While keeping an eye on the outputs from his business Lindy's AI assistants, Flo Crivello observed something odd. When a new client requested a video instruction from her Lindy AI assistant to help her better grasp how to utilize the platform, Lindy dutifully responded; this is when Crivello realized something wasn't quite right. A video lesson is not available.
"We noticed this and thought, 'Well, what kind of video did it send?' and then we realized, 'Oh no, there's a problem,'" Crivello said in aImagine Future interview.
The music video for Rick Astley's 1987 dance-pop hit song "Never Gonna Give You Up" was what the AI provided the client. Put more simply, the client was Rickrolled. by an artificial intelligence.
The bait-and-switch craze known as "rickrolling" dates back more than 15 years. When Rockstar Games posted the highly anticipated "Grand Theft Auto IV" trailer on their website, the site collapsed due to the overwhelming amount of traffic, which is what made the meme viral. To allow others to view the trailer, some individuals were able to download the video and upload it to other websites, such as YouTube. They then shared the URLs. However, one 4chan user chose to pull a practical joke by posting the link to Rick Astley's song "Never Gonna Give You Up." Even after seventeen years, individuals are still pulling practical jokes on their pals by playing the Astley song when it's not appropriate – the music video has had over 1.5 billion views on YouTube.
Because this internet joke is so common, big language models—like ChatGPT, which runs Lindy—inevitably noticed it.
According to Crivello, "the way these models work is they try to predict the most likely next sequence of text." "Oh, I'm going to send you a video!" is how it begins. What is then most likely to happen next? YouTube.com. What is most likely to happen next?
Crivello revealed to Imagine Future that Lindy had only twice rickrolled clients out of millions of responses. It was still necessary to fix the problem.
“The really remarkable thing about this new age of AI is, to patch it, all I had to do was add a line for what we call the system prompt — which is the prompt that’s included in every Lindy — and it’s like, don’t Rickroll people,” he said.
Given that AI models are frequently trained on vast portions of the internet, Lindy's error raises concerns about the extent to which online culture will be incorporated into these algorithms. The reason Lindy's unintentional Rickroll is so noteworthy is that the AI's illusion was fueled by an organic reproduction of this particular user activity. However, AI is not immune to internet comedy; Google discovered this the hard way when it used Reddit data to train its AI. Google's AI eventually informed a user that adding glue would improve the cheese's ability to attach to pizza dough because the platform is a hub for user-generated content, a lot of which is humorous.
"It wasn't exactly making stuff up in the Google case," Crivello stated. "It was content-based; the content itself was subpar."
Crivello believes that since LLMs advance quickly, mistakes like these won't happen as frequently in the future. Furthermore, Crivello claims that fixing these errors is now simpler than before. When Lindy first started out, one of its AI assistants would tell the user that it was working on a task, but it would never finish it. (Interestingly, that sounds rather human.)
"We had a lot of difficulty fixing that problem," Crivello remarked. However, we simply added a popup that said, "If the user asks you to do something you're not able to do, just tell them you can't do it," when GPT-4 was released. And it resolved the issue.
The good news is that the consumer who was tricked by Rickroll may not even be aware of it at this moment.
“I don’t even know that the customer saw it,” he said. “We followed up immediately like, ‘Oh hey, this is the right link to the video,’ and the customer didn’t say anything about the first link.”