GPT-5.6 had no problem seeing the text when giving it a video recording of it. "No problem" as in using temporal analysis with optical flow and vertical-displacement maps to estimate how the image moved, and combine those into a motion map with increased contrast to see the text. I didn't give it any instructions though, just asked it what it said.
I threw it at Fable with the prompt "What does the message say?" and it also figured out without issue. Thought process is interesting https://imgur.com/a/GToXs6W
SMH.
Maybe AI will take over after all, just not due to the reasons we think.
It is smart, but it is not impossible to crack algorithmically imo. In particular, one can take two consecutive frames and run perturbations moving one frame around (ie shift the indexes) to find where the difference between the two consecutive frames minimises. Then subtract these two (perturbed) frames and ocr it. Works easily if the movement is linear/one-directional. I did this in 20 lines of code (checking only…
This is not resistant to false positives, and is expensive all things considered.
there's already captcha systems like this and are already easy to beat. You wouldnt buy much time, if any, with this. the reason is that all it takes to defeat is a human style temporal averaging vision system, take any 4 connected frames and take the average, text stands out and even quite bassic llm's can read. works not just for this but for a slew of other "ai" combating methods, so likely scrapers and the like wouldnt be slowed by this at all. Decreasing SNR to force longer frame windows to get avceptable snr for reading makes it harder for humans than it does bots where dynamic temporal averaging via tool call is trivial. run tool, it temporally averages until snr in the region of interest improves, measured by a plateu in change of importance of high frequency terms in fourier space, since if the moving background has been blurred to a smooth gray, the noise text stands out clearly in contrast. Not something an LLM could solve when asked, but now that I've made this comment publically for nothing other than a few internet points and ego stroking, it's a matter of time before some of the larger llms start suggesting this solution. He'll claude probably already will, since I got it to write code for this series of tests/experiments during early opus 4.x era. and I know other ideas LLM's were shit at that I discussed heavily with claude ended up being the go-to recommendation a generation later, despite no "public" discourse on the matter.
Is the answer correct? I don't seem to see any demo video with "this is a ghost font" encoded
It is actually correct but not in the intended way. Delete all the sample text. If you look at your screen from a distance you'll see a subtle ghost like text on the noise pattern. It says "this is a ghost font".
It’s the “decoy text” intended to trick LLMs, from the arricle:
> The decoy message serves as a final trick for a determined agent. When looking for a hidden message, it might first find the decoy message and think that that is the real embedded message in the video.
Took me a long time to realise that "Written In Ghost Text" wasn't actually the text I was meant to be reading, and that was only the decoy message. I can barely read the actual message, and it's about as "readable" to me as the Magic Eye 3D pictures. Actually I think I have a headache from looking at it on a mobile screen. As a research idea it's cool though. But I do wonder if/when AI models will figure out how to…
It depends on how big/dense is your screen, high frequency message is actual text while low frequency message is decoy text. Probably so that it shows when frames are averaged together without checking actual motion vectors. It is similar effect like that image showing either young woman or old man depending on distance you are viewing it from (or other one with either Einstein or Marylin Monroe)
We came up with a hacked font a few months back called "Noroboto"[1]. It just lies about its underlying Unicode representation which fools most of the naive ones and is crystal clear to humans. The better harness/model combos will render the text to a picture and OCR that, but many of them fall for the swapped characters. Was kind of hoping to see an extension of that idea here. https://tritium.legal/blog/noroboto
this exact trick is used in some drm schemes (to protect either text or font file, or both), where codepoints are either mixed up or font is converted to vector curves and assigned random indexes
GPT-5.6 had no problem seeing the text when giving it a video recording of it. "No problem" as in using temporal analysis with optical flow and vertical-displacement maps to estimate how the image moved, and combine those into a motion map with increased contrast to see the text. I didn't give it any instructions though, just asked it what it said.
For people thinking it detected the decoy text. It did not. It was the text I entered myself. Here are the thinking steps: Inspected video metadata, extracted frames, and analyzed content Detected and analyzed periodicity in stereogram image data Blurred, downsampled, and analyzed image frames Computed frame pixel shifts and cross-correlation Checked OpenCV installation, optimized and analyzed optical flow processing…
Im curious share the link - i cannot reproduce. Gemini 3.5 flash gets it. For those curious on ios you can record your screen using control center
Edit: interestingly gpt5.6 needed a slightly longer video and got it too
