Deepseek and Chatgpt can't solve simple newspaper crytogram.

[Marc51]

There is a lot of hype about AI taking over the world. I asked the above listed sites to solve an easy puzzle and it didn't get any letter right! It looks like they were trying to match quotes found in the literature. It also tried frequency analysis to no avail. Here is the puzzle:

YXOP TXZWT KNGMKTT SUK CMKHFT HZC GMHLKMT HZC UXGKT XY SUK EXMPAZW GKXGOK.
— UHZP EAOOAHFT

[kb83]

Folk songs express the dreams and prayers and hopes of the working people. Hank Williams
"people" always helps me get started.
I have to think that AI will get there to solving crypto's, only a matter of time.

[cc_frazer]

Chat GPT is really bad at cryptology. I wondered if it was deliberately made like that to stop it being used by hackers to break data encryption

[cc_frazer]

Chat GPT can code a puzzle solver GUI though. It doesn't solve the puzzle for you but lets you change the letters and solve it yourself. I made one on github but it seems I cant share it on here. It has some useful features like letter frequency counter and when you replace one letter it auto replaces all that letter in the cypher.

[Brewer]

ChatGPT can't even solve simple logic puzzles. It's unable to satisfy more than about 4 clues at the same time.

[Anachronismatic]

Cryptograms are a warm-up to the concept of cryptography, but if data were stored in that fashion, it would be so easily stolen that no one would bother even trying to store sensitive data.

Well-implemented cryptographic algorithms store data in what looks like completely random blocks of hexadecimal data (256-letter "alphabet" rather than 26). There are no patterns to find.

The German "Enigma" device used in World War II (and cracked by a team of experts in England) is an example of a more sophisticated letter substitution. It's rudimentary compared to modern cryptography, but the concept (create what looks like a random string of letters, no patterns to analyze) is solid. It was cracked because some of the Germans who used the code were quite sloppy in implementing it, allowing the team to reverse-engineer (on a good day) the set of wheels used for encoding. No one sat there looking at messages, deducing what was going on.

ChatGPT (and other LLMs) are not code-breakers because they analyze text and predict how a response should look. It's not surprising they take a rudimentary approach to cryptograms - that approach would eat up too much time if applied more thoroughly. However, if the Baron's set of quotes were added to the training database and identified as such, ChatGPT (or any first-year coder's programming attempts) could solve anything instantly.

There are far more sophisticated tools out there. I wrote one a while back for logic problems based on how the Baron phrases clues. I decided not to use it to set records because that seems like a terrible thing to do to the person who put in all the work to create this site.

Using ChatGPT as a coder to create one of those tools is possible. Since I am a coder, I wouldn't, but as long as you understand GIGO (garbage in, garbage out), it's viable for certain types of projects. I don't think my job is in any danger in the near future, though, because writing code that seems to work is only part of the job. Maybe some day.