On The Net
Welcome to Just Okay
by James Patrick Kelly
ai blues
Several years ago, while casting around for a column idea on a cold Sunday afternoon in December, I decided to interview Chatbot 3.0 partly as an experiment, mostly as a joke. My expectations were low because AI had been perpetually “ten years away” since the turn of the century. However, I was immediately taken aback by the bot’s flawless grammar and amused by its clever answers. In less than an hour it had generated a just okay column in response to my questions. I cut a couple of sentences and “An Interview With ChatGPT (12/18/2022)” ran otherwise unedited in the September/October 2023 issue.
Little did I know that three years later, pulling a trick like that would get me banned from this magazine! As each new iteration of LLM technology has gotten more sophisticated, its drawbacks have tarnished the worthwhile things it does that are just okay. We worry about the hallucinations and the biases in training data and the disturbing energy footprint. Artists, writers, and other creatives are suing AI companies for copyright infringement over the theft of their works to train generative AI. Earlier this year there was an uproar in the science fiction and other writing communities when The Atlantic published a tool to search the infamous Library Genesis database that contains 7.5 million books and 81 million research papers. Tech giant Meta used LibGen, which is full of pirated material, to train its AI systems. According to court documents, rather than pay for licenses, Our Friend Zuck decided it was faster and cheaper to go with the pirates. When I searched with the tool, it spit out ninety-five citations of my work, mostly stories from these pages. And if that isn’t scary enough, AI pioneers like Nobel Prize winner Geoffrey Hinton are worried that AI may pose an existential threat to humanity sooner than we think.
slop
Not an existential threat, but nonetheless annoying, is AI slop, a kind of digital pollution churned out with effortless speed by LLM-powered applications. It most often takes the form of just okay images and videos that can take up as much as a third of your social media feeds. These might be surreal, silly, mawkish, humorous, or provocative but they work to grab your attention and possibly sell you something.
AI slop also comes in narrative form. In 2023 Amazon was overwhelmed by sham books written by AI. Some were sheer nonsense “prose” while others were scam rewrites of legitimate work. To protect its store, Amazon issued a disclosure policy for AI generated work. “We define AI-generated content as text, images, or translations created by an AI-based tool. If you used an AI-based tool to create the actual content, it is considered ‘AI-generated,’ even if you applied substantial edits afterwards.” However, Amazon decreed that AI assisted work that has been brainstormed, edited, refined, error-checked, or otherwise improved by bots need not be disclosed. Sheila has also struggled with AI slop stories. Hundreds of unintelligible, unreadable, and horrifyingly clichéd stories generated by AI continue to clog her screen. This is why you’ll find a harsher warning on the ‘Mov’s submissions portal. “We will not consider any submissions written, developed, or assisted by these tools. Attempting to submit these works may result in being banned from submitting works in the future.”
Meanwhile, there is a new kind of slop, thanks in part to NotebookLM a free document analysis tool created by Google. NotebookLM digests documents you load into it and tries to summarize, organize, and explain them. It is the quintessential just okay AI app, but I find it useful. Crucially, one of its outputs is a podcast style overview that features a male and a female AI voice discussing the material you’ve selected. The two speakers chat believably; they interrupt, finish each other’s sentences, repeat themselves, and evoke surprise with the occasional Oh wow or Really? Click over to the Deep Dive into Brain and Behavior for an typical example. Notebook LM has spawned clones designed primarily to create podcasts that are totally AI—text and voice. Podcast slop has already spread to Spotify, Apple Podcasts, and You Tube.
talkies
I was more surprised by the verisimilitude of the AI narrators in NotebookLM than I was by its analytical abilities. The technology has levelled up from the robotic and clumsy voice assistance of Siri and Alexa, and has opened new use cases. For instance, among trusted news sources, the BBC, the New York Times, Reuters, the Economist use just okay AI voices to read stories from their print editions. Then there’s voice cloning. In as little as five minutes you can train a LLM to imitate your speech; more robust training produces better impersonation. Celebrities (or their estates) are licensing their voices for specific uses like James Earl Jones and Judy Dench. Unfortunately, the cloning tech is so simple to use that unauthorized deepfakes of all kinds have proliferated.
I’ve done a little narration as a side hustle and what concerns me about this tech is its impact on the audiobook industry. All of the largest audiobook retailers— Audible, Apple Books, Google Play Books, Kobo, and Barnes & Noble—allow AI voice narration in some of their products. Audible, with a two-thirds share of the audiobook market, features over forty thousand AI narrated titles. Spotify also is aggressively expanding into AI narration.
Why Is AI Narration for Audiobooks a Bad Thing? makes a persuasive case against AI narration since it results in audiobooks that are, at best, just okay. Meanwhile there are simple and fair ways for professional human narrators and writers—especially indy writers—to collaborate. However, many writers who are not traditionally published chose just okay AI narrators to reach the audiobook audience, currently estimated to account for 25 percent of publishing sales.
golden
To get a sense of how professional narrators are reacting to this AI incursion into their business, I turned to my friend Stefan Rudniki, a Grammy and Audie award winning voice artist. Stefan, sometimes referred to as “the Golden Voice” and Gabrielle de Cuir are the principal vocal talents of Skyboat Media with a decades long connection to science fiction.
I started our interview by describing a narrator industry that was roughly separated into three tiers. One might include all audiobooks with just okay AI narration. A second and much larger tier might include those produced by independent voice artists, many solo, who work either with publishers or individual writers and who do narration and postproduction. A third might include more established vocal artists who work with a production team, mostly for audiobook publishers.
Stefan considered this description. “Simplistic, but yeah, it’s certainly a viable way of looking at the industry in this present moment.” I asked if he had ever collaborated with individual writers for projects destined for Audible’s ACX store, where writers and narrators negotiate to share revenues on their audiobooks. Some compare this to Amazon’s indy-favorite Kindle Direct Publishing, except for audio. He grimaced. “Not with the revenue share system, because I have expenses upfront. Most of the recording done in the business is done by solo narrators, essentially doing their own editing. For myself, I find that is debilitating to any kind of really effective performance technique. I work almost exclusively with a director. I need an audience. I need somebody out there. And I work with a post (production) team. The one we’ve been with for over a decade now is in the UK. They’ve got proofers because you’ve got to have that as well. Somebody has to listen to it. With Audible they hire somebody or an AI to listen and catch mistakes.”
We talked about voice cloning and I asked him if he’d ever consider it. “No, because my voice is the primary thing that I have as a resource that I can count on. In fact I had some health issues about fourteen months ago where I thought I was losing it. I mean I got hoarser. It got very difficult to maintain the degree of flexibility the voice has to have. And it was by dint of extreme discipline and various kinds of vocal exercise and the help of several doctors that I got back to what I had two years ago.
I pressed him on this. “What if you hadn’t got back to where you wanted to be? There’s a vast library of your audio work recorded that would be perfect for training a quality clone of your voice. That way you continue to earn a living. Would that interest you?”
“No,’ he said, speaking with obvious passion, “and this is a different reason, a different question now. Because it’s not going to be my voice, right? Maybe it’ll sound like it, but it will not have the emotional underpinning. You know, we regularly work with large casts. All of the people who we work with, the actors, they have a commitment not just perform, but to act. It’s not accidental that the keywords in actor training involve things like intention. Purpose. You take it seriously. Devote yourself fully to the moment. The whole emotional context, you don’t put it away, you bring it with you wherever you go. It demands that you use certain kinds of emphasis. These are not intellectual choices and AI is not going to do that.”
exit
I asked Stefan whether he thought there would be as many human narrators in twenty years as there are now. “It’s a difficult question.” He paused. “Because in a way there are too many now.” He confided that he’d been hired to redo at least a half a dozen major books over the years because the original performance had been rejected by the publisher. We talked about how often Audible listeners give bad reviews to narrators. It turns out that nearly 60 percent of listeners have ditched an audiobook because they didn’t enjoy the narrator.
Yikes! As Steven Wright famously observed, “Half the people you know are below average.” The LLMs, despite all their drawbacks, have set a minimum standard for us. The challenge for writers and narrators and artists of all kinds is to be better than just okay in this AI Age.