Is This Beer Review Real or AI?
You’re no doubt familiar with Alan Turing’s test for synthetic intelligence. By his standard, if language created by software could fool a human, it could be described as “intelligent.” For years I’ve been hearing about AI used for highly formulaic articles, like sports summaries, and I’ve dismissed it. Sure, taking scores from a game and creating little synopses is easy, because the parameters are so fixed. But real writing—you know, the kind I do—is far too precious and human to duplicate, or so I assumed. Encountering a demonstration of one version of AI writing, composed by GPT-3, however, I was left a little shaken. It’s not quite as good as Rushdie, but it’s far from text Siri, either. Care for a demonstration? Select the human and AI-composed review of Sam Adams Boston Lager:
This is a beer that I could drink over an hour or 45 minutes. It has a nice color to it, and a crisp flavor that hits your tongue right away. Had a dry feel to it with some bitterness at the sides of your mouth. You can definitely taste the hops, leaving it more honestly bitter than I expected and some spice notes with that.
Boston Lager is a crisp and refreshing amber lager that is easy-drinking. It is perfect for a hot day, and it quenches your thirst. This beer is perfect for a beginner beer drinker, and it is a great gateway beer to other styles of beer.
Well?
One of those reviews was written by software called GPT-3, which harnesses the brute force of modern storage and processing power. In order to make the program run, you do have to enter some information, including the kind of writing it’s performing, along with a few bits of data to work with. This makes sense—dialogue is quite different than technical writing or, say, blog posts. The company behind the tech offers dozens of different formats, including, amazingly, a restaurant review function. It’s so sophisticated, though, that when I replaced the word “restaurant” to “beer” in the parameter, it shifted gears effortlessly. Below is a direct transcript.
Name: Sierra Nevada Hazy Little Thing
[Word Inputs]: Hazy, fruity, refreshing
Review: "If you're looking for a refreshing, fruity beer, look no further than Sierra Nevada's Hazy Little Thing. This hazy IPA is the perfect summer beer, with its citrusy and hoppy flavor. It's the perfect beer to drink on a hot day."
From just three words, it managed to describe the beer (accurately) as a hazy IPA, one with citrusy flavor. Impressive! In order to get the Sam Adams review at the top of the post, I fed the computer these prompts: “crisp, quenching, amber, easy-drinking.” Now can you guess which one it generated?* (I figured the use of “gateway” might throw you. I was certainly surprised to see beer-appropriate slang included.) Here are a few more:
Jever Pilsner. If you're looking for a sharp, bitter, and crisp pilsner, Jever is a great option. It's got a strong flavor that can hold its own against heavier dishes. However, it's also light and refreshing enough to enjoy on its own. (prompts: sharp, bitter, crisp).
Guinness Stout. If you're a fan of dark, roasty beers, you'll love Guinness Stout. This beer is smooth, satisfying, and sessionable, meaning you can drink it all night long. It's the perfect beer for those cold winter nights. (Roasty, smooth, satisfying, sessionable)
Augustiner Helles. I'm a fan of Augustiner beers, and this Helles did not disappoint. It's a bit spicy and rustic, with a lot of malt flavor. It's not too sweet, and has a nice, crisp finish. I would definitely recommend this beer to anyone looking for a good, all-around lager. (Spicy, rustic, malt-forward)
Saison Dupont. “This is a great example of the style. It's phenolic, effervescent, and dry, with a nice peppery flavor. It's the perfect beer for a warm day, and would be great paired with some light fare like a salad or a sandwich.” (Phenolic, effervescent, dry)
What impresses me the most is how it has been programed to fill in details. For Guinness, I gave it the word “sessionable,” in-group slang among drinkers, and it actually defined the term accurately. It knew helles was a lager. It knew phenolics were peppery. Are these the most trenchant reviews? The most elegant? The most specific and detailed? No—but they certainly pass the accuracy and humanity test. They seem as intelligent as the average beer rater (the human review of San Adams came from Beer Advocate).
The program even provides multiple responses. Below are three versions of Heady Topper, generated using these prompts: “Hazy, tropical, aromatic, strong.” Note that it added the location of the brewery and the strength of the beer, plus the adjectives “peach, mango, and pineapple.” It also inferred we were talking about hops, a word I didn’t mention, and knew it was a double IPA. I’ll bold the original info GPT-3 came up with when crafting these.
Brewed in Vermont, Alchemist’s Heady Topper is an 8% ABV double IPA that is hazy, tropical, aromatic, and strong. This beer is definitely hopped up, with a big tropical fruit flavor and aroma. It's a bit too strong for my taste, but fans of IPAs will likely love it.
Heady Topper is an absolutely amazing beer! It's hazy, tropical, and aromatic, with a strong flavor that really packs a punch. It's one of the most delicious beers I've ever had, and I would highly recommend giving it a try if you get the chance.
If you're a fan of hazy, tropical IPAs, you'll love Alchemist’s Heady Topper. This beer is aromatic and strong, with flavors of peach, mango, and pineapple. It's definitely a delicious brew, but be warned - it's also 8% ABV!
Any of the above would work well enough, but you could mix and match and improve the computer’s work with another minute or two of effort. For example: “If you're a fan of hazy, tropical IPAs, you'll love Alchemist’s Heady Topper. Brewed in Vermont, it’s an 8% ABV double IPA that really packs a punch. It's one of the most delicious beers I've ever had, with flavors of peach, mango, and pineapple.”
If you know beer and have just sampled Heady Topper, you can probably write a better review. What’s mind-boggling to me is that someone could offer this review without having gotten anywhere near Vermont. If I were planning to get into high school or college education, this would make me really nervous.
And if anything, these examples tend to lowball GPT-3’s capacity for writing human-like prose. I heard about the software from Vauhini Vara in this piece. In it, she describes how the software helped her write about her sister’s death. As she reveals, the prose is sometimes janky—at one point it gets stuck in a very 1950s loop of endlessly repeating the same sentences. But sometimes it’s downright exquisite. In one passage, Vauhini starts with these words: “I remember the same small set of details” before passing it of to GPT-3:
“Her dark eyes, her long thin fingers. But these are just the details that are universal to all sisters. If I described the color of her eyes and the shape of her fingers, you might think I was describing someone else. Even if you knew my sister, you might not recognize her in these details. (Why is it that, though we look at our siblings so closely, we forget them so quickly?) So I can’t describe her to you. But I can describe what it felt like to have her die.”
That’s fine, emotionally-astute writing. I’d have gotten that little endorphin rush of satisfaction had I composed it. Because the software is still quirky and mechanical, it produces too much poor writing to take my job. Yet. I’m no longer so confident it won’t take it eventually.
*GPT-3 generated the second review.