How can you tell if AI is lying to you?
Nastassia
32 replies
Do you have any methods to identify deception when working with AI, ChatGPT in particular? Have you noticed any areas where AI frequently provides false information? Are there any phrases or expressions that clearly indicate a mismatch with the facts?
Replies
Oleg Naumenko@naou
Scade.pro
I don't have any secret recipes, everything is more or less the same as vis-a-vis humans: if the other side states facts that are plainly incorrect according to my own knowledge. Or if I can just subtly feel that something doesn't add up. But, in my experience, unlike human beings, AI less often has flaws in logic and reasoning
However, once I've performed an experiment, I asked ChatGPT to write a little bit peculiar, not so generic code for me. First couple of attempts turned out to be a complete disaster. Then I tried to provide my AI assistant with test cases: you do X, you get Y as a result. At first it seemed that step by step we were moving to the victory. But after some steps ChatGPT started to forget what was discussed in the beginning. After 13th message I surrendered
So, what to do if it's something important? Check the facts, test your code, etc
Share
Scade.pro
Scade.pro
@mksaint13 thanks Max. Agree like in any research you have to check different sources to know what’s right
AINave
Yes, I observe this quite frequently nowadays even with ChatGPT 4. (proving popular theory of "ChatGPT is being nerfed" as true day by day).
I use ChatGPT most of the time for coding assistant and content generation. For coding, earlier, it used to give almost error free responses with bare minimum changes to be done by me. But now, I have to be particularly very cautious and don't easily believe what ChatGPT provides. I read the response thoroughly and verify (since I am a professional). So, that's how I deal with it.
For any factual information, I can't even rely on ChatGPT + Browsing mode. I use Perplexity AI which almost provides error free results every time. Nowadays, Google Search GenAI is providing good results as well.
Scade.pro
@ramitkoul That's an incredible power of symbiosis between artificial intelligence and an expert in the field. AI simplifies and speeds up your work, while you have the knowledge and skills to check the correctness of the code
@ramitkoul AI may provide inaccurate or incorrect information. This could be related to flaws in the training data or errors in the algorithms. If the information provided by AI seems dubious, it's best to verify it from another reliable source.
Scade.pro
In this case I'm looking for the information in Google😀
Scade.pro
@alexanderptitsyn don't you use Bard instead? Or does it also seem inaccurate?
I've noticed that ChatGPT's information isn't up to date beyond 2021, so it might not always be accurate. That's why I sometimes have to double-check things by Googling them.
Scade.pro
@keshini_dabare2 yep it really helps, but takes time :)
No real special method, but I try to use multiple AI platforms that draw from different sources. I find that Anthropic's Claude AI is pretty accurate and less likely to be nerfed, so I will sometimes compare questionable ChatGPT responses to Claude's. BARD seems to love to suggest answers from unproven journals and research documents and call it truth. I've found the most deception/mistakes in that platform. Lastly, I found that simply implying an answer is true in my own responses, will often be met with agreement, even though I know it was false. Seems like building your own context around a deception can mask it or get an agreement from AI; especially if it is a well known - easily mistaken topic.
Scade.pro
@reconcatlord Thanks, Nick. I haven't tried Claude yet, but it sounds interesting. It's quite disappointing that Bard is so unreliable.
The idea of tricking AI with false information spoken in a confident way seems like a great method to determine if AI is trustworthy or not. Or is this agreement common to all neural networks?
@reconcatlord Unfortunately, there are no specific phrases or expressions that would indicate a mismatch with the facts. It's important to always be critical of the information AI provides and remember that it may be inaccurate or incomplete.
@nastassia_k I'm not sure, when I had a hard fact that I knew to be true and AI gave me a incorrect answer I've had two separate responses when suggesting the correct answer, (1) a reiteration of the incorrect info, (2) and a "I'm sorry you're correct, I now see my mistake, thank you for your correction, it helps me train to be a better AI system", I'm paraphrasing but that was the gist of it.
@mksaint13 I've had some pretty blatant errors as AI responses that actually surprised me, wondering how that level of incorrect info could possibly be trained or sourced into the system. Another oddity was AI presenting info as if it was true but then admitting it was synthesized to represent the correct answer the system could not find. Ask BARD for the names of specific book chapters, it will give you a good list for most requests, yet they rarely are the real chapter titles. When questioned if these were the actual chapter titles or made up, the system will concede to manifesting them through journal articles and book reviews. It's a weird grey area of deception/incorrect info.
@reconcatlord I have a psychiatrist. When he tried to get different descriptions of medicinal substances through ChatGPT, he always laughed at the result.
You're absolutely correct. Training and correcting AI models are crucial steps in improving their performance and accuracy. AI models like ChatGPT learn from vast amounts of data during their training process, but they may not always provide perfect or up-to-date information.
😀 To enhance their performance, AI models need continuous training and feedback. Developers and researchers actively work on training these models on new data and incorporating user feedback to address inaccuracies and improve their responses.
Scade.pro
@helplook_101 thanks for sharing.
I wonder, who are the researchers involved in the development of AI models? Do they come from diverse scientific fields or are they predominantly linguists?
GPTBLOX-ChatGPT Save Data
A good way to ensure accuracy with AI is to ask the same question in different ways or verify its answers using different models/sources. This allows you to cross-check its consistency and reliability.
Scade.pro
@barryzheng568 Great method. Thank you! I'll give it a try
Scade.pro
One of the ways to check AI answers in this case is to throw the answer into another LLM for a quick check. For example, from ChatGPT to Llama and vice versa.
Scade.pro
In these cases, I look for information in different sources. For example, original research articles.
It's important to understand that artificial intelligence, including ChatGPT, doesn't lie in the conventional sense of the word. AI doesn't have emotions or intentions, so it cannot deliberately deceive. AI generates responses based on the data it was trained on.
Scade.pro
@mksaint13 Max, thanks for your interesting answers. This one is great as well!
AppSave by Appscribed
Interesting question. My take on this is two-fold:
1. You need to be reading a lot of content in the past(pre-chatgpt) and that's really engaging. You will ultimately train your mind on what good content is and written by mostly humans although there were AI models (GPT1 & 2) before chatgpt (GPT 3 & 4). Having done that (that's done over a long period of time), you will know what's written by AI when you see it. But again, you will have to read a lot of AI-generated content.
2. If you want to skip 1 & 2, tools like AI detectors and tools like zeropgpt can be used but believe me, AI can dodge another AI once it gets more mature.
While I am not against AI-generated content, if you are really smart, you can leverage the AI-generated content and upgrade it, and make it most engaging for your audience in the least possible time.
Makes sense?
I've shared this article with my colleagues, and we're already discussing how to incorporate the strategies mentioned here into our practices.
https://crackdeck.com/sigmakey-b...