Memes Totally Confuse AI—And That Just Might Save Humanity From the Singularity (2024)

Artificial intelligence is supposedly drawing closer to singularity—the point at which machine intelligence surpasses human intelligence. That’s sparked debates about how we can best control AI’s capabilities; while superintelligent systems could revolutionize life as we know it, experts are still concerned they could do us more harm than good. Identifying AI’s limitations is one of the best tools we have for seeing how big the gap truly is between machine and human intelligence.

On the path to receiving a Ph.D. in electrical and computer engineering at Delaware University in 2021, Ishaani Priyadarshini wrote her dissertation on the topic of AI’s inability to decipher memes, and how that could help us in the battle toward singularity. “Internet memes serve as excellent checkpoints to ensure humans have the upper hand over machines, thereby preventing machines from understanding memes and surpassing human intelligence,” she tells Popular Mechanics.

Memes vs. Superintelligent Machines

Before pursuing artificial intelligence, Priyadarshini’s background was in cybersecurity, and she observed that over 50 percent of the internet is made up of bots—some of these bots are good, while others have malicious intent. This confirmed her worst fear: if AI were to achieve sentience and go rogue, it could spell disaster for your online data and privacy. “They would be able to gain every piece of information that is out there in the cyberspace … it could lead to catastrophic results in the future,” says Priyadarshini.

Her goal was to find a security measure that could defeat artificial intelligence before it ever reaches this point. The solution? Memes. These surprisingly complex amalgamations of text and images often pick apart the more humorous bits of pop culture and current events—and are usually posted across multiple social media platforms. (Picture Grumpy Cat and Doge.)

Priyadarshini wanted to find out why AI was so poor at decoding memes, along with how the humorous posts could be used as a cybersecurity tool in the future. AI’s inability to comprehend memes is multifaceted, but much of it has to do with the fact that they tend to be ambiguous. Priyadarshini says that our understanding of memes relies heavily on our own life experiences. “For a machine, memes are merely a bunch of text and images,” she explains.

We already have AIs that are able to recognize text and facial expressions, but there are really no rules with memes. Sure, Optical Character Recognition (an advanced system, able to recognize text in images) can already be used to accurately read license plates, but this type of machine learning already has a very good idea of where to look—decoding the text from memes is a completely different story.

“There’s no template that memes follow, which is why extracting the text from them is really challenging,” says Priyadarshini. You could argue that most memes are just variations on the same template, but the same meme could still be interpreted wildly differently depending on the content that’s been injected.

Memes Are Much More Complex Than We Think

We currently have no idea if AI has the ability to be sentient, meaning it would be able to think for itself and make its own decisions; even if we did know, it’s unclear how we’d measure just how sentient an AI is. That being said, we can measure AI’s ability to extract emotions from text and facial expressions—and it’s already pretty awful at picking up emotions from snippets of text. The billions of permutations you’ll find in both text and facial expressions make it difficult for AI to crack the code. That’s before we even bring in the context of the meme itself.

“Understanding memes could forever be a challenge for AI and remain an unsolvable problem,” Priyadarshini says in her dissertation.

Despite memes’ reputation as nothing more than mindless images on the internet, the process of unpacking them is actually much more complex than we thought. The human brain has an uncanny knack for parsing through an unimaginable amount of data very quickly—helping us understand memes in a matter of seconds. “Superintelligent machines” ... not so much.

✅ Understanding Context

Without life experiences, AI doesn’t have much context to lean on. Remember all the Suez Canal memes from back in 2021? Let’s use the meme below as an example:

View full post on X

An AI might be able to recognize that there’s a container ship in the video. However, can it properly identify it as the Ever Given that had run aground in the Suez Canal? To add another layer of complexity on top of that, why is Austin Powers in the background making a 27-point turn? Does it know that the Austin Powers movie series is supposed to be funny? “It may not be an easy task for AI to comprehend internet memes given the plethora and variety of memes found on the internet, along with the different patterns each of the memes can generate,” Priyadarshini says in her dissertation, which was published in the journal Computers, Materials, and Continua.

Meanwhile, the human brain is able to pick up the pieces with very little effort and register the clever parallel between Austin Powers’ multi-point turn and the container ship maneuvering out of the Suez canal.

An NPR exclusive with behavioral and data scientist Pragya Agarwal, Ph.D., reveals that the human brain can process roughly 11 million bits of information every second—that’s 0.011 gigabits per second. However, our conscious mind (filled with thoughts we’re aware of) can only grab onto about 40 to 50 bits per second. Thankfully, our brain’s amygdala is able to drink from this firehose of information by taking shortcuts to get the bigger picture; Agarwal mentions that the amygdala (the part of your brain responsible for your “gut” feeling) is able to match new information with pre-existing stereotypes and templates to make sense of it.

AI Databases Are Difficult and Labor Intensive to Train

Even if AI could successfully decode memes, that’s only one part of the puzzle. Internet memes are moving at, well, the speed of the internet—meaning AI has to keep up. Something that was funny last week could be considered old news in a matter of days. To offer some perspective, ChatGPT (an AI chatbot) has only been trained with a snapshot of the internet from 2021. Anything after that simply doesn’t exist to it.

Think of the memes that we saw in 2022, like those about the distribution of 1 billion free COVID-19 tests. Sure, the meme involves a photograph of the family from Full House, which AI would likely be able to recognize. However, without any context that the U.S. government was giving out free COVID-19 test kits, how would AI know what on Earth was going on in this image?

It would have to extrapolate the fact that every home in the United States is only eligible for just four free covid tests. We can clearly deduce that four test kits won’t nearly be enough for most families, but AI isn’t that sophisticated. Plus, artificial intelligence doesn’t really have a sense of humor.

Humor and AI Don’t Mix (Yet)

Being funny is an incredibly sophisticated skill that can mean different things to different people. This subjectivity makes it difficult for AI to understand humor—remember that only certain people might find one piece of content inherently funny. Priyadarshini mentions that AI would be able to parse through screen grabs from The Office and know that it’s a comedy, but it wouldn’t be able to unpack the humor within each scene; it has more of a binary is-this-funny-or-not type of understanding.

“The real challenge comes from identifying the different kinds of humor that are in those sentences or the videos,” says Priyadarshini. We’re just talking about general humor here, but there are also jokes that might involve dark humor, witty humor, or even dry humor. It really brings into sharp focus the psychology of what makes something funny. “We also don’t know what is the true sense of humor … it is different for everyone,” says Priyadarshini.

Without AI being sentient, it’s not clear if it will ever be able to have a complete understanding of humor.

A Futuristic CAPTCHA

Riffing on Full House is all well and good, but what will this next-level CAPTCHA—short for Completely Automated Public Turing to tell Computers and Humans Apart—look like in a real-world application? Priyadarshini simply thinks of it as a much more three-dimensional test to differentiate humans from machines; instead of reading weird text or picking which boxes have bridges in them, you’ll instead have to decode the meaning of a meme.

She gave a theoretical example of an authentication system for a nuclear power station. The employee would begin as usual by entering their username and password. Once that’s done, they’ll be prompted with a meme. In this instance, Priyadarshini used the case of the Chubby Bubbles meme (see below) where the user will be given four prompts to identify what is going on in the meme. There are still more questions than answers about how this type of test would work in practice, but it’s a novel idea that could revolutionize the future of cybersecurity.

We love the remote possibility of nuclear power station workers deciphering random memes before getting their day going. Sure, we’re most likely not giving up the nuclear football system anytime soon, but it certainly poses the question of improving security measures elsewhere. Above all, we’re happy to report the future of cybersecurity is looking bright—and full of hilarious memes.