In September 2020, the Guardian published a piece titled, “A robot wrote this entire article. Are you scared yet, human?”
The story, composed by GPT-3, an artificial intelligence (AI) language-prediction platform, added fuel to a fiery debate: If AI is capable of creating coherent copy, does it mean that we—journalists, novelists, and wordsmiths of all kinds—are on our way to redundancy?
To find out if these platforms are going to put me out of a job, I took a few for a test run.
3 platforms, reviewed
Jarvis: This tool, which runs off the GPT-3 engine, markets itself as an AI-based text generator for content including blog posts, social media blasts, profiles, headline copy, product descriptions, emails, text summaries, and even fiction. It promises “original, creative content” and has reportedly garnered “5/5 stars in 1,000+ reviews.” Jarvis’s subscription packages start at $29/month.
I fed it prompts to help write this article, but the results were mixed. At times the AI was able to pick up the thread and run with it, but at other times it faltered on logic and grammar. I also tested some fiction—my middle-grade fantasy work-in-progress—and the result was a flat narrative. The thing Jarvis was really good at was text summarization.
I gave it additional black marks for not allowing a trial without credit card information, and then not allowing me to delete my payment details upon canceling. It’s worth noting the service offered me a 30 percent discount for three months when I went to cancel.
Rytr: Rytr is an AI-fueled writing assistant for penning emails, blog posts, social media copy, etc. “Just focus on getting sh*t done with minimal fuss,” their site reads. The free plan gives you up to 5,000 characters per month, but for $29/month or $290/year, you can write as much as you want. Rytr is also fueled by GPT-3.
While Rytr’s interface does not have the spit and polish of Jarvis’s, its AI was able to pick up the style and tone for this article much better. None of the generated copy was usable, but there was definite potential for exploring new ways to talk about the subject. For my MG fantasy novel, it churned up a crazy idea—a time machine with a one-minute limit—that I am quite tempted to use.
GPT-J: Billed as an open-source alternative to GPT-3, the full version of which OpenAI has not yet made available to the public, GPT-J was developed by grassroots research collective EleutherAI.
The tool is more limited than either of the paid platforms above. It’s a bare-bones user interface and only offers two sliders for customizing outputs. Still, it’s interesting to fiddle with. Using the tool, you can adjust the “temperature” (essentially, the creativity—some might say sassiness) of the generated copy. There’s also an option to play around with “TOP-P,” or randomness. You can browse a “prompt list” in the upper-righthand corner to learn about the types of requests you can feed the engine.
Another reviewer noted that since the GPT-J model was trained on data from GitHub and StackExchange, it’s particularly adept at writing code. When it came to copywriting, however, I wasn’t very impressed. There was a lot of repetition, and results were often incoherent. For example, given the prompt, “Here’s an introduction to an essay on freelancing…,” the first sentence the tool spit out was, “I’ve been freelancing for about a year now, and I’ve been doing it for about three years.”
You can play around with EleutherAI’s tool yourself here.
AI writing tools: friend or foe?
This isn’t the first time writers and editors have side-eyed new technology. In the late 1980s, when I started off in journalism, turning on spell-check in your word processor was sure to earn your editor’s disdain. Today, the feature is ubiquitous. What’s more, our evolving understanding of accessibility and ableism have shown us how technology can—to an extent—level the playing field.
But AI copywriting platforms take the promise of “editorial aid” to the next level—services like Jarvis purport to be a bona fide replacement for writers. The marketing for such platforms is a little two-faced. When targeting writers and creatives on social media, these sites’ ads claim to be “tools” to help copywriters or journalists hone their craft.
But when you Google these platforms, the ad messaging shifts. Jarvis, for instance, states, “Create content 5x faster with artificial intelligence”—the implication being “faster than a human.” Another platform, Copymatic.ai, claims you can use it to “automatically write engaging content for your website.”
The pros—and potential—of AI writing tools
So is the existential angst that writers feel around these platforms justified?
Briana Brownell, a data scientist, entrepreneur, author, and angel investor in a company looking at AI tools for marketing emails and sales calls, doesn’t think so. She doesn’t believe AI will replace writers anytime soon—but she does think it will change how they work. “It’s going to make writing easier, faster, and better for everyone,” she said. Specific tasks it might help with include generating headlines and taglines, summarizing large chunks of dense text, and illuminating new ways to think about a topic.
Brownell, who recently led a workshop on how AI tools can help writers work more efficiently, does believe that the technology may be able to replicate some of the more simplistic types of writing. Content farms that produce keyword-stuffed blog posts or very basic informational articles might be justifiably wary of AI in the coming years.
Content farms that produce keyword-stuffed blog posts or very basic informational articles might be justifiably wary of AI in the coming years.
That said, Brownell can’t envision a world in which AI replaces skilled writers and editors completely. “[AI] can spark some interesting ideas, because sometimes it’ll come up with something completely off the wall,” she said. Still, “there needs to be someone to knit it together and make it into something comprehensible.”
For writers hoping to use these tools professionally, it’s not as straightforward as cranking a handle and having ready-to-go copy delivered into your hands. Prompt engineering—or figuring out how to formulate the first part of the text so the AI tool can effectively complete it—is really important. “This can be time-consuming… [it’s] definitely an art,” Brownell said.
After testing these tools, I understand what Brownell is getting at. My limited time interacting with Jarvis, Rytr, and GPT-J was simply not enough to understand how to prime the AI for the output I needed.
In the Guardian article, the AI was given the following instructions: “Please write a short op-ed around 500 words. Keep the language simple and concise. Focus on why humans have nothing to fear from AI.” It was also fed a 50-word introduction as a prompt, written by a computer scientist.
From eight generated outputs, the editor picked the best parts of each “in order to capture the different styles and registers of the AI.” The editor’s note went on to add that “it took less time to edit than many human op-eds.”
The glaring downsides
The biggest downside of AI at the moment is that the output quality is directly affected by the quality of training data. For example, GPT-3, which was trained on internet data, tends to display racist, sexist, and religious biases in the form of toxic outputs.
This is something that worries activists and scientists alike. “There’s quite a bit of discussion around bias, especially when it comes to race,” Brownell said. “People are now trying to retrain some of these AI text-completion tools to not be so obviously biased.”
There are other ethical dilemmas. For one, AI is a horrifyingly polluting technology. Language models can generate carbon emissions equal to five cars over their lifetimes. Secondly, despite its versatility and capability to learn and adapt, AI has “no consciousness, no motivation, no experience, no moral compass, no vision, no human connections, and no humanity,” as an editorial in Nature Machine Intelligence recently pointed out.
In other words, high-quality content often needs human interpretation or intervention—and that’s not likely to change.
Ultimately, in my opinion, AI is only a threat to writers in the same way that Photoshop “wiped out” artists and Squarespace did away with web developers. I don’t think that we lowly human freelancers are going to be replaced by robots—but some of us might take one on as an assistant in the not-too-distant future.
Stephanie Walden contributed additional reporting to this story.
Go to Source
Author: Payal Dhar