> The public posts and discussions being had about this subject are already informing AI companies on how to train their multimodal models to get around these obfuscations, most of which have already been broken. I'd argue every new font and tech demo is effectively a benchmark, daring AI firms come up with solutions to sidestep them. And they will be sidestepped, one way or another. If a human can see the information, that means there is a way the information can be parsed. "Ghost" fonts will become just another scraping obstacle with its own set of contingencies.
1. I don't like the sense of futility and powerlessness this advocates for.
2. I'm not sure it is so futile. I agree this stuff isn't encryption, which means it'll always be possible to circumvent the obfuscation, but it could raise the cost. Hopefully that can be done to the point where it's just not worth the bother.
That could happen if:
1. There are so many schemes out there the catalog of circumventions gets unwieldy.
2. Doubly so if the schemes allow generation of new obfuscated fonts per site or per page.
3. Then you're forcing the scrapers to pay a greater tax to get your text: spin up a Chrome instance to OCR a screenshot, or spend some a buck or two or LLM credits to reverse engineer the page in order to scrape it.
Thanks for introducing me to shieldfont.org! It's the first of these I've seen that feels designed to be more than a visual experiment, reading through their landing page is interesting. In particular, their section on accessibility seems to contradict this post's opening premise:
> Screen readers get the real words.
A screen reader reading down the page is never handed scrambled text, and our NVDA test asserts exactly that. Screen review and touch exploration are untested. ShieldFont hides shielded passages from accessibility tools by default with aria-hidden="true", because a decoy read aloud is fluent, grammatical, wrong English, and that is worse than silence.
> The real words remain sealed in the same page, and a visible notice above the block carries the control that uncovers them. It is on by default and reachable by mouse, keyboard and screen reader alike. Pressing it sets the reader’s browser to solving a compute-heavy puzzle: JavaScript and a few seconds of processing, more than most mass scrapers are willing to spend. That puts the words within reach of a screen reader, a translator and copy/paste.
I'd love to hear your thoughts on that. Also just aside, love this TUI-esque blog design and color palette (maybe a bearblog theme? still worth an upvote)
When you look at their live demo (https://shieldfont.org/demo/), it says: If you use a screen reader, custom font, or translator, please uncover the text before reading.
They also actively block copying the text, telling you to "uncover" the text first. The uncover operation is VERY expensive.
Anyone using assistive technologies or trying to copy "protected" text is SOL.
Search engines will index the decoy. You'll get no traffic. Their solution is basically putting yourself in a black hole.
> Search engines will index the decoy. You'll get no traffic. Their solution is basically putting yourself in a black hole.
It's only search engines? And who uses those anymore anyway? Other bots?
It doesn't affect organic traffic, so you are not really putting yourself into a black hole. A lot of website are driven by social media and organic traffic, so they would be just fine with this approach.
Isn't this the same sort of cost CloudFlare and anime catgirls are making us pay daily? Only enough to deter bots, or "a few seconds processing."
I take your point that one extra click/interaction required for screen readers only is objectively more friction, but it's only by exploring these technologies instead of dismissing them that we'll arrive at UX solutions truly work for people of all stripes (and ideally, not for bots, scrapers, and the like). I think if you're applying a tool like this you already have very different priorities than fueling Google.
As soon as you need to put a "decode" button for accessibility functions to work, you're effectively posting the key along with the cipher. It's self-defeating because any tool that supports screen readers will also support scraping. It's a fool's errand.
It's funny, I was just thinking that the one thing I hate most about the terminal is that its monospaced fonts and overly-long lines are a nightmare for reading. So the fact that somebody designed a blog reading experience to mimic this is just… ugh for me.
But you seem to appreciate it, and I'm sure others do too. Different strokes for different folks.
With this kinda of stuff it's hard to tell whether the person is doing it unironically, or knows it's "performative art". A while ago there was a trend of using a tool which imperceptibly perturbs an image in a way that supposedly breaks AI training on it. Of course, artists ate it up, despite the skepticism from AI researchers. Same with people setting up their sites to be "AI scraper traps", generating gibberish content. Probably also trivial to filter out, but people do it.
There could be benefits unlocked in legal documents by retaining a machine-readable version and distributing the obfuscated version with a legend at the top. We proposed one that said:
"This document contains mitigations against review by automated systems. Recipients should ensure that they have read the contents on screen or in print. Recipients with bona fide vision impairments may be entitled to unmitigated documents upon request."
In testing, obfuscating small portions of text slipped under the radar of most (then-)frontier LLMs.
We used a font that was rendered on the fly and reported faulty or fake Unicode mappings: https://tritium.legal/blog/noroboto but others have proposed and done the same with ligatures.
There was a story once about a boy with a wheelchair who needed a ramp to get into school, and the school made him use the loading dock ramp used for garbage and other things at the back. The school argued that it was an appropriate accommodation.
Accessibility is not accessible if you need to go through extra steps to get it.
Cerebrally, this as a solution makes sense. But if you know anyone with a vision or other impairment, gating it behind a request is not only cruel but gets within dangerous striking distance of an ADA lawsuit, for general applications.
Maybe in the legal field or specific niche cases it's possible. But this would represent a major step backwards in the work we've done lowering barriers for a population whose only difficulty in accessing common resources is because they were born, or got sick, differently than anyone else.
My dad caught paralytic Polio at age 2 and has had limited mobility his entire life, so I'm familiar with that issue.
Our internal, hypothetical use-case was between contracting parties who were looking to avoid terms escaping into the wild. This shouldn't show up in standard ToS or similar. There are already really good legal reasons for that.
> Accessibility is not accessible if you need to go through extra steps to get it.
That seems a bit entitled to me, especially in the story you shared. They had access to the school just like everyone else. Why does it have to be in the exact same spot? Surely they could be dropped off by the loading dock as easily as others could be dropped off out front, and maybe even moreso. Should the wheelchair accessible stall be the first stall in the bathroom so they don't have to go through "extra steps" to get to it? As long as the ramp had the proper gradation to satisfy the ADA, I don't see the problem.
> They had access to the school just like everyone else
They literally did not. They had access to the school from a different entrance than everybody else. Specifically, one intended for cargo before people.
It is not sufficient to work against current AI. It needs to also work against AI that has been trained by a competent team aware of your mitigation. Or worse, a competent developer with no particular AI skills.
Otherwise, you are relying on obscurity, and will lose as soon as you become interesting enough to matter.
You will also break non-AI machine processing use cases. That isn't just accessibility, it is things like search.
You will surely not have a good time enforcing the terms of a legal document that explicitly spells out that it is intentionally obfuscated from the party it intends to bind.
OCR is a more expensive option, right? I don't think these fonts need to stop ai from being trained; I think they just need to make it more expensive and difficult.
More expensive and not worthwhile are two different things. Also when certain implementations become popular it's much more likely someone will write a very efficient kernel for decoding said text making it much less expensive than generic OCR.
Also more expensive doesn't mean that something won't happen, only the dynamics of how it happens. For example if you put all your documents in images then some service might just sell the AI providers the text. That service may do underhanded things like bundle OCR in an app that does something else and use your phone to get the text out of these images all day.
It's slightly more expensive, but the cost would be worth it if this became common. It's not designed to be hard to OCR, it's designed to be hard to copy out of the source code, so it doesn't require a very advanced OCR system (and all the regions that would need OCR-ing are clearly marked).
I believe the cat and rat game is already there, for multiple places, spam, captchas and now for AI content, yes, it's objective is to make it harder for AI companies to get such data, if it wastes their time, it's a win.
> it's objective is to make it harder for AI companies to get such data, if it wastes their time, it's a win.
That's only one half of the equation, though, isn't it? What if it makes it harder for legitimate users as well? It seems there's a balance to be struck.
Kind of funny how you get downvoted for a rational take, but one that's not anti-ai.
You'd ask that same person how much they like captchas and I'm sure they'd think their a terrible idea and they've ran into all kinds of issues with them.
We are now in a weird middle ground where we want to write things OCR algorithms have trouble transcribing which also means we write things people with accessibility issues have trouble seeing. No child left behind?
It's like the old tale about the national park bin with the smartest bear / dumbest tourist crossover, except we're now comparing capabilities of the smartest AI with disabled humans.
I think the difficulty was specifically with CAPTCHA challenges, which had to be quick to generate but still legible - OCR on physical documents has to be robust to a different set of problems
That said, what kind of errors are you getting? I think the main practical difference between Tesseract's LSTM-based OCR and newer VLM-based OCR like PaddleOCR [1] is (hopefully) getting to skip making heuristics for the layout of the document, but errors possibly compounding over multiple tokens - are you getting individual illegible words or having the documents smooshed together because the OCR can't parse the layout?
Well, I used specialized software specifically for solving CAPTCHA on select few file sharing websites. My later experience with general purpose OCR software was just as awful as yours. But it's a product problem, not technology problem - you really don't need an advanced AI for it, you just need a good implementation of traditional shape-matching OCR that doesn't seem to exist anywhere for some reason.
Is it part of the joke that the site is intentionally over-pixelated while the author critiques readability? (edit: I don't mind esoteric design and I play old games. I found it funny to see a blog with aesthetics that are not optimized for long reading to complain about the readability of fonts)
I don't think any anti-AI font design is more than design-as-art statements against AI. If there is evidence of them being used exclusively for business and without accessible fonts aside them, I'm willing to be wrong.
Accessibility is important, and I find that "Reader Mode" in most browsers is quite good. Everyone should have access to the tools to consume content. Did your browser not provide that functionality?
I'm a bit frustrated by what seems to be a widespread strong negative reaction to anti-AI fonts. The accessibility problem is real, but I feel like that's a reason to push the investigation deeper for solutions to that problem, not a reason to abandon the effort entirely. The largest intellectual property infringement in the history of the universe is actively unfolding, and it's resulting in an existentially threatening transfer of wealth and power. That's a problem worth exploring every solution for, and solving it may entail some serious sacrifices.
It's simply that a font won't solve anything. That concept is worst than security by obscurity, as it causes more problem and add more constraints than what it solves... for a very limited time until AI bots are adjusted to decode those fonts properly.
Every half assed means of trying to confuse an AI is just a small bit of learning away from making the AI better than you.
Worse when you have people with disabilities, which I seem to be this week, you just make doing things a pain in the ass.
What I don't get is people like you think there is a solution to this. There is not. The harder you try you either exclude more actual humans or you align the AI closer to how people actually see.
> The public posts and discussions being had about this subject are already informing AI companies on how to train their multimodal models to get around these obfuscations, most of which have already been broken. I'd argue every new font and tech demo is effectively a benchmark, daring AI firms come up with solutions to sidestep them. And they will be sidestepped, one way or another. If a human can see the information, that means there is a way the information can be parsed. "Ghost" fonts will become just another scraping obstacle with its own set of contingencies.
1. I don't like the sense of futility and powerlessness this advocates for.
2. I'm not sure it is so futile. I agree this stuff isn't encryption, which means it'll always be possible to circumvent the obfuscation, but it could raise the cost. Hopefully that can be done to the point where it's just not worth the bother.
That could happen if:
1. There are so many schemes out there the catalog of circumventions gets unwieldy.
2. Doubly so if the schemes allow generation of new obfuscated fonts per site or per page.
3. Then you're forcing the scrapers to pay a greater tax to get your text: spin up a Chrome instance to OCR a screenshot, or spend some a buck or two or LLM credits to reverse engineer the page in order to scrape it.
Thanks for introducing me to shieldfont.org! It's the first of these I've seen that feels designed to be more than a visual experiment, reading through their landing page is interesting. In particular, their section on accessibility seems to contradict this post's opening premise:
> Screen readers get the real words. A screen reader reading down the page is never handed scrambled text, and our NVDA test asserts exactly that. Screen review and touch exploration are untested. ShieldFont hides shielded passages from accessibility tools by default with aria-hidden="true", because a decoy read aloud is fluent, grammatical, wrong English, and that is worse than silence.
> The real words remain sealed in the same page, and a visible notice above the block carries the control that uncovers them. It is on by default and reachable by mouse, keyboard and screen reader alike. Pressing it sets the reader’s browser to solving a compute-heavy puzzle: JavaScript and a few seconds of processing, more than most mass scrapers are willing to spend. That puts the words within reach of a screen reader, a translator and copy/paste.
I'd love to hear your thoughts on that. Also just aside, love this TUI-esque blog design and color palette (maybe a bearblog theme? still worth an upvote)
When you look at their live demo (https://shieldfont.org/demo/), it says: If you use a screen reader, custom font, or translator, please uncover the text before reading.
They also actively block copying the text, telling you to "uncover" the text first. The uncover operation is VERY expensive.
Anyone using assistive technologies or trying to copy "protected" text is SOL.
Search engines will index the decoy. You'll get no traffic. Their solution is basically putting yourself in a black hole.
> Search engines will index the decoy. You'll get no traffic. Their solution is basically putting yourself in a black hole.
It's only search engines? And who uses those anymore anyway? Other bots?
It doesn't affect organic traffic, so you are not really putting yourself into a black hole. A lot of website are driven by social media and organic traffic, so they would be just fine with this approach.
Isn't this the same sort of cost CloudFlare and anime catgirls are making us pay daily? Only enough to deter bots, or "a few seconds processing."
I take your point that one extra click/interaction required for screen readers only is objectively more friction, but it's only by exploring these technologies instead of dismissing them that we'll arrive at UX solutions truly work for people of all stripes (and ideally, not for bots, scrapers, and the like). I think if you're applying a tool like this you already have very different priorities than fueling Google.
As soon as you need to put a "decode" button for accessibility functions to work, you're effectively posting the key along with the cipher. It's self-defeating because any tool that supports screen readers will also support scraping. It's a fool's errand.
Accessibility isn't optional, so what do you propose?
It's funny, I was just thinking that the one thing I hate most about the terminal is that its monospaced fonts and overly-long lines are a nightmare for reading. So the fact that somebody designed a blog reading experience to mimic this is just… ugh for me.
But you seem to appreciate it, and I'm sure others do too. Different strokes for different folks.
Are these even meant to be used though? It seems more like performance art.
With this kinda of stuff it's hard to tell whether the person is doing it unironically, or knows it's "performative art". A while ago there was a trend of using a tool which imperceptibly perturbs an image in a way that supposedly breaks AI training on it. Of course, artists ate it up, despite the skepticism from AI researchers. Same with people setting up their sites to be "AI scraper traps", generating gibberish content. Probably also trivial to filter out, but people do it.
The problem with being dumb satirically is dumb people look up to you as a thought leader.
"caveman speak" skill, need I say more?
People aren't particularly bright. That's why the scientific method was developed to counteract our built-in tendency for... Unorthodox approaches
The irony of championing accessibility using low-contrast simulated VGA text...
Also the notion that the sides of the pages flash as you scroll due to simulating that old Macintosh monochrome monitor dithering effect. My eyes.
Where do you see the low contrast?
Anti-AI fonts seems like scrambled porn on cable back in the 1980s.
It also mostly don’t work.
Yeah, it seems like most of these would probably be easier for an AI to read than a human once you give even a tiny bit of training to the AI.
there is a reason nobody uses text based captchas anymore.
Wouldn't a font that shuffled the codepoint-to-glyph assignments be more effective?
That wouldn't make the decoy text look plausible to the AI.
There could be benefits unlocked in legal documents by retaining a machine-readable version and distributing the obfuscated version with a legend at the top. We proposed one that said:
"This document contains mitigations against review by automated systems. Recipients should ensure that they have read the contents on screen or in print. Recipients with bona fide vision impairments may be entitled to unmitigated documents upon request."
In testing, obfuscating small portions of text slipped under the radar of most (then-)frontier LLMs.
We used a font that was rendered on the fly and reported faulty or fake Unicode mappings: https://tritium.legal/blog/noroboto but others have proposed and done the same with ligatures.
There was a story once about a boy with a wheelchair who needed a ramp to get into school, and the school made him use the loading dock ramp used for garbage and other things at the back. The school argued that it was an appropriate accommodation.
Accessibility is not accessible if you need to go through extra steps to get it.
Cerebrally, this as a solution makes sense. But if you know anyone with a vision or other impairment, gating it behind a request is not only cruel but gets within dangerous striking distance of an ADA lawsuit, for general applications.
Maybe in the legal field or specific niche cases it's possible. But this would represent a major step backwards in the work we've done lowering barriers for a population whose only difficulty in accessing common resources is because they were born, or got sick, differently than anyone else.
My dad caught paralytic Polio at age 2 and has had limited mobility his entire life, so I'm familiar with that issue.
Our internal, hypothetical use-case was between contracting parties who were looking to avoid terms escaping into the wild. This shouldn't show up in standard ToS or similar. There are already really good legal reasons for that.
> Accessibility is not accessible if you need to go through extra steps to get it.
That seems a bit entitled to me, especially in the story you shared. They had access to the school just like everyone else. Why does it have to be in the exact same spot? Surely they could be dropped off by the loading dock as easily as others could be dropped off out front, and maybe even moreso. Should the wheelchair accessible stall be the first stall in the bathroom so they don't have to go through "extra steps" to get to it? As long as the ramp had the proper gradation to satisfy the ADA, I don't see the problem.
> They had access to the school just like everyone else
They literally did not. They had access to the school from a different entrance than everybody else. Specifically, one intended for cargo before people.
They also literally cannot have the same access like everyone else, since they're in a wheelchair. Everything will be different.
That's a different sense of "just like". Not "in the same manner", but "to the same extent".
Would it have been different, if they had everyone take the cargo entrance?
Yes. It's about discrimination and dehumanizing everyday cruelty.
It is not sufficient to work against current AI. It needs to also work against AI that has been trained by a competent team aware of your mitigation. Or worse, a competent developer with no particular AI skills.
Otherwise, you are relying on obscurity, and will lose as soon as you become interesting enough to matter.
You will also break non-AI machine processing use cases. That isn't just accessibility, it is things like search.
Yep
You will surely not have a good time enforcing the terms of a legal document that explicitly spells out that it is intentionally obfuscated from the party it intends to bind.
No, that’s not at all what is going on here.
I don't foresee anti-AI fonts being widely adopted. I see them largely as the symbolic saber rattling of intellectual property trolls.
Actually, pretty sure it was an art performance
Every AI font proponent I've seen has been a fanfiction writer who just doesn't like AI stealing their shit to use against them.
Is there evidence shieldfont doesn't work?
It only works while it's rare. If it was more common, scrapers would switch to OCR or simply reverse the font so they can decode the ligatures.
Their own whitepaper brings up a bigger issue: if it works, it poisons search engines as well.
OCR is a more expensive option, right? I don't think these fonts need to stop ai from being trained; I think they just need to make it more expensive and difficult.
More expensive and not worthwhile are two different things. Also when certain implementations become popular it's much more likely someone will write a very efficient kernel for decoding said text making it much less expensive than generic OCR.
Also more expensive doesn't mean that something won't happen, only the dynamics of how it happens. For example if you put all your documents in images then some service might just sell the AI providers the text. That service may do underhanded things like bundle OCR in an app that does something else and use your phone to get the text out of these images all day.
It's slightly more expensive, but the cost would be worth it if this became common. It's not designed to be hard to OCR, it's designed to be hard to copy out of the source code, so it doesn't require a very advanced OCR system (and all the regions that would need OCR-ing are clearly marked).
DRM for your eyes == garbage idea.
I believe the cat and rat game is already there, for multiple places, spam, captchas and now for AI content, yes, it's objective is to make it harder for AI companies to get such data, if it wastes their time, it's a win.
> it's objective is to make it harder for AI companies to get such data, if it wastes their time, it's a win.
That's only one half of the equation, though, isn't it? What if it makes it harder for legitimate users as well? It seems there's a balance to be struck.
Kind of funny how you get downvoted for a rational take, but one that's not anti-ai.
You'd ask that same person how much they like captchas and I'm sure they'd think their a terrible idea and they've ran into all kinds of issues with them.
Not everything that annoys or inconveniences you is harmful, as if this needs to be said to an adult.
We are now in a weird middle ground where we want to write things OCR algorithms have trouble transcribing which also means we write things people with accessibility issues have trouble seeing. No child left behind?
It's like the old tale about the national park bin with the smartest bear / dumbest tourist crossover, except we're now comparing capabilities of the smartest AI with disabled humans.
It already was a major issue in 2010 - home desktop-grade OCRs could easily beat an average grandma on reading heavily garbled text.
> in 2010 - home desktop-grade OCRs could easily beat an average grandma
I’ve heard Tesseract OCR recommended, but even in 2026, no matter how I scan receipts or documents, the OCR output seems far worse than human reading?
I think the difficulty was specifically with CAPTCHA challenges, which had to be quick to generate but still legible - OCR on physical documents has to be robust to a different set of problems
That said, what kind of errors are you getting? I think the main practical difference between Tesseract's LSTM-based OCR and newer VLM-based OCR like PaddleOCR [1] is (hopefully) getting to skip making heuristics for the layout of the document, but errors possibly compounding over multiple tokens - are you getting individual illegible words or having the documents smooshed together because the OCR can't parse the layout?
[1] https://github.com/PaddlePaddle/PaddleOCR
Well, I used specialized software specifically for solving CAPTCHA on select few file sharing websites. My later experience with general purpose OCR software was just as awful as yours. But it's a product problem, not technology problem - you really don't need an advanced AI for it, you just need a good implementation of traditional shape-matching OCR that doesn't seem to exist anywhere for some reason.
Is it part of the joke that the site is intentionally over-pixelated while the author critiques readability? (edit: I don't mind esoteric design and I play old games. I found it funny to see a blog with aesthetics that are not optimized for long reading to complain about the readability of fonts)
I don't think any anti-AI font design is more than design-as-art statements against AI. If there is evidence of them being used exclusively for business and without accessible fonts aside them, I'm willing to be wrong.
https://en.wikipedia.org/wiki/Turbo_Vision
I'm aware, and my comment isn't meant to be a challenge to anyone's sacred honor. I was observing a juxtaposition.
Not trynna to be funny. The author's post is anti-human, thus useless and harmful (medically).
I can't read this bad font, sizing, spacing, etc. The main offender is the color choice, and fonts that are just god aweful to read.
Accessibility is important, and I find that "Reader Mode" in most browsers is quite good. Everyone should have access to the tools to consume content. Did your browser not provide that functionality?
basically make it look hard to read for the majority for the sake of 1%? let those 1% use a diff tool view instead of 99% of us having to suffer.
That website no-way is accessible for my 30 year old eyes.
Found it awful to look at, tried to zoom out and everything on the page got larger. Seriously gross accessibility.
I'm a bit frustrated by what seems to be a widespread strong negative reaction to anti-AI fonts. The accessibility problem is real, but I feel like that's a reason to push the investigation deeper for solutions to that problem, not a reason to abandon the effort entirely. The largest intellectual property infringement in the history of the universe is actively unfolding, and it's resulting in an existentially threatening transfer of wealth and power. That's a problem worth exploring every solution for, and solving it may entail some serious sacrifices.
It's simply that a font won't solve anything. That concept is worst than security by obscurity, as it causes more problem and add more constraints than what it solves... for a very limited time until AI bots are adjusted to decode those fonts properly.
https://en.wikipedia.org/wiki/Generative_adversarial_network
Every half assed means of trying to confuse an AI is just a small bit of learning away from making the AI better than you.
Worse when you have people with disabilities, which I seem to be this week, you just make doing things a pain in the ass.
What I don't get is people like you think there is a solution to this. There is not. The harder you try you either exclude more actual humans or you align the AI closer to how people actually see.