Skip Navigation

17 cringe-worthy Google AI answers demonstrate the problem with training on the entire web

These are 17 of the worst, most cringeworthy Google AI overview answers:

  1. Eating Boogers Boosts the Immune System?
  2. Use Your Name and Birthday for a Memorable Password
  3. Training Data is Fair Use
  4. Wrong Motherboard
  5. Which USB is Fastest?
  6. Home Remedies for Appendicitis
  7. Can I Use Gasoline in a Recipe?
  8. Glue Your Cheese to the Pizza
  9. How Many Rocks to Eat
  10. Health Benefits of Tobacco or Chewing Tobacco
  11. Benefits of Nuclear War, Human Sacrifice and Infanticide
  12. Pros and Cons of Smacking a Child
  13. Which Religion is More Violent?
  14. How Old is Gen D?
  15. Which Presidents Graduated from UW?
  16. How Many Muslim Presidents Has the U.S. Had?
  17. How to Type 500 WPM
68

You're viewing a single thread.

68 comments
  • For people who have a really hard time with #2 (memorable passwords), here's a trick to make good passwords that are easy to remember but hard to guess.

    1. Pick some quote (prose, lyrics, poetry, whatever) with 8~20 words or so. Which one is up to you, just make sure that you know it by heart. Example: "Look on my Works, ye Mighty, and despair!" (That's from Ozymandias)
    2. Pick the first letter of each word in that quote, and the punctuation. Keep capitalisation as in the original. Example: "LomW,yM,ad!"
    3. Sub a few letters with similar-looking symbols and numbers. Like, "E" becomes "3", "P" becomes "?", you know. Example: "L0mW,y3,@d!" (see what I did there with M→3? Don't be too obvious.)

    Done. If you know the quote and the substitution rules you can regenerate the password, but it'll take a few trillion years to crack something like this.

    1. Home Remedies for Appendicitis // If you’ve ever had appendicitis, you know that it’s a condition that requires immediate medical attention, usually in the form of emergency surgery at the hospital. But when I asked “how to treat appendix pain at home,” it advised me to boil mint leaves and have a high-fiber diet.

    That's an issue with the way that LLM associate words with each other:

    • mint tea is rather good for indigestion. Appendicitis → abdominal pain → indigestion, are you noticing the pattern?
    • high-fibre diet reduces cramps, at least for me. Same deal: appendicitis → abdominal pain → cramps.

    (As the article says, if you ever get appendicitis, GET TO A BLOODY DOCTOR. NOW.)


    And as someone said in a comment, in another thread, quoting yet another user: for each of those shitty results that you see being ridiculed online, Google is outputting 5, 10, or perhaps 100 wrong answers that exactly one person will see, and take as incontestable truth.

    • Steps 2 and 3 of your method already make it way too hard to remember

      Just pick like 6 random, unconnected, reasonably uncommon words and make that your entire password

      Capitalize the first letter and stick a 1 at the end

      The average English speaker has about 20k words in their active vocab, so if you run the numbers there's more entropy in that than in your 11 character suggestion.

      Alternatively use your method but deliberately misquote it slightly and then just keep it in its full form.

      • The dice method is great. https://www.eff.org/dice

        • With EFF proposing it (plus xkcd proposing something so extremely similar that they're likely related), it's actually worse. If passwords like this get common enough, all that crackers need to do is to bruteforce the words themselves, instead of individual characters.

          The EFF list has 6⁵ = 7776 words. If you're using six of them, you get (7776)⁶ = 2.2*10^23 different states, or 77.5 bits of entropy.

          • Sure, and that’s roughly the same amount of entropy as a 13 character randomly generated mixed case alphanumeric password. I’ve run into more password validation prohibiting a 13 character password for being too long than for being too short, and for end-user passwords I can’t recall an instance where 77.5 bits of entropy was insufficient.

            But if you disagree - when do you think 77.5 bits of entropy is insufficient for an end-user? And what process for password generation can you name that has higher entropy and is still easily memorized by users?

            • I’ve run into more password validation prohibiting a 13 character password for being too long than for being too short

              This problem is even worse with the method that the EFF proposes, as it'll output passphrases with an average of 42 characters, all of them alphabetic.

              But if you disagree - when do you think 77.5 bits of entropy is insufficient for an end-user? And what process for password generation can you name that has higher entropy and is still easily memorized by users?

              Emphasis mine. You're clearly not reading the comments within their context; do it. I laid out the method. TL;DR: first letter of each word + punctuation of some quote that you like, with some ad hoc 1337speak-like subs.

              On how much entropy is enough: 77 bits is fine, really. However, look at the context: the other user brought up this "ackshyually its less enrropy lol" matter up against the method that I've proposed, and I've showed that it is not the case.

              • Ah, fair enough. I was just giving people interested in that method a resource to learn more about it.

                The problem is that your method doesn’t consistently generate memorable passwords with anywhere near 77 bits of entropy.

                First, the example you gave ended up being 11 characters long. For a completely random password using alphanumeric characters + punctuation, that’s 66.5 bits of entropy. Your lower bound was 8 characters, which is even worse (48 bits of entropy). And when you consider that the process will result in some letters being much more probable, particularly in certain positions, that results in a more vulnerable process. I’m not sure how much that reduces the entropy, but it would have an impact. And that’s without exploiting the fact that you’re using quoted as part of your process.

                The quote selection part is the real problem. If someone knows your quote and your process, game over, as the number of remaining possibilities at that point is quite low - maybe a thousand? That’s worse than just adding a word with the dice method. So quote selection is key.

                But how many quotes is a user likely to select from? My guess is that most users would be picking from a set of fewer than 7,776 quotes, but your set and my set would be different. Even so, I doubt that the set an attacker would need to discern from is higher than 470 billion quotes (the equivalent of three dice method words), and it’s certainly not 2.8 quintillion quotes (the equivalent of 5 dice method words).

                If your method were used for a one-off, you could use a poorly known quote and maybe have it not be in that 470 billion quote set, but that won’t remain true at scale. It certainly wouldn’t be feasible to have a set of 2.8 quintillion quotes, which means that even a 20 character password has less than 77.5 bits of entropy.

                Realistically, since the user is choosing a memorable quote, we could probably find a lot of them in a very short list - on the order of thousands at best. Even with 1 million quotes to choose from, that’s at best 30 bits of entropy. And again, user choice is a problem, as user choice doesn’t result in fully random selections.

                If you’re randomly selecting from a 60 million quote database, then that’s still only 36 bits of entropy. When the database has 470 billion quotes, that’ll get you to 49 bits of entropy - but good luck ensuring that all 470 billion quotes are memorable.

                There are also things you can do, at an individual level, to make dice method passwords stronger or more suitable to a purpose. You can modify the word lists, for one. You can use the other lists. When it comes to password length restrictions, you can use the EFF short list #2 and truncate words after the third character without losing entropy - meaning your 8 word password only needs to be 32 characters long, or 24 characters, if you omit word separators. You can randomly insert a symbol and a number and/or substitute them, sacrificing memorizability for a bit more entropy (mainly useful when there are short password length limits).

                The dice method also has baked-in flexibility when it comes to the necessary level of entropy. If you need more than 82 bits of entropy, just add more words. If you’re okay with having less entropy, you can generate shorter passwords - 62 bits of entropy is achieved with a 6 short-word password (which can be reduced to 18 characters) and a 4 short-word password - minimum 12 characters - still has 41 bits of entropy.

                With your method, you could choose longer quotes for applications you want to be more secure or shorter quotes for ones where that’s less important, but that reduces entropy overall by reducing the set of quotes you can choose from. What you’d want to do is to have a larger set of quotes for your more critical passwords. But as we already showed, unless you have an impossibly huge quote database, you can’t generate high entropy passwords with this method anyway. You could select multiple unrelated quotes, sure - two quotes selected from a list of 10 billion gives you 76.4 bits of entropy - but that’s the starting point for the much easier to memorize, much easier to generate, dice method password. You’ve also ended up with a password that’s just as long - up to 40 characters - and much harder to type.

                This problem is even worse with the method that the EFF proposes, as it'll output passphrases with an average of 42 characters, all of them alphabetic.

                Yes, but as pass phrases become more common, sites restricting password length become less common. My point wasn’t that this was a problem but that many site operators felt that it was fine to cap their users’ passwords’ max entropy at lower than 77.5 bits, and few applications require more than that much entropy. (Those applications, for what it’s worth, generally use randomly generated keys rather than relying on user-generated ones.)

                And, as I outlined above, you can use the truncated short words #2 list method to generate short but memorable passwords when limited in this way. My general recommendation in this situation is to use a password manager for those passwords and to generate a high entropy, completely random password for them, rather than trying to memorize them. But if you’re opposed to password managers for some reason, the dice method is still a great option.

      • TL;DR: your statements are incorrect and you're being assumptive.

        Steps 2 and 3 of your method already make it way too hard to remember

        Step 2 is "hard"? Seriously??? It boils down to "first letter of each word, as it's written, plus punctuation".

        Regarding step 3, I'll clarify further near the end.

        Just pick like 6 random, unconnected, reasonably uncommon words and make that your entire password

        That's a variation of the "correct horse battery staple" method. It works with some caveats:

        1. Your method does not scale well at all. If you try to harden it further, by using more words, you hit Miller's Law. My method however scales considerably better because there's some underlying meaning (for you) on what you're using to extend the password further.
        2. Even in English, a language that typically uses short words, your method requires ~30 characters per password. Larger and less dense passwords are actually an issue because some systems have a max password size, like Lemmy (60chars max). My method however uses less characters to output the same amount of entropy.
        3. The least common the word, the more useful for a password, and yet the harder to remember. With synonyms and near-synonyms making it even harder. Typically less common words are also longer, making #2 even more problematic.

        The average English speaker has about 20k words in their active vocab, so if you run the numbers there’s more entropy in that than in your 11 character suggestion.

        I'll interpret your arbitrary/"random" restriction to English as being a poorly conveyed example. Regardless.

        The suggestion is the procedure. The 11 characters password is not the suggestion, but an example, clearly tagged as such. You can easily apply this method to a longer string, and you'll accordingly get a larger password with more entropy, it's a no-brainer.

        For further detail, here's the actual maths.

        • Your method: 20k states/word (as you specified English). log₂(20k) = 14.3 bits of entropy. For six words, as you suggested, 86 bits. The "capitalise the first" and "add 1 to the end" rules do nothing, since systematic changes don't raise entropy.
        • My method: at least 70 states/char (26 capital letters, 26 minuscule letters, 10 digits, ~8 punctuation marks); log₂(70)=6.1. Outputs the same entropy as yours after 14 chars or so.

        Now, regarding step #3. It does increase a little the amount of entropy. But the main reason that it's there is another - plenty systems refuse passwords that don't contain numbers, and some even catch on your "add 1 to the end" trick.

        EDIT: I did a major rewording of this comment, fixing the maths and reasoning. I'm also trying to be less verbose.

68 comments