answer that was never in the data to begin with.

For a lot of our users, that's not a harmless quirk. A personal injury firm reviewing footage, an investigator going through hours of audio, a compliance team checking a recording: they need to know what the recording shows, not what a model imagines it might have shown. An invented detail doesn't just fail to help. It can wreck a case.

So we built Echosaw around a simple rule: work from the evidence, and point back to it. Every finding ties back to the moment in the file it came from, so you can open it and check it yourself. When the recording supports an answer, you get the answer and the receipt. When it doesn't, the right answer is that the recording doesn't say, and we would rather tell you that than tell you a story.

The takeaway

The question isn't how smart the AI is. The question is what was captured.

Light that never reached the sensor can't be recovered. Voices that were never separated can't be pulled apart with certainty. No model, however large, changes that, because it isn't a model problem. It's an information problem.

AI is extraordinary at finding what's in your media. It is not a time machine, and it is not AGI. The tools worth trusting are the ones that know the difference, and show you their work.

Stop watching hours of footage to find one moment.

Echosaw transcribes every word and identifies what's on screen (vehicles, people and visible text), and every answer links to the exact moment. Start with our Starter plan for $9/month — usage-based pricing with no per-seat fees, so you only pay for what you analyze.