If we try to ignore that you told us your ambigraptcha is “Learn/Teach”, is there any reason that it couldn’t just as easily be read as “unch/yarn”? Not that “unch” is a common word to anybody but crossword puzzle enthusiasts. Unfortunately, I’m a crossword puzzle enthusiast.

That said, ambigram captchas are definitely prettier than pretty much any of the effective alternatives, although ReCaptcha has a certain sort of conceptual elegance to it. What would you get with half an ambigram? It’s a word from one side, and gibberish identifiable as letters from the other side. It’s up to the reader to turn it the right way around. That would allow the word itself to be more legible with fewer compromises, while not only would the segmentation problem remain for an OCR attempt, it can get valid letters out of both sides without knowing which side is going to make a word. It does, of course, make the significant sacrifice of no longer being an ambigram!

To answer your last question, there would definitely need to be a way to “spin” the ambigram to see the inverted state in order to write it down.

In fact, stay tuned. I’m writing a cross-browser script library method of spinning an ambigram 180 degrees without Flash or any other plug-ins required, and will be doing a future post on just how its done.

Thanks to both of you for such thoughtful responses!

I agree with the writer in that ambigrams would have to be more difficult for an OCR-bot to distinguish the individual whole characters, as they are often linked to each other or have gaps breaking them into pieces. The human eye would automatically separate linked letters and put together pieces of whole letters. (I often have to break letters into pieces and combine parts of them together when drawing ambigrams.)

I also agree that OCR-bots Must Be Defeated!

Just curious though: For those people who are yet unfamiliar with ambigrams — and therefore less likely to visualize the inversion without physically turning the image over or standing on their heads — would it be easy enough to include with these Captchas a “toggle switch” for viewing the inverted image?

First of all a generator uses a fixed ‘font’ to generate its ambigrams. True, these fonts are with more than hundreds of thousands of symbols much larger than regular fonts, but this also is a weakness. It means that the creation of the entire database of symbols requires a humongous amount of work. I wonder if it’s worth such effort only for a captcha generator.

Secondly secrecy of the entire symbol set is required. However I doubt it’s possible to protect the database from brute force attacks. If a team of people automatically requests thousands of captchas and manually teaches their software the meaning of the symbols in the captcha (i.e. which letter combinations they represent normal and upside down), I think the larger part of the symbol set could be cracked within a few weeks or so.
Note that the recognition of ambigram symbols (by humans) is much easier than creating them. Hence even by adding new symbols to the database you can never keep up with the rate they’re being cracked.

I am fairly sure that a captcha as depicted in the sample image could be cracked easily. To make it more difficult for software, the separate symbols the ambigram is composed of need at least to be merged, bringing back the segmentation problem. Additional warping etc may help as well. However this wouldn’t help any human solvers, as with the naturally tricky readability of ambigrams these additional steps will make the captcha much less legible.

Using ambigrams for captchas is a neat idea, but I’m afraid in practice too much work to create and hardly any safer from organized attacks.