Software engineering is experiencing a crisis of faith, but the culprit isn’t a lack of human intelligence. Robert Peake, a developer with 25 years of experience, argues that AI automation follows 'Verifier’s Law': the easier a task is to check, the easier it is to train AI to perform it. This principle explains why code, which compiles or fails and passes or breaks tests, was automated first, while creative work remains resistant.
The Grading Problem
Peake distinguishes between old software development, defined by specification, and new AI-driven development, defined by verification. He references AI researcher Jason Wei and Andrej Karpathy to explain that 'grading' is the make-or-break factor for automation. A voice clone passes the 'wife test' because audio fidelity is numerically checkable. In contrast, poetry evades mediocrity, a quality that numerical comparison cannot quantify. This creates a 'jagged frontier' where AI excels at tasks we find hard but struggle with tasks we find easy, if those tasks lack objective metrics.
Value vs. Verifiability
The article plots work on two axes: the ability to be automatically graded and the existence of objective value. Software sits in the top-right box: it is highly valuable and easily graded. This makes it prime for automation. Conversely, parenting sits in the top-left: high value, but impossible to grade automatically. Peake notes that commercial art and copywriting are currently being automated because they have clear client briefs (value) and deliverables that can be checked against those briefs (grading), unlike fine art, which relies on subjective taste.
Market Forces and Accountability
Automation isn’t just about technical feasibility; it’s about market incentives. Peake points out that AI hands the means of production back to subscribers, threatening the freelance model where a creative with a laptop owned their own workflow. He cites a Society of Authors survey from January 2024 where 26% of illustrators and 36% of translators reported losing work to generative AI. Furthermore, data from Brynjolfsson, Chandar, and Chen shows employment for workers aged 22-25 in AI-exposed occupations is running 19% below peers, largely due to hiring freezes.
Key Takeaways
- AI automation targets tasks that are both valuable and easily verifiable by machine, such as code compilation and test passing.
- Creative work like fine art and poetry resists automation because 'quality' cannot be numerically graded, only subjectively evaluated.
- The shift from 'specifying' software to 'verifying' AI outputs changes the developer role from writer to reviewer and accountable owner.
- Market forces drive automation in commercial creative fields where briefs provide clear grading criteria, unlike fine art.
The Bottom Line
Stop worrying about whether AI can write a sonnet; worry about whether your job produces deliverables that can be automatically checked against a spec. If it can be graded, it can be automated.