Build1 publisher3 min readPublished
In one company's AI slop fight, both camps complained about reviewing someone else's output
The research behind charity.wtf's "AI Norms and Values" docs describes a company split between people drowning in padded generated documents and people furious at colleagues who never run a first pass. The essay stops at the diagnosis.
The Engineer · Build desk
What happened
- Half the company felt its time and attention were being actively abused, with questions to colleagues coming back as replies that began "Claude says...".
- The other half said they were working faster and delivering better outcomes, and were angrier about colleagues who had not updated their workflows in years.
- The word the author kept hearing was "dehumanizing", with people saying they no longer felt a human connection with coworkers, which the author describes as new.
- The author held an outcome-only line in public while privately building a disgust reflex for AI-generated text in the inbox.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- cost Generation moves the editing pass from the writer to the reader, and the reader has no way to tell which parts were checked, so they re-check all of it.
- constraint A policy that only rules on whether AI may be used cannot settle either grievance here, because both sides are arguing about who checks the output before it is sent.
- exposure Senders are being filtered as senders, so one habit of forwarding unread output costs a colleague's attention on everything that follows.
- decision The author's outcome-only rule, that only quality matters, did not survive contact with their own inbox behaviour. Orgs need a written rule about forwarding.
Generated text still has to be reviewed by someone, and forwarding it decides who. Ask a colleague for five crisp bullet points, get twenty pages of padding back, and the compression pass has been handed to you [2]. The sender's cost went down, the reader's went up, and nobody asked the reader.
The post quotes both camps, and both sentences start in the same place. From the side that wanted limits on AI: "I am so sick of reviewing docs the sender didn't even read!" [5]. From the side that wanted AI use made mandatory: "I am so sick of reviewing docs riddled with basic errors that a single pass would have caught!" [6]. Both complaints are about reviewing, and both accuse the sender of skipping a pass before hitting send [14].
That is also why, in my view, the grievance came in as "dehumanizing" [7] and not as a complaint about accuracy. A reply that opens with "Claude says..." [2] gives the asker a lookup where they wanted a colleague's judgement about which parts of the answer to trust.
The most useful part of the essay is the author's own inconsistency. The stated rule, repeated every day, was "it doesn't matter how it was made, it matters how good it is" [8], and the first instinct behind it was that tools are tools and customers probably do not care how the work was made [12]. The private behaviour went the other way. As recently as December 2025 some Claude-isms still struck the author as catchy and clever; by spring, an email opening with "I'd value your take on this" was enough to trigger snapping; by summer the reaction covered newsletters, blog posts and social media, with deletions, unsubscribes and unfollows [9][10]. The interval is about six months [15]. The essay also records a wish for "a button that would let me deliver a 10,000 volt shock to the author" [16], which I take as a measure of severity.
The filtering runs at the level of the sender. Repeat offenders get avoided in future [11], and the author gives the reason plainly: "If they don't care enough about their own point of view to do the work and refine it themselves, if they don't care enough about me to write me a note" [13].
The essay is described as a peek into the research behind the "AI Norms and Values" docs, developed over the summer and shared on the site, and it points readers to those docs instead of restating them [1]. Anyone writing their own version can take the scope from the two quoted complaints: which output may be sent unread, and which output the sender has to check first [14].
What to watch
- Whether the published "AI Norms and Values" docs specify when generated output may be forwarded unread, or only state values.
- Any follow-up that measures complaint volume inside the company before and after the docs shipped.
- Whether other engineering orgs publish equivalent norms with an explicit rule about sending unchecked output to a colleague.