Your AI Enthusiast Is Right. So Is Your AI Skeptic. That's the Problem.
Charity Majors identified why AI disagreements stay unresolved on most teams: there's no natural feedback loop between the person who moves fast with AI and the person who deals with the quality failures. Here's how to build one.
By Forge Team
If you manage a team that uses AI, you probably have both types: the person who uses it constantly and wants everyone to keep up, and the person who uses it sparingly and keeps flagging problems with the outputs. The uncomfortable part is that both of them are describing reality accurately. The problem isn't that one side is wrong. It's that their feedback loops don't connect — and without a deliberate structure, they never will.
What Majors named
Charity Majors — founder of Honeycomb and one of the more rigorous thinkers on engineering practice — published an essay last week (shared widely by Simon Willison, June 4) with a specific diagnosis. AI enthusiasts face competitive extinction: if they don't adopt AI fast enough, competitors who do will price or outpace them out of the market. Skeptics face entropy: if AI outputs erode the quality of work they're responsible for, the systems they maintain quietly degrade. Both risks are real. Neither side is wrong.
Her diagnosis: there is no natural feedback loop between them.
The enthusiast who uses AI to finish a first draft three hours faster doesn't experience the downstream corrections that fall to the skeptic. The skeptic who catches errors in that draft doesn't see the timeline comparison — how long the work would have taken without AI, or whether the error rate is meaningfully different from pre-AI drafts. Each side has evidence. Neither side has the other's evidence.
The HN thread "Why is HN anti-AI?" (435 points, June 6) confirmed the same pattern across a broader technical community. The responses weren't from people who had never tried AI tools — they were from people who observed specific failure patterns that weren't visible to the AI advocates in the same thread.
What this means Monday morning
The problem isn't convincing your skeptic to adopt AI or your enthusiast to slow down. The management task is building a channel between the two sets of observations so both can update their positions on evidence rather than anecdote.
Three things that actually create that channel:
Make quality failures visible to the enthusiast. If someone downstream is catching errors in AI-assisted work, the person who produced that work needs to see that data — specifically, which task types generate error patterns. Without this, the enthusiast has no reason to adjust their approach.
Make speed data visible to the skeptic. Time savings that aren't shared stay invisible. If AI-assisted drafts are arriving significantly faster but the skeptic only sees the errors, they have an incomplete picture of the tradeoff they're evaluating.
Use checkpoints as the meeting point, not the battleground. A checkpoint where skeptic reviews enthusiast's AI-assisted work isn't a concession to either side — it's the mechanism that generates the shared data both need. The checkpoint catches errors and makes them visible. Over time, that data tells both sides where AI is and isn't working.
Sarah: one shared log, six weeks
Sarah is a marketing director at a 60-person professional services firm. Her team has an AI enthusiast — a senior content writer who uses Claude for every first draft — and a skeptic, the head of client communications, who reviews content before it reaches clients.
The enthusiast's speed was real: documents that took half a day now took under two hours. The skeptic's concerns were real: two client-facing documents in the previous quarter had contained incorrect references to client project timelines. The enthusiast didn't know about those corrections. The skeptic didn't know about the time savings.
Sarah added one thing: a shared log where every significant AI-assisted piece got two data points recorded — production time and any corrections flagged at client review. After six weeks, both sides updated. The enthusiast saw that the error pattern was concentrated in one content type (client-specific timeline summaries) and started writing tighter briefs for those. The skeptic saw that the overall error rate wasn't higher than pre-AI work, and that revision time had dropped by about 40%.
Design a checkpoint structure for one AI-assisted workflow on your team — what gets reviewed, by whom, and what data gets logged from the review.
The version without the log
At a 90-person fintech, the same split played out without the shared data structure. The head of product used AI to generate feature specifications at roughly triple the previous pace. The head of compliance caught more issues in those specs than before — in compliance's experience, quality had dropped. The product lead disagreed; in their experience, output was higher and faster than ever.
Three months of escalating tension, no resolution. The product lead didn't have compliance's error data. Compliance didn't have the volume or timeline data. Neither side was fabricating their experience. They were each describing a real part of the same process, with no mechanism to compare notes.
The disagreement wasn't about AI. It was about an information gap that nobody had thought to close.
Given a specific AI-assisted workflow, decide which checkpoints are essential versus which create friction without catching real errors.
The one sentence
The enthusiast and the skeptic on your team are each holding half of the information you need to run AI well. The management task is building the channel between them — not picking a side.
Like this post?
Get the next one in your inbox. Practical AI skills, no filler.