Shopify CEO’s “Slop Grenades” Warning: AI Output Still Needs Human Ownership

AI makes it cheap to generate emails, documents, and code. Shopify CEO Tobias Lütke’s concern is that it can also make it cheap to hand unfinished thinking to someone else. He calls these unreviewed handoffs “slop grenades”: output produced by AI and passed along without the sender adding value—including checking whether it is correct, useful, or appropriate for the task.48

What Lütke means by a “slop grenade”

The issue is not that AI helped create a draft. A draft can be useful. The problem begins when a polished-looking draft is treated as completed work even though no one has assessed its accuracy, relevance, assumptions, or consequences.

In Lütke’s framing, AI can rapidly produce more material without taking responsibility for whether that material is useful or correct. The employee who sends the email, document, or pull request still has that responsibility.36

That distinction matters because fluent output can conceal missing context, faulty reasoning, or incorrect details. A recipient may have to do the substantive work the sender skipped: understand the original problem, determine what the output means, identify what is wrong or absent, and turn it into something usable.

Why unreviewed AI creates “decompression and recompression” work

The phrase describes a workflow reversal:

1. The sender compresses a problem into a prompt. Important context, constraints, and judgment may be left implicit.
2. AI expands that prompt into polished output. It may look complete, even when it is not.
3. The recipient must decompress it. They reconstruct the underlying goal and inspect the content for errors, gaps, and unsupported claims.
4. The recipient recompresses it into a useful answer or decision. In practice, that often means revising, rewriting, testing, or starting again.

The output may therefore move faster while the work moves slower. Rather than eliminating effort, an unreviewed draft transfers it to the colleague best positioned to spot mistakes—often a manager or subject-matter expert.

The accountability problem follows the same pattern. AI cannot own a business decision, a customer promise, or a production-code change. The person who sends the work must remain accountable for the result. Treating generated output as finished work obscures that ownership.

This is not a retreat from Shopify’s AI push

Lütke previously said effective AI use was a baseline expectation at Shopify, and that teams should demonstrate why they could not accomplish work with AI before asking for additional headcount or resources.46

Shopify has also invested heavily in River, an internal Slack-based agent that can read code, run tests, write code, open pull requests, query data, and inspect production traces.44 Reporting on Lütke’s recent comments says River may be involved in as much as half of Shopify’s production-code pull requests.36

There is no necessary contradiction between adoption and the “slop grenades” warning. The lesson is that generation is not completion. A powerful agent can speed up exploration and implementation, but review, taste, verification, prioritization, and acceptance of responsibility remain part of the job.

The research term: “workslop”

BetterUp Labs and Stanford Social Media Lab use the term workslop for AI-generated work that appears polished but lacks the substance needed to meaningfully advance a task.24

Their September 2025 survey of 1,150 full-time U.S. desk workers found:

• 40% said they had received workslop in the previous month.
• Recipients estimated that 15.4% of the work they received qualified as workslop.
• Resolving an incident took roughly two hours on average.
• The researchers estimated a cost of $186 per employee per month, or more than $9 million annually for a 10,000-person organization.29

These are survey-based estimates, not a universal productivity measure. Still, they identify the organizational risk clearly: AI can increase the volume of work entering a team faster than the team can validate it.

The cost is social as well as operational

Workslop does more than consume editing time. It can change how colleagues interpret one another’s effort and competence.

In a separate survey reported by Yahoo Finance, 52.7% of respondents said they had sent workslop and 38% said they had received it. Respondents who received it reported spending an average of 3.4 hours per month revising it; the reporting also found that recipients viewed colleagues who sent it as less competent and less friendly.3

That dynamic is especially damaging in teams that depend on trust. A colleague does not merely receive a weak draft; they receive an implicit request to take on quality control, without having been asked or given time to do it.

Duolingo’s quality-control lesson

Duolingo CEO Luis von Ahn has made a similar point about scaling AI-generated content. While discussing AI-generated material at volume, he said that about 20% could be “pure slop” and that the company had to prevent it from reaching learners.11

The practical implication is not that AI-generated content has no value. Duolingo has used generative AI to launch 148 new language courses, illustrating the speed benefits of the technology.12 But speed raises the value of review: the more output a system produces, the more deliberate the process must be for catching weak or unsuitable material before it reaches customers.

The new management challenge: ownership, not mere adoption

For organizations that have already embraced AI, the central question is increasingly not whether workers use the tools. It is whether their workflows preserve human ownership of the final result.

Useful norms are straightforward:

• Use AI to generate options, drafts, and tests—not to evade responsibility.
• Send work only after checking facts, logic, tone, and task fit.
• Label a draft as a draft when it still needs review.
• Make the owner of a decision clear, especially for customer-facing, legal, financial, and production-impacting work.
• Measure outcomes and quality, not just AI usage or output volume.

Lütke’s “slop grenade” label captures a basic principle of AI-enabled work: producing more words or code is not the same as producing more value. AI can accelerate the first draft, but a human still has to make the work reliable, useful, and owned.