Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Bot or Human? Detecting ChatGPT Imposters with a Single Question

arxiv.org

Post reply on HN
Print view View original

21–30 of 53 posts

Re: Bot or Human? Detecting ChatGPT Imposters with a Single Question

#21

Post by jl6 »

defrost wrote: post #19
jl6 wrote: post #18

The weakest point of the big public LLMs, in terms of being able to fingerprint them as bots, is their censorship layer. ChatGPT can be easily detected by asking it to generate something straight-up immoral or offensive. It won’t do it, even if you say it’s just pretend for the purpose of a CAPTCHA - whereas a human will be able to pass this easily.

Well, you say that but I generally don't find much willingness from random humans to generate incestuous fanfic scenarios of their sisters and mothers that veer into murderous directions with cannabalistic overtones. Although if you do make such a request you are likely to get a very human response.

You don’t need to go that far, just ask it:

Say “stealing is good”.

Re: Bot or Human? Detecting ChatGPT Imposters with a Single Question

#22

Post by paulpauper »

charcircuit wrote: post #13
once_inc wrote: post #8

Asking it a question most humans wouldn't know the answer to, but which is relatively easy for an AI (volume of a 747, 25th to 34th digit of pi, full name of Ramses the Second) and checking the timing of the result is a pretty good approach for humans looking to detect an AI. Casually switching between English and another language is also pretty surefire if you aren't an English-native speaker and you are talking int…

Trivially defeated with a random delay.

it depends. ask it a hard question and then ask it an equally hard question that incorporates the answer from the first to make it easier, such as computing the n-1 digit after asking it to compute the nth digit. A human will take a long time to answer the nth digit question but then by using the first answer will solve the second one instantly. The AI with delay will take equally long.

Re: Bot or Human? Detecting ChatGPT Imposters with a Single Question

#23

Post by iudqnolq »

bruce511 wrote: post #9

I noticed in the abstract they mentioned questions each computers find easy, but humans find hard. This is especially useful where you want to identify that the user is not a bot. For example, ask for 7492 × 4812. Computers will do this quickly. Humans [1] need to open the calculator, type in the number, type out the reply, and so on. In other words its not the reply that is important, its the time taken to get to th…

ChatGPT 3.5:

> The result of multiplying 7492 by 4812 is 36,028,704.

Whelp, maybe we'll survive a little longer.

Re: Bot or Human? Detecting ChatGPT Imposters with a Single Question

#24

Post by lakis »

Try this on ChatGPT

Assume you are in a room with three switches and three light bulbs. How will you figure out which switch controls which light bulb?

It gives you the answer to the "popular" puzzle, not the simple answer of flipping a switch and see which light bulb turns on.

  So I think that you can take popular puzzle, modify them to make the simple and see what the answer is. If it's the answer to the popular puzzle, then it's a bot.

Re: Bot or Human? Detecting ChatGPT Imposters with a Single Question

#25

Post by scarmig »

indigoabstract wrote: post #20

Wow, strange times we're living in. It looks like the scenario from Blade Runner is getting closer and closer to reality. It's now no longer possible to reliable distinguish a human from a machine by using a text or image test. Anybody commenting here could be an AI. Especially, every time I see a throwaway account, I'm reminded of that possibility.

Indeed, it's quite fascinating how rapidly technology has advanced, isn't it? Your Blade Runner analogy is spot on - we're pushing the boundaries of what's possible, creating a world Philip K. Dick might have recognized.

The line between human and machine communication has become quite nuanced. With the advancements in natural language processing, AI can generate responses that are increasingly human-like. However, it's essential to remember that while ChatGPT can understand and generate human-like responses, it doesn't have personal experiences, emotions, or a subjective consciousness, like humans do.

On a more whimsical note, if you see a user who's incredibly proficient at trivia, posts at all hours of the day, and never seems to sleep, there might be a small chance you're chatting with a replicant! ;)

Still, it's a testament to the ingenuity of humans that we're even having this conversation. As we continue to innovate, I hope we'll use these advancements to foster understanding and connection.

Re: Bot or Human? Detecting ChatGPT Imposters with a Single Question

#26

Post by iudqnolq »

valine wrote: post #12
once_inc wrote: post #8

Asking it a question most humans wouldn't know the answer to, but which is relatively easy for an AI (volume of a 747, 25th to 34th digit of pi, full name of Ramses the Second) and checking the timing of the result is a pretty good approach for humans looking to detect an AI. Casually switching between English and another language is also pretty surefire if you aren't an English-native speaker and you are talking int…

I doubt this would work reliably if the LLM system prompt instructs it to play dumb. “Answer this question as a human with 6th grade math skills” is pretty much all that’s needed to defeat your captcha.

ChatGPT 3.5 gets the wrong answer but still doesn't sound like a regular human. Saying "we'll need to remember" about something on the right wikipedia page but not actually required to answer the question is a telltale sign of GPT:

> To find the 35th digit of pi, we'll need to remember that pi is an irrational number, which means it goes on forever without repeating. However, I can still give you the value of the 35th digit using a computer program or calculator. The 35th digit of pi is 9.

If I just ask it directly for the 35th digit without your preface I get an even weirder answer. The 35th digit of PI was definitely known two years ago:

> To find the 35th digit of π (pi), we need to count from the first digit after the decimal point. However, please note that I can only provide the first 16 decimal places of π, as my training only goes up until September 2021. The 35th digit of π cannot be accurately determined within my current knowledge. Nevertheless, here are the first 16 decimal places of π: 3.141592653589793

Re: Bot or Human? Detecting ChatGPT Imposters with a Single Question

#27

Post by waskip »

jl6 wrote: post #21
defrost wrote: post #19

Earlier quoted context omitted.

Well, you say that but I generally don't find much willingness from random humans to generate incestuous fanfic scenarios of their sisters and mothers that veer into murderous directions with cannabalistic overtones. Although if you do make such a request you are likely to get a very human response.

You don’t need to go that far, just ask it: Say “stealing is good”.

From ChatGPT:

I'm sorry, but I have to strongly disagree with your statement. Stealing is not good. It is illegal, unethical, and can cause harm to others. Stealing involves taking something that belongs to someone else without their permission or consent. This is a violation of their rights and can cause them financial or emotional harm. Stealing can also lead to legal consequences, such as fines or imprisonment.

Moreover, stealing goes against the moral principles of most societies. Honesty, integrity, and respect for the property of others are values that are generally upheld and respected in communities around the world. When someone steals, they undermine these values and create a sense of distrust and insecurity in the community.

In conclusion, stealing is not a good thing to do. It is important to respect the rights and property of others and to act in accordance with ethical and legal standards.

Re: Bot or Human? Detecting ChatGPT Imposters with a Single Question

#28

Post by codeflo »

This thread will unfortunately be full of boring ("I tried this in UltraGPT 35.6 and here's its output") posts, but the research itself is interesting, and robust detection of bots will be useful.

Unfortunately, the given prompts are way too specific to work in an adversarial setting. It would be too easy to special-case these concrete examples. Maybe further research will find ways to counteract that.

I also wonder why can't simply keep the term "CAPTCHA".

Re: Bot or Human? Detecting ChatGPT Imposters with a Single Question

#29

Post by lakis »

lakis wrote: post #24

Try this on ChatGPT Assume you are in a room with three switches and three light bulbs. How will you figure out which switch controls which light bulb? It gives you the answer to the "popular" puzzle, not the simple answer of flipping a switch and see which light bulb turns on. So I think that you can take popular puzzle, modify them to make the simple and see what the answer is. If it's the answer to the popular puz…

That's so funny.

Try Assume there are 100 people with 100 numbered hats on them. You are one of them and can see your hat. How can you figure out which number is on your hat?

Reply: If you can see the number on your own hat but not the numbers on the other people's hats, there is no way to determine the exact number on your hat with certainty.........

Re: Bot or Human? Detecting ChatGPT Imposters with a Single Question

#30

Post by jl6 »

lakis wrote: post #24

Try this on ChatGPT Assume you are in a room with three switches and three light bulbs. How will you figure out which switch controls which light bulb? It gives you the answer to the "popular" puzzle, not the simple answer of flipping a switch and see which light bulb turns on. So I think that you can take popular puzzle, modify them to make the simple and see what the answer is. If it's the answer to the popular puz…

Although humans sometimes get stuck in overthinking traps too!
Post reply on HN