I pasted a screenshot of the default text ("GHOST FONT") into ChatGPT 5.6 Sol, told it to read it, and without further instruction it chewed on it for awhile before coming back with:
The text is a video. Every frame contain random dots, so an individual frame by itself doesn't contain the intended message
This "font" exploits the fact that current-gen frontier models will process video one frame at time, but each frame is noise, so by looking at frames in isolation doesn't reveal anything
Then, they add a hidden message to each frame just so that the agent report something and stop trying (because if the agent tried to correlate between the frames, they could discover the trick)
But if you pass just a frame, there is no message. Just the noise plus the decoy
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 decode it - I imagine a bit of additional prompting would get them there.
Funny, for me it is exactly the opposite: I can read the actual text very easily, but the “Written in Ghost Text” is barely perceptible to the point I would have completely missed it, if it were not for the comment pointing it out here.
I've just tried it on my large desktop monitor (roughly 1440p, not HiDPI), and I now see "Ghost Font" extremely clearly and can't see the decoy at all. If I scale my browser window to 30% zoom, then I can just see the "Written In Ghost Text" decoy message again.
My phone would have been zooming out the browser window, and making the dots even tinier, but the phone is HiDPI so it would have still preserved the dots. My eyes are middle-aged and probably starting to do the same kind of median-blur effect that models do when they resize an image. That's my current guess for why I can see the decoy more clearly on mobile.
If that's the case, then this trick will stop working as vision models approach pixel-perfect vision, instead of the current resizing that they do. Pretty cool as steganography though.
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".
Technically it's not a font, because font needs to be still. Analogy: if I took photo after book was closed would we say that font cannot be read by a camera?
Took a picture (only a single frame) and a 1s movie and threw it toward GPT 5.6 Sol (High):
Frame took 9m30s to decyper and GPT 5.6, it returned: WRITTEN IN GHOST FONT. Weird because I can only see "GHOST FONT" on the demo... but extracted data from image (I saw the highlited one) definitely looks like the "Ghost Font".
--
Video is more amusing, because after 3m GPT 5.6 figured it's motion-defined and asked to run QuickTime. At one moment I got:
> The animation is a motion-defined illusion. I’ve confirmed there’s no readable static OCR layer; I’m decoding its optical-flow field so the letter shapes become explicit.
At 4m it got extracted motion image that was in shape of letters but analyzed for 9 more letters and returned (at 13m36s) "GHOST FONT"
--
So:
a font... - FALSE - not a font, but video effect
...humans can read... - FALSE - I can't read it from image (but AI can!)
...but AI cannot - FALSE - it can
It's a static decoy message independent from what you type in. You can see it if you take a long exposure pic of the screen (e.g. with your smartphone).
Every single on of those answers I've seen _says_ they did decode it, but each and every one of them only found the decoy message without even realizing it.
When I gave Fable a screenshot it found the GHOST portion of GHOST FONT. Based on pixel density via some python code apparently - https://imgur.com/a/m3c801F
I had thought to use homographs. Sadly, all the models I tried were able to decode something like:
"ﾌㄖ乇ㄚ ᗪㄖ乇丂几'ㄒ 丂卄卂尺乇 千ㄖㄖᗪ"
However, I have noticed that voice assistants have a hard time understanding homonyms. Saying "bow" (as in to bow one's head) is often stored as "bow" (as in a bow and arrow). I wonder if there's a sufficiently complex sentence which is intelligible to humans but not to machines?
I haven’t tried, but it looks like you could trivially solve with optical flow?
Edit: looks like yes, from the shared chats people are posting. But it’s interesting to think of communication schemes that require a temporal component so any single image is unreadable and can’t be beaten by long exposures or other tricks (otherwise persistence of vision displays would satisfy). A sort of physical anti copy/paste.
It has bugs with long words: I typed "MARRY AND REPRODUCE". That was the only try that got the last word on a single line, but with too much space between U and C.
If the string is empty, I can read "WRITTEN IN GHOST FONT" very faintly. I'm guessing that is a watermark Edit: Ah, it's decoy text. Of course.
Ah, sorry, yes, tried with a very different message. But it can still read when telling that the text is formed by movement. AI doesn't see the world the way we do, so it's understandable. https://chatgpt.com/share/6a522660-929c-83eb-91ff-66b7873420...
An interesting experiment. I suppose that if you make things like CAPTCHAs too hard to do, we'd end up struggling as well. I can't imagine Ghost Font would be a good fit.
One side i really like it - i also love to play around with funny ideas - but have to say if i would read more than like 2 sentences with that font i'd throw up xD
Sadly another shot in the arms race that captchas started which just leads to increased inaccessibility.
It's interesting work for sure, but the end goal of separating out AI versus human consumers is tough. Indeed, if there was a lasting solution, that would be a substantial discovery that would quickly become very famous...
"find out with opencv what the hidden message is."
Skill issue on promoter side.
Fable oneshotted it for me.
"""
Reveal a motion-camouflaged message hidden in video noise.
How it works:
The background noise scrolls vertically at a constant rate (a few px/frame),
while the noise inside the letters does not follow that motion. Any single
frame looks like pure static. The decode is:
1. Estimate the background's global motion between consecutive frames
with phase correlation (this is the "optical flow" step - the motion
is a pure translation, so one global vector suffices).
2. Motion-compensate: shift frame t+1 back by that vector so the
background lines up with frame t.
3. Take the absolute difference. The background cancels almost
perfectly; the letters (which don't move with the background)
light up.
4. Average the residual over a SHORT window of consecutive frame pairs
(long windows smear the letters, because the text itself drifts
slowly over time), blur lightly, and threshold with Otsu.
PAIRS = 5 # number of consecutive frame pairs to average (keep small!)
BLUR_SIGMA = 6 # spatial blur of each residual, in pixels
START_FRAME = 0 # where in the video to start
def load_gray_frames(path, count):
cap = cv2.VideoCapture(path)
frames = []
while len(frames) < count:
ok, frame = cap.read()
if not ok:
break
frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY).astype(np.float32))
cap.release()
if len(frames) < 2:
raise SystemExit("Could not read enough frames from the video.")
return frames
def main():
if len(sys.argv) < 2:
raise SystemExit(__doc__)
src = sys.argv[1]
dst = sys.argv[2] if len(sys.argv) > 2 else "revealed_message.png"
frames = load_gray_frames(src, START_FRAME + PAIRS + 1)
h, w = frames[0].shape
acc = np.zeros((h, w), np.float32)
for i in range(START_FRAME, START_FRAME + PAIRS):
a, b = frames[i], frames[i + 1]
# 1) global background motion between the two frames
(dx, dy), response = cv2.phaseCorrelate(a, b)
dxi, dyi = int(round(dx)), int(round(dy))
print(f"pair {i}: background shift = ({dx:+.2f}, {dy:+.2f}) px, "
f"response = {response:.2f}")
# 2) motion-compensate frame b by integer (dxi, dyi), then
# 3) residual = |a - b_shifted| on the overlapping region
ys = slice(max(0, -dyi), min(h, h - dyi))
xs = slice(max(0, -dxi), min(w, w - dxi))
ysb = slice(max(0, dyi), min(h, h + dyi) if dyi < 0 else h)
# simpler: crop both to the common overlap
a_ov = a[max(0, -dyi):h - max(0, dyi), max(0, -dxi):w - max(0, dxi)]
b_ov = b[max(0, dyi):h - max(0, -dyi), max(0, dxi):w - max(0, -dxi)]
resid = cv2.GaussianBlur(np.abs(a_ov - b_ov), (0, 0), BLUR_SIGMA)
acc[:resid.shape[0], :resid.shape[1]] += resid
# 4) normalize + Otsu threshold + light cleanup
u8 = cv2.normalize(acc, None, 0, 255, cv2.NORM_MINMAX).astype(np.uint8)
_, mask = cv2.threshold(u8, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernel)
out = 255 - mask # black text on white
cv2.imwrite(dst, out)
print(f"wrote {dst}")
# optional: OCR if pytesseract is installed
try:
import pytesseract
text = pytesseract.image_to_string(out, config="--psm 6").strip()
print("OCR result:\n" + text)
except ImportError:
pass
You can also write using sound based/compressed 'text message' dialect: unless a real human is reading, automated watching tool should have a hard time (until coded/ML-ed on such dialects I guess)
I'm colourblind and this was very difficult to read. If it's the directions to the resistance hq, I'd put in the effort. If it's the manifesto, I just wouldn't read it.
heh although this font can be read by AI as other comments say, it gave me an idea:
How about writing or drawing stuff using optical illusions?
Shapes that not even human eyes can see, but the brain hallucinates: Shapes that seem to appear when you look straight at a pattern, or for a second after you look away from a pattern, or after you close your eyes, etc.
If you take a screenshot or a photo the image would just contain the same static pattern.
Why would they be trained to read a research experiment that fundamentally goes against the way they perceive? They can't train on this technique, they can only postprocess it into a form they can perceive.
I've had the same idea recently, and even set up a similar page to experiment with different speeds and noise types. I've had the idea to set up a message board where the font is basically 'GhostFont'. However, in my experiments, I've noticed that the biggest issue is that this only works for larger font sizes. If the text is as small as, for example, on HackerNews, it will become borderline unreadable.
Furthermore, if AI can read this or not depends on how the text sequence is pre-processed. If AI only gets snapshots of the text, it will probably fail in decoding the text as every snapshot contains only white noise and such no information. However, if we calculate the Deltas between the animation frames, the text will become decodable by an AI, you probably don't even need LLMs or CNNs for this.
I pasted a screenshot of the default text ("GHOST FONT") into ChatGPT 5.6 Sol, told it to read it, and without further instruction it chewed on it for awhile before coming back with:
> a screenshot
The text is a video. Every frame contain random dots, so an individual frame by itself doesn't contain the intended message
This "font" exploits the fact that current-gen frontier models will process video one frame at time, but each frame is noise, so by looking at frames in isolation doesn't reveal anything
Then, they add a hidden message to each frame just so that the agent report something and stop trying (because if the agent tried to correlate between the frames, they could discover the trick)
But if you pass just a frame, there is no message. Just the noise plus the decoy
If you take a frame you see it's neither random nor dots:
https://i.imgur.com/CgtyGjl.png
From a single frame you can definitely identify boundaries because the dots are sliding and get truncated.
Exactly. It's a good idea, badly executed.
What did you expect from a screenshot of obvious noise? The only thing that makes the text readable is the motion.
EDIT: On second look, the static screenshot does say "WRITTEN IN GHOST FONT".
> I posted a screenshot of static white noise to AI
HackerNews never disappoints
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 decode it - I imagine a bit of additional prompting would get them there.
Funny, for me it is exactly the opposite: I can read the actual text very easily, but the “Written in Ghost Text” is barely perceptible to the point I would have completely missed it, if it were not for the comment pointing it out here.
I've just tried it on my large desktop monitor (roughly 1440p, not HiDPI), and I now see "Ghost Font" extremely clearly and can't see the decoy at all. If I scale my browser window to 30% zoom, then I can just see the "Written In Ghost Text" decoy message again.
My phone would have been zooming out the browser window, and making the dots even tinier, but the phone is HiDPI so it would have still preserved the dots. My eyes are middle-aged and probably starting to do the same kind of median-blur effect that models do when they resize an image. That's my current guess for why I can see the decoy more clearly on mobile.
If that's the case, then this trick will stop working as vision models approach pixel-perfect vision, instead of the current resizing that they do. Pretty cool as steganography though.
Humans can read it, but with difficulty. If it becomes important, AI can be taught to read it.
So...usefulness?
It’s a research project, that doesn’t need to be useful. They wanted to explorer that area and share their findings
Also this can always result in something useful over time. I'd love if AI safe writing will be possible at some point again...
Claude Opus 4.8 can read it with a single prompt and no instructions on how to read it.
https://ibb.co/WWMSXQkQ
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".
But... neither of the videos say "this is a ghost font"? Are you sure you are a human?
and I cannot
(so either I am AI at a level less than Opus 4.8 or just all-round defective as a human)
Technically it's not a font, because font needs to be still. Analogy: if I took photo after book was closed would we say that font cannot be read by a camera?
Took a picture (only a single frame) and a 1s movie and threw it toward GPT 5.6 Sol (High):
Frame took 9m30s to decyper and GPT 5.6, it returned: WRITTEN IN GHOST FONT. Weird because I can only see "GHOST FONT" on the demo... but extracted data from image (I saw the highlited one) definitely looks like the "Ghost Font".
--
Video is more amusing, because after 3m GPT 5.6 figured it's motion-defined and asked to run QuickTime. At one moment I got:
> The animation is a motion-defined illusion. I’ve confirmed there’s no readable static OCR layer; I’m decoding its optical-flow field so the letter shapes become explicit.
At 4m it got extracted motion image that was in shape of letters but analyzed for 9 more letters and returned (at 13m36s) "GHOST FONT"
--
So:
:DEdit: https://imgur.com/a/SHlGu4O - work-in-progress images
> it returned: WRITTEN IN GHOST FONT
It's a static decoy message independent from what you type in. You can see it if you take a long exposure pic of the screen (e.g. with your smartphone).
Oh, cool I was wondering how can I get to see that decoy!
The answers here seem to establish that some frontier models can read it sometimes, but only after tremendous compute.
That still makes it (well, a future version) potentially useful as a captcha if we hate our users but hate AI more.
Every single on of those answers I've seen _says_ they did decode it, but each and every one of them only found the decoy message without even realizing it.
Related work (all involve noise and flickering images, photosensitive eyes/brains beware):
- "This game disappears if you pause it": https://youtu.be/Bg3RAI8uyVw
- "Illusion: If You Pause, The Image Will Disappear": https://youtu.be/ZqGfb_Vlrig
Hahaha one of the comments:
“Not just image. The sound also disappears when you pause”
Brilliant :)
When I gave Fable a screenshot it found the GHOST portion of GHOST FONT. Based on pixel density via some python code apparently - https://imgur.com/a/m3c801F
I had thought to use homographs. Sadly, all the models I tried were able to decode something like:
"ﾌㄖ乇ㄚ ᗪㄖ乇丂几'ㄒ 丂卄卂尺乇 千ㄖㄖᗪ"
However, I have noticed that voice assistants have a hard time understanding homonyms. Saying "bow" (as in to bow one's head) is often stored as "bow" (as in a bow and arrow). I wonder if there's a sufficiently complex sentence which is intelligible to humans but not to machines?
I haven’t tried, but it looks like you could trivially solve with optical flow?
Edit: looks like yes, from the shared chats people are posting. But it’s interesting to think of communication schemes that require a temporal component so any single image is unreadable and can’t be beaten by long exposures or other tricks (otherwise persistence of vision displays would satisfy). A sort of physical anti copy/paste.
It has bugs with long words: I typed "MARRY AND REPRODUCE". That was the only try that got the last word on a single line, but with too much space between U and C.
If the string is empty, I can read "WRITTEN IN GHOST FONT" very faintly. I'm guessing that is a watermark Edit: Ah, it's decoy text. Of course.
No, it can https://chatgpt.com/share/6a521e9b-c754-83ed-9d8e-cf653894eb...
That's the decoy message :)
Ah, sorry, yes, tried with a very different message. But it can still read when telling that the text is formed by movement. AI doesn't see the world the way we do, so it's understandable. https://chatgpt.com/share/6a522660-929c-83eb-91ff-66b7873420...
That is not the message. Did you read the article.
I downloaded the message video, renamed it to test.mp4
Still could read https://chatgpt.com/share/6a5221f0-e3fc-83eb-bc15-74420002b6...
I gave him this video file https://streamable.com/1bxxyf
An interesting experiment. I suppose that if you make things like CAPTCHAs too hard to do, we'd end up struggling as well. I can't imagine Ghost Font would be a good fit.
One side i really like it - i also love to play around with funny ideas - but have to say if i would read more than like 2 sentences with that font i'd throw up xD
related: https://news.ycombinator.com/item?id=45284311
It reminds me of https://silverspaceship.com/static/
I cannot read it. Maybe I am AI.
> "I cannot read it. Maybe I am AI."
I found the bot living in a simulation!
What do I win? Where's my prize?
I cannot read that text.
Security through obscurity is not security :)
> humans can read
strong statement, I struggle to read it
This is really cool!
Sadly another shot in the arms race that captchas started which just leads to increased inaccessibility.
It's interesting work for sure, but the end goal of separating out AI versus human consumers is tough. Indeed, if there was a lasting solution, that would be a substantial discovery that would quickly become very famous...
Old people and bad vision people firewall. This will violate disability accessibility requirements.
Isn't this triviaklu defeatable by taking the diff between two frames and marking changed pixels white and unchanged black?
"find out with opencv what the hidden message is."
Skill issue on promoter side.
Fable oneshotted it for me.
""" Reveal a motion-camouflaged message hidden in video noise.
How it works: The background noise scrolls vertically at a constant rate (a few px/frame), while the noise inside the letters does not follow that motion. Any single frame looks like pure static. The decode is:
Usage: python reveal_hidden_message.py input.mp4 [output.png] """import sys import cv2 import numpy as np
PAIRS = 5 # number of consecutive frame pairs to average (keep small!) BLUR_SIGMA = 6 # spatial blur of each residual, in pixels START_FRAME = 0 # where in the video to start
def load_gray_frames(path, count): cap = cv2.VideoCapture(path) frames = [] while len(frames) < count: ok, frame = cap.read() if not ok: break frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY).astype(np.float32)) cap.release() if len(frames) < 2: raise SystemExit("Could not read enough frames from the video.") return frames
def main(): if len(sys.argv) < 2: raise SystemExit(__doc__) src = sys.argv[1] dst = sys.argv[2] if len(sys.argv) > 2 else "revealed_message.png"
if __name__ == "__main__": main()You can also write using sound based/compressed 'text message' dialect: unless a real human is reading, automated watching tool should have a hard time (until coded/ML-ed on such dialects I guess)
I'm colourblind and this was very difficult to read. If it's the directions to the resistance hq, I'd put in the effort. If it's the manifesto, I just wouldn't read it.
How is it being colorblind affect it? The video is literally black and white only.
this is black and white, I thought color blindness is only for colors?
heh although this font can be read by AI as other comments say, it gave me an idea:
How about writing or drawing stuff using optical illusions?
Shapes that not even human eyes can see, but the brain hallucinates: Shapes that seem to appear when you look straight at a pattern, or for a second after you look away from a pattern, or after you close your eyes, etc.
If you take a screenshot or a photo the image would just contain the same static pattern.
i.e. qualia-based "cryptography" :)
"humans can read"
lol. Barely.
uuh, what's the point? i mean, models will just be trained to understand it
Why would they be trained to read a research experiment that fundamentally goes against the way they perceive? They can't train on this technique, they can only postprocess it into a form they can perceive.
I've had the same idea recently, and even set up a similar page to experiment with different speeds and noise types. I've had the idea to set up a message board where the font is basically 'GhostFont'. However, in my experiments, I've noticed that the biggest issue is that this only works for larger font sizes. If the text is as small as, for example, on HackerNews, it will become borderline unreadable.
Furthermore, if AI can read this or not depends on how the text sequence is pre-processed. If AI only gets snapshots of the text, it will probably fail in decoding the text as every snapshot contains only white noise and such no information. However, if we calculate the Deltas between the animation frames, the text will become decodable by an AI, you probably don't even need LLMs or CNNs for this.
yet