AI’s writing and storytelling abilities are advancing far more slowly than many expected. “The technology isn’t progressing in quite the way it was sold,” said Hollywood Star Ben Affleck.
“If you try to get ChatGPT or Gemini or Claude to write you something, it’s really shitty,” Affleck says.
When ChatGPT launched three years ago on OpenAI’s GPT-3.5, the reaction was borderline hysterical. And for good reason: it was a genuine leap over earlier large language models like GPT-3, which powered the first wave of mainstream AI copywriting tools such as Jasper.
GPT-3 was like Steve—the C-minus freelance copywriter you’d quietly stop hiring after two or three assignments. ChatGPT, by contrast, was Ted: a B-minus writer with lots of ideas and the occasional spark of insight.
Some of us have worked with a Ted before and thought, “Yeah, he’s… fine.” Once AI cleared the Ted Chasm, the hype exploded.
That frenzy was amplified by bold promises from AI leaders who argued that progress would be exponential, driven by scaling laws. Sure, AI might be Ted today, but give it a few months and it would be Toni Morrison.
That didn’t happen.
In reality, progress in AI as a writer and storyteller has largely stalled. Part of that comes down to the limits of the training process itself. But as I explored in my deep-dive essay Storytelling Is the New Coding, there’s another force quietly shaping the outcome.
As Dan Balsam, co-founder and CTO of Goodfire, explained, AI developers are being forced into hard trade-offs. Constraints around power consumption, chip availability, and sheer cost mean models can’t be optimised for everything at once.
So the biggest players are following the money. Rather than pushing hard on prose and narrative, they’re prioritising areas like coding and technical analysis, where the economic upside is far greater. Replacing software engineers and data scientists is simply more lucrative than replacing copywriters.
It also helps that code is easier to evaluate. AI can test and retest its output until something works. Writing doesn’t offer that kind of objective feedback loop, and the quality of training data for code is generally higher.
All of which makes the outlook fairly clear: AI’s writing capabilities are likely to plateau, improving only in small, incremental steps.
Storytelling is a harder problem than it looks, and since AI already produces passable, jargon-heavy business writing, there’s little commercial pressure to push it much further.
OpenAI now carries a valuation north of $500 billion on roughly $13 billion in annual revenue, while planning to burn through an estimated $17 billion this year alone.
At those numbers, the only narrative that supports the scale of investment pouring into AI is one of sweeping disruption—AI replacing human workers at massive scale.
A far less dramatic story, in which AI behaves like most “normal” technologies—slow, uneven adoption, messy organisation-wide transformations, and incremental productivity gains—doesn’t justify hundreds of billions in capital. And it’s certainly not a story you can afford to tell if you’re Sam Altman or Dario Amodei.
