Writing may represent one of the few knowledge work domains resistant to AI displacement, according to a blog post by software engineer Murat Buffalo. While large language models have rapidly advanced at code generation—leaving programmers struggling to meet inflated productivity quotas—AI writing remains trapped in what Buffalo calls the “uncanny valley,” producing what he characterizes as “robotic” prose with “tired vocabulary” and lack of genuine insight.
Buffalo grounds his argument in three observations. First, he notes that AI labs have already attempted to improve prose generation but hit a plateau, unlike image, voice, and video models that improved rapidly through increased parameters and compute. This suggests fundamental differences in how language models approach text versus other modalities.
Second, Buffalo frames writing as a “wicked problem”—a systems theory term describing problems lacking definitive formulation, clear stopping rules, or objectively correct solutions. Math sits at the opposite extreme, with binary verification. Code relies on formal logic and automated testing. But prose success is “fundamentally subjective,” depending on resonance with a reader’s mind. This, Buffalo argues, explains why AI excels at math and code but produces “utter slop” at writing.
Third, Buffalo suggests writing may be an “AI-complete problem” requiring Theory of Mind—the ability to simulate a specific reader’s mental state in real time. Good writers track what readers know, manage cognitive load sentence by sentence, and predict how arguments land. LLMs, by contrast, optimize for statistical probability of the next word across vast datasets without empathizing with or understanding a particular human audience.
Buffalo concludes by invoking comparative advantage from classical economics. Even if AI has absolute advantage in typing speed and volume at zero cost, human labor maintains value where opportunity cost is lowest. As AI-generated content saturates the web, authentic human voices become costly signals—economically valuable precisely because they’re expensive to produce and impossible to fake, drawing a parallel to peacock tails in nature.
Unlike programmers, Buffalo argues, writers need not change their workflows to maintain comparative advantage. While software engineering reveals its inherent complexity under AI pressure, writers have always wrestled with communication’s inherent complexity at “the wicked frontier” and remain largely untouched.
Key facts
- LLMs have plateaued on prose quality despite AI labs’ efforts to improve writing capabilities
- Writing is a ‘wicked problem’ lacking objective verification, unlike code which has automated testing
- AI writing lacks Theory of Mind—the ability to simulate a specific reader’s mental state
- Authentic human writing becomes increasingly valuable as AI-generated content saturates the web
- Writers maintain comparative advantage without needing to change their workflows