Every's CEO on cloning editor taste and AI's productivity paradox

Dan Shipper runs a company that reviews AI models while building products on top of them—a dual mandate that puts him at the center of automation's contradictions. His firm, Every, operates as both a technology publication and a product studio, employing roughly 30 people who produce journalism, software tools, and early assessments of frontier models from labs like OpenAI and Anthropic.

In a recent interview with Casey Newton of Platformer, Shipper described how AI now writes essentially all of Every's code while humans still handle most essay writing. But the boundary is shifting. The company has begun automating editorial judgment itself by training an agent on 30,000 historical edits from editor-in-chief Kate Lee, then back-testing the system against her past work to refine its accuracy.

Why it matters

Every's experiment captures a broader tension facing knowledge work: AI can replicate patterns from solved problems, but novel challenges still require human experts. That dynamic explains why companies automate aggressively yet keep hiring—a pattern with implications for workforce planning across industries as models improve.

Automating taste, not just tasks

Shipper told Newton that Every's editorial agent can now review Google Docs and make suggested changes that match Lee's style and standards. The system isn't perfect, but it distributes her expertise across landing pages, launch emails, and other content that previously required her direct attention. Shipper frames this less as replacement and more as leverage: "Kate has a specific set of skills as an expert inside of Every that she can only apply right now by spending her time," he said. The tool lets her "spread it throughout more of the org, where she doesn't have to spend her time to do more work."

The copy-editing agent only recently became viable. Shipper said he'd been attempting to automate Lee's role "since GPT-3" without success until models gained sufficient instruction-following ability and browser control to operate inside collaborative documents.

The paradox: more automation, more humans

Every doubled from approximately 15 employees to 30 over the past year while loudly pursuing automation. Shipper attributes the growth to what he calls AI's structural limitation: models are "trained on the residue of human expertise" and can only solve problems similar to those already documented. When users apply AI to novel situations, the output is often close but inadequate—what Shipper calls "slop." Experts must then step in to bridge the gap between generic solutions and specific needs.

This creates two roles for human specialists. First, they build infrastructure that channels non-expert AI use into approved patterns—ensuring that when colleagues query data or generate code, the results align with organizational standards. Second, they pursue ambitious projects that weren't feasible before AI raised both the floor and ceiling of what's possible.

Navigating conflicts of interest

Every's position as both reviewer and customer of frontier labs introduces tension. The company published a critical assessment of Anthropic's Claude Sonnet 5 under the headline "a model pitched for everyone impresses no one." Shipper told Newton that labs typically ask for feedback before launch because "they'd rather know beforehand, honestly, than find out from a ton of other people who use it." He argued that Every's credibility as an arbiter—"no one trusts a model company to tell you where they objectively sit"—represents durable value even as labs encroach on application territory.

On the risk of building atop platforms that might absorb his products as features, Shipper acknowledged uncertainty but likened model makers to oven manufacturers: making the hardware doesn't guarantee knowing how to cook. He said Every operates in a zone where "you have to both really want to make something awesome and high quality, and be willing to throw it out every three to six months as the capabilities change."

These details were first reported by Casey Newton in Platformer's interview series on productivity in the AI era.