There’s this meme that’s been going around the internet for a little while now. The format of the meme is “I forced a bot to watch X amount of Y, and then asked it to write Y – here’s what it produced”. There are countless examples of this, but I want to write about one that I came across this summer.
This past summer, on June 13, 2018, a tweet following the format of the meme started making its way around the internet. Twitter user Keaton Patti @KeatonPatti posted this tweet, which started with “I forced a bot to watch over 1,000 hours of Olive Garden commercials and then asked it to write an Olive Garden commercial of its own. Here is the first page.” This “first page” displays a TV script format, listing characters such as WAITRESS and FRIEND 1, with dialogue such as the WAITRESS saying “Pasta nachos for you”, stage directions like “Friend 4 smiles wide. Her mouth is full of secret soup”, and, at the end of the “commercial”, the announcer saying (in a “wet voice”) “Olive Garden. When you’re here, you’re here.”
It’s extremely funny. And, of course, it’s fake. There’s no way a bot wrote that.
I’m the first to admit I’m cranky when I see things like this. When I see even the beginning of memes like this, and read the words “I forced a bot to–”, I immediately say to myself, often out loud, “No you didn’t.” Because the language describing the task that the bot is doing –”forcing” the bot to “watch” or “read” something –is describing a process that isn’t possible for current artificial intelligence.
In her 2018 book Artificial Unintelligence, Meredith Broussard describes a helpful distinction between what she calls general AI and narrow AI. General AI, she writes, is like the AI we see in the movies; it’s sentient robots with consciousness, or machines that can “think”. Narrow AI, on the other hand, is a mathematical model for prediction. It can give you a prediction about how likely an answer to a question is, so long as that question can be represented by a number. Narrow AI is, as I like to say to my students, just really fast statistics.
And I, knowing that not everyone knows the distinction between what kind of AI is currently possible (narrow AI), and what kind of AI is currently idealized (general AI), become frustrated and grumpy that people without this knowledge are being “duped” by this meme instead of recognizing the human behind the humor, and the realities behind AI. Is it that something is perceived as funnier if a bot wrote it instead of a human? And if so, why is that? Neither of which I have answers for.
So I came across this particular meme and I became cranky and grumpy and frustrated, with a desire to correct the inaccuracies, but unsure of what to say that wouldn’t make me sound like the pedant that I turn into when encountering this meme.
Fortunately, though, the world is full of wonderful and intelligent people like Janelle Shane (@JanelleCShane on Twitter), who, in a series of excellent series of tweets, concisely described why it is extremely likely that the ‘I forced a bot to watch X’ memes are, in fact, written by humans. Her tone is informative and accurate, and I recommend reading this whole thread. She has the right idea, of course; instead of getting annoyed, educate. So I’ll take a moment here to summarize a few of her comments.
These memes are almost always written by humans. There a few ways that we can tell.
— Neural networks (programming paradigms that enable computers to learn from observational data) give out the data that they take in. They learn by what samples they’re taking in. A neural network that’s given a video input will produce a video output, not a script (and certainly not a perfectly formatted one).
— Neural networks need data to learn from. It’s unlikely that there are, as Shane calculates, 120,000 unique 30-second Olive Garden commercials in existence, which would be a reasonable estimate of how much data a neural network would need to learn from. (And if there are, then WOW have I underestimated Olive Garden’s advertising budget.)
— Neural networks aren’t good at outputting complex sentence structure. While they might be able to pick up and “remember” the vocabulary in the data, neural networks tend to do poorly with complex grammar and syntax. The result is sentences that read like word salad. (See Shane’s Harry Potter bot examples.)
— Words that don’t appear in the data you’re training your network on will not appear in any of your output. Therefore the word “nacho” would be extremely unlikely to show up in a script, unless Olive Garden has made a massive change to their image.
Essentially, as Shane writes, actual AI-written text isn’t just isn’t this coherent. Not that it won’t be some day. But we aren’t there yet.
Does this mean that you can’t enjoy memes like this? Of course not, and I hope you still do. Things that you find funny are still funny, regardless of how they were produced. A bot failing to reproduce human language and culture is funny. A clever human taking these known failings and turning them into something relatable is also funny. It doesn’t have to be one or the other.
As for me, Our Grumpy Pedant? I’m in solidarity with Shane in wishing that people wouldn’t present these memes as bot-written. But I don’t disagree that they’re funny, and I give a lot of credit to Keaton Patti (who is a professional writer and comedian) for writing this fake bot-written Olive Garden ad. It really is funny. I particularly like “When You’re Here, You’re Here”, which is how I’ve felt in every Olive Garden I’ve ever been in. And this meme has also given way to this blog post, which hopefully has served as an entry point for readers to learn a little bit about how AI does and doesn’t work.
And who knows? Maybe I can learn to appreciate this meme after all. At the very least, Twitter user Michael Lynch (@deliberatecoder) had a follow-up tweet that I thought was pretty great.