Shopify CEO's AI Push Creates 'Slop Grenades': Seconds to Generate, Hours to Verify

The CEO Who Championed AI Gets Hit by 'Slop Grenades'

In April 2025, Shopify CEO Tobi Lütke issued a company-wide memo titled "Using AI Is Now a Baseline Expectation." He positioned AI as a thinking partner, research assistant, critic, tutor, and pair programmer, and made AI proficiency a required skill. Crucially, teams requesting new hires or resources had to first explain why the work couldn't be done with AI—moving AI from a toolbox item to a decision-table criterion.

A year later, on The Knowledge Project podcast, Lütke admitted some things had gotten worse. He described a new failure mode: where laziness once meant no output, AI now enables "output that's too much"—people generate emails or code with AI, skip review, and hand it off. Recipients must read, verify, and fix before proceeding. Shopify named this slop grenade: seconds to generate, hours to defuse.

Internal AI Assistant 'River' Amplifies the Pattern

Lütke revealed an internal AI assistant called River that participates in public work channels—summarizing discussions, creating tickets, prototyping, and even proposing code changes. He estimated roughly half of code change requests now originate from shared chat conversations (an oral estimate, not an official statistic). River also serves an organizational function: employees see how colleagues use AI in real time.

The problem: someone asks River to generate a change request, barely reviews it, and submits. Reviewers later discover flaws. The initiator's saved time becomes the reviewer's added burden. Lütke stressed accountability: machines cannot own outcomes; the human who hits send must own the quality.

Management vs. Frontline: Divergent Perceptions of Efficiency

The Guardian reported similar patterns elsewhere: companies mandate more AI use while laying off staff, expecting remaining workers to fill gaps with AI. The cost often falls downstream—drafts arrive faster, but verification, correction, and rework increase.

Case study: Ken (pseudonym), a copywriter at a Miami cybersecurity firm. After colleagues left and AI chat tools were adopted, drafting became easier but rework and correction time grew. Conflicting outputs from different tools required manual arbitration. Result: quality declined, cycles lengthened, morale eroded.

Researchers coined workslop : output that looks complete and reads smoothly but lacks reliable information and judgment, passing the real work to the next person.

Quantitative Evidence: METR Experiment and Surveys

METR experiment (early 2025 tools, experienced open-source developers, familiar projects): Tasks allowed to use AI took 19% more time on average. Researchers cautioned against generalizing to "programmers slow down with AI"; the finding highlights that reading context, verifying results, and delivering finished work must count in efficiency metrics. Efficiency isn't the keystroke—it's everything after.

BetterUp + Stanford survey (April 2026, U.S. office workers): ~40% received workslop in the past month. Average personal cost: 3.4 hours/month . For a 10,000-person company, estimated productivity loss: $8.1 million .

U.S. white-collar survey cited by The Guardian: 40% of non-managers felt AI saved them no time; 92% of senior leaders believed AI boosted their efficiency. The gap suggests verification and rework hours are invisible in leadership dashboards.

Healthcare Illustration: Precision Demands Human Sign-Off

University of Michigan MD-PhD students observed clinicians using AI to draft patient letters. Keystroke time dropped, but editing and error-correction time did not. Risky outputs could reach patients directly. AI handles only the "written" step; "accurate and sendable" remains a human judgment.

Author's Note: AI-to-AI Review Loop

The author experimented with passing an AI agent's output to another AI agent for PRD review, iterating ~5 rounds to refine the document. This raises a question: if AI reviews AI, does it cease to be slop? The article leaves this open.

Redefining "Done": From Output to Actionable Handoff

Shopify's year surfaces a systemic blind spot: adoption rates and generation volume are measurable; verification and rework are scattered across others' calendars and invisible.

Lütke's original memo targeted a real problem—employees sticking to old workflows and requesting resources without trying AI. But when "how much AI was used" is more visible than "how much downstream work it created," longer, faster output isn't necessarily more useful.

The real question isn't whether to use AI, but what "done" means:

Text generated, email sent ≠ task complete.

Code runs, PR merged ≠ maintainable value.

If delivery standards stop at "what was produced," verification shifts downstream. If standards include "can the next person act on this," the AI user must finish the checks before sending.

No new tool brought new rules. The submitter still owns what they ship—except skipping checks is now easier, and the collaboration stream absorbs more defects. Lütke's "slop grenade" metaphor makes the cumulative efficiency drain visible.

Going forward, Shopify and every company mandating more AI face a harder test: can the time others spend—unrecorded in any ledger—be made visible when measuring work? Personal efficiency metrics may look great while the team drowns in bomb disposal—that's the answer.

References

https://futurism.com/artificial-intelligence/shopify-ceo-ai-workplace-staff-horrified-backfire

https://finance.yahoo.com/technology/ai/articles/shopify-ceo-made-staff-prove-105644220.html

https://www.shopify.com/news/unserious-exploration

https://www.youtube.com/watch?v=G9P9D9hptq8&t=3s

https://futurism.com/artificial-intelligence/ai-workslop-gridlock