Rickrolling AI: What It Reveals About Human Intelligence

Language models playing internet pranks

Some trends and language patterns are so deeply embedded in the language and culture of the internet that even the most advanced language models struggle to fully extract or contextualize them.

Humor, especially the internet’s unique and eccentric styles of jokes can help to explain the difference between human and machine intelligence. Apparently, large language models have a Rickrolling problem.

Rickrolling is an internet prank that's been an internet joke for decades. It uses hyperlinks to create a surprise connection. A link that appears to direct you to a meta-analytic research article on humor styles may actually redirect you to Rick Astley's 1987 hit "Never Gonna Give You Up."

Rickrolling and internet humor

This bait-and-switch is one of the most enduring internet pranks. It plays with the way technology (a hyperlink) is embedded in language. But it also is a subtle warning. Clicking a hyperlink on the internet may take you somewhere you didn’t expect. Rickrolling is a light and humorous way to remind people not to take everything online at face value.

There have been instances of large language models inadvertently Rickrolling users or chatbots pulling off a prank on unsuspecting clients of company chatbots.

This reveals much about human intelligence as well as the way language models work.

The culture of the internet is a vast repository of language, context, and subtlety. Simple jokes and cues require both an understanding of context and social situations. Human intelligence has evolved to quickly pick up subtleties from limited and ambiguous information.

Social functions of humor

Jokes and pranks convey information on multiple levels and serve social functions. On the surface, they can showcase the personality of the individual telling the joke, whether they’re easy-going, friendly, or mischievous. But jokes can accomplish other objectives: signaling levels of familiarity, setting boundaries for conversation topics, and even sending covert messages about power dynamics.

A joke that demeans a person or group often reveals more about the joke-teller’s social intentions and perceived status. Whereas a self-deprecating joke can cut through social divisions and formalities, serving as a way to defuse tension or bring people closer. In both cases, humor is a social tool—either to reinforce or challenge social hierarchies.

Rickrolling is a perfect example of how an internet joke can operate on multiple levels. While it’s often just a light-hearted prank, it also carries a subtle warning: On the internet, you can’t always trust what you see. This reflects a broader truth and warning about the online world: What you encounter might not always be what you expect.

Humor reveals layers of language and culture

But it’s also revealing about how our language shapes the internet. The web is a vast network of language connected into networks of relationships and meaning. A prank like Rickrolling is used so often that it is deeply embedded in the language and code of the internet.

Large language models trained on vast datasets harvested from the web cannot simply ignore these elements of digital culture. Rickrolling is so ubiquitous that it shows up in the output of large language models that have failed to decode the context or social meaning of the joke.

Human vs. machine intelligence

Artificial Intelligence Essential Reads

The Great AI Detection Dilemma

Are Large Language Models More Liberal?

Neil Lawrence, DeepMind Professor of Machine Learning at the University of Cambridge, explains the difference between human and machine intelligence in his book The Atomic Human. Lawrence deliberately uses the term ‘machine intelligence’, not 'artificial intelligence,’ to emphasize that machine intelligence is fundamentally different from how humans process information.

Human intelligence is not just about processing data; it’s deeply embodied and evolved to handle ambiguity, make quick decisions with limited information, and rely on situational cues and social context. Machine intelligence can process vast amounts of data extraordinarily quickly to mimic human outputs. Yet it lacks true understanding, embodied knowledge, or the capacity to focus on specific information based on subtle situational cues. That means human intelligence is uniquely suited for environments and situations where nuance and context are key. Humor is the perfect example.

This is why a machine might replicate a Rickroll without any real understanding of the joke. It’s a combination of trust, deception, and unexpected results that is so attractive to human intelligence.

A caution embedded in the context

Rickrolling is a harmless prank that hints at a greater potential security risk. The fact that this has been incorporated into machine intelligence is extraordinarily funny, but it is also a warning.

The culture and the language of the internet are now deeply embedded in everything that is produced and created from it. All the human elements of trust, deception, unintended consequences, and uncertain relationships are wired into the connections of the network.