Do you generally trust people who speak confidently, without indecision, with not a word out of place? Of course you do. We’re conditioned to it. But some people lie. And this is one of the challenges with a large language model (#LLM) like #ChatGPT. Among all their stunning capabilities—and they are a mind-blowing achievement—the capacity to #lie confidently is one of the most impressive. Part of their confidence, even when they’re wrong, is the perfection of the writing mechanics. ChatGPT spells correctly, and gets the grammar right. It uses subordinate clauses. It punctuates well. It writes in structured paragraphs! And it’s fast. This is great, but it can lull the user into forgetting it’s a machine that very often makes mistakes. Fortunately, there are countermeasures you can employ to detect these lies. I cover many of them in this article: https://lnkd.in/gZmuxNYd This is an in-depth article—and beautifully formatted 😀—so I can't reproduce it here, but these are the tactics I cover: • trust no one • look it up • use logic • ask follow-ups • pit LLMs against one another I give examples from ChatGPT, #bingchat, Google's #bard, and #chatsonic. If reading's not your bag and you want an 8-word tl;dr: the most important thing is to apply #skepticism. #ai #ml #artificialintelligence #machinelearning #bots #writing #largelanguagemodels #beware #logic Image credit: #bluewillow (another amazing machine that can confidently lie)
Glen Ford
1y
Tagging some folks who saw me give this presentation at #productcampaustin 28: Drew Bixby, Roger Cauvin, Jason Carrier, Dan Corbin
• Skip Balch
1y
Well, color me impressed! It seems our beloved ChatGPT has mastered the art of confidently lying, just like a smooth-talking politician. 😉 Apparently, we should put on our detective hats and question every word it says, like a cross between Sherlock Holmes and a skeptical grandma. Who knew that impeccable grammar and perfect punctuation could be so deceiving? 🤔 But fear not, fellow humans! With the tactics provided in this article, we shall fight the machine and unveil its lies. Just remember: trust no one, not even this comment! 😜 #SkepticalYetAmused #TrustIssues #ChatGPTDetectives #AIConfidenceGames #SarcasmAndSkepticismUnite
Dan Corbin
1y
Fantastic presentation!
Collaborative, startup-bred product leader, driven by empathy ❤️ — 🤖 AI/ML for 6 years
1yDrew Bixby I think it was you who brought up some points about ChatGPT's use of links as citations, or maybe Roger Cauvin. Since the talk I found a better example that you can see in the Tactic 4 section: https://austinproduct.pro/blog/how-to-catch-chatgpt-lie#tactic-4 It gave me one bad link and one good one. I still don't know why it gets it right sometimes, but it does seem to.