Three new cards from one source. Dan Shipper's interview with Platformer (August 20) covers the AI copy-editing agent Every built for its editor in chief. The batch is small but atomically clean: a specific hill-climbing method for encoding expert taste, an observation about the hidden complexity in rule-based tasks, and a candid point about AI adoption in professional writing.
Encoding expert taste from demonstrated decisions
Dan Shipper describes building a copy-editing agent for Every's editor in chief, Kate Lee, by collecting 30,000 of her historical edits, building a prompt from that dataset, and then back-testing the prompt against every previous document the team could reach. The goal was not just to replicate her rules but to replicate her judgment.
"I collected a dataset of 30,000 of her historical edits, used that to make a prompt, and then back-tested the prompt on all of the previous documents to hill-climb and make it better and better."
The back-testing is the mechanism. Each iteration of the prompt is scored against what Kate would actually have done on held-out documents. The loop tightens further after deployment: every correction she makes feeds back as new training signal, updating the agent's sense of what it missed. Building a copy-editing agent from 30,000 historical edits and hill-climbing the prompt against all past decisions encodes durable expert taste.
The same project surfaced something Shipper did not initially expect: copy editing, which appears rule-based, resists full automation even when the rules are known and the model is capable. Exception density is the reason. Every application of a rule requires a judgment call about context, author voice, and house style. A model that applies rules reliably still fails at the exception layer. The hill-climbing method sidesteps this problem by encoding past decisions rather than rules, but it does not eliminate it.
"Even copy editing, which is rules-based, is super, super complicated and not fully automatable even now."
- Dan Shipper, Building a copy-editing agent from 30,000 historical edits and hill-climbing the prompt against all past decisions encodes durable expert taste. Hill-climbing a prompt against 30,000 historical edits encodes expert taste with continuous self-improvement as new corrections accumulate.
- Dan Shipper, Even rule-based tasks like copy editing carry enough context-sensitivity to resist full automation by current AI. Even explicitly rule-based tasks carry enough exception-density and context-sensitivity to resist full automation by current AI.
The silence that tracks AI use in professional writing
Alongside the technical observations, Shipper makes a sociological one. He argues that AI use among professional writers is already near-universal in practice, and that the gap between private use and public acknowledgment is maintained by stigma, not by non-adoption.
"I think there is a real dirty secret right now, which is that almost every writer is using it. Just, most of them are not saying so."
The mechanism is self-reinforcing: rare disclosure reads as rare use, which preserves the social cost of disclosing, which keeps disclosure suppressed. The norm stabilizes at widespread use with widespread silence. Near-universal AI use among professional writers is not surfacing publicly because social stigma around authenticity suppresses disclosure.
- Dan Shipper, Near-universal AI use among professional writers is not surfacing publicly because social stigma around authenticity suppresses disclosure. AI use among professional writers is near-universal in Shipper's observation, with disclosure suppressed by authenticity stigma rather than by actual non-adoption.