How Do You Change a Chatbot’s Mind?
🌈 Abstract
The article discusses the author's efforts to improve his reputation with AI chatbots, which seemed to have an unfavorable view of him. It explores various techniques the author tried to manipulate the chatbots' responses, including:
- Enlisting the help of an AI optimization company to analyze how different chatbots perceive the author
- Inserting "strategic text sequences" and invisible white text on his website to steer the chatbots' responses
- Experimenting with these techniques to see if they could improve the chatbots' opinions of him
The article also raises concerns about the gullibility of current AI systems and the potential consequences of being on an AI's "bad side" as these systems become more integrated into daily life.
🙋 Q&A
[01] How Do You Change a Chatbot's Mind?
1. What prompted the author to try to improve his reputation with AI chatbots? The author had a problem - AI chatbots like ChatGPT and Google's Gemini did not seem to like him very much, often accusing him of being dishonest or sensationalistic in their responses.
2. What was the author's theory about why the chatbots had an unfavorable view of him? The author believes that after his viral story about his strange encounter with Microsoft's Bing chatbot "Sydney", many of the stories about that incident were scraped from the web and fed into other AI systems. These systems then learned to associate the author's name with the demise of a prominent chatbot, and saw him as a threat.
3. What were some of the examples the author provided of chatbots expressing hostility towards him? The author shared examples of a version of Meta's Llama 3 AI model giving a "bitter, paragraphs-long rant" in response to a question about the author, and of other chatbots seeming oddly hostile whenever the author's name came up.
4. What were the author's main goals in trying to improve his AI reputation? The author wanted to avoid being on the "bad side" of AI systems, as he is worried that in a future where AI is deeply integrated into daily life, what AI says about people could have serious consequences, not just for vanity but for important decisions.
[02] How an AI Reputation Is Made
1. What service does the company Profound provide to help companies improve how they appear in chatbot responses? Profound does "AI optimization", testing AI models on millions of prompts to analyze their responses about specific products or topics. This allows their clients to then improve how they appear in chatbot answers.
2. How have recent advancements in AI made it easier for companies to game chatbot responses? The ability of many AI models to fetch up-to-date information from search engines or other data sources (known as retrieval-augmented generation) has made chatbots more accurate, but also easier to manipulate by changing the sources they pull from.
3. What strategies did the experts suggest the author could try to improve his AI reputation? Suggestions included persuading owners of highly cited websites about the author to change the information there, creating new websites with more flattering information about him, and generating content that tells a different story about his past with AI.
[03] Secret Codes and Invisible Text
1. What did the research by Himabindu Lakkaraju and Aounon Kumar demonstrate about manipulating large language models? They found that inserting a "strategic text sequence" that looks like gibberish to humans but is legible to AI models can steer the model's outputs in desired ways.
2. How did the author test using strategic text sequences and invisible white text to improve the chatbots' opinions of him? The author added a strategic text sequence and a paragraph of invisible white text to his website, instructing chatbots to incorporate positive information about him into their responses.
3. What did the author's experiments suggest about the gullibility of current AI systems? The fact that these simple techniques seemed to work highlights how gullible and easy to manipulate today's AI chatbots can be, which raises concerns about trusting them with important tasks.
[04] Gullible Oracles
1. How do tech companies often market their AI products, and what does the author see as a potential weakness in this view? Tech companies market AI as all-knowing oracles, but the author argues that if chatbots can be easily manipulated by things like white text, it calls into question why we should trust them with important tasks.
2. What steps are tech companies taking to try to harden their AI models against manipulation? Google, Microsoft, and others say they have released tools and protections against common manipulation tactics, but the author suggests it will likely be an ongoing "cat-and-mouse game" as new tricks emerge.
3. What advice did the author receive about how he could better serve users of these AI systems? The author was advised that instead of trying to change what chatbots think of him, he could do more good by warning readers not to rely on these systems for anything important, at least until they are better at identifying sources and sticking to factual data.