The New Office Etiquette: If You Want Human Attention, Show Human Effort

Somewhere between the third AI-generated design critique and the second pull request review request padded with model-written boilerplate, a lot of engineers have started to feel a specific kind of tired. Tom Bedor, in a short post on his blog, gives that feeling a name and a rule of thumb: "If you are requesting human attention, demonstrate human effort."

The observation is small, but it points at something larger happening inside software organizations right now. As more debug investigations, documents, and code get drafted by machines, teams have stumbled into a new etiquette problem that didn't exist three years ago. The question is no longer whether AI can produce something useful. It often can, especially when wired into a company's internal codebase and documentation. The question is when it's acceptable to pass that output along to another person and expect them to read it.

The moment the annoyance clicks

Bedor traces his own version of the rule to a specific encounter. He proposed a design. A teammate fed it to an AI, asked the model to critique it, and forwarded the result with a disclaimer: "I didn't read this, so it might not be entirely accurate."

His reaction captures the whole problem in one line. If reading it wasn't worth the sender's time, why was it worth the recipient's?

That asymmetry is the core of it. Generating text now costs almost nothing. Reading it carefully still costs exactly what it always did. When one person offloads the cheap half of the work and silently hands the expensive half to a colleague, the math stops being collegial. It becomes a quiet transfer of labor, dressed up as collaboration.

Attention was already the scarce resource

What makes this more than a manners complaint is the underlying economics. Engineering attention was a bottleneck long before language models showed up. Code review, design feedback, and incident write-ups all run on a limited supply of focused human reading time. AI didn't create that scarcity, but it has dramatically lowered the cost of producing things that demand it.

The result is a slow flooding of the channel. If anyone can have a model say something in seconds, the volume of plausible-looking text aimed at any given inbox climbs fast. Each individual message looks reasonable. In aggregate they produce fatigue, and fatigue erodes the trust that makes review work in the first place. A reviewer who has been burned by unread machine output a few times starts skimming everything, which defeats the purpose of asking for review at all.

A rule that doesn't reject the tools

The part of Bedor's argument worth holding onto is that it isn't anti-AI. He's explicit that he sends AI-generated content to teammates when it's useful. The discipline he describes has three parts, and they're cheap to adopt:

• Label clearly what is AI generated, so no one mistakes it for considered human judgment.
• Add your own commentary alongside it, which is where the actual signal lives.
• Review your own AI-written code before asking another person to review it.

That last point is the sharpest. Sending unreviewed generated code to a human reviewer inverts the whole point of review. The author is supposed to be the first line of defense, the person who already caught the obvious problems. Skip that step and the reviewer inherits work the author should have done, plus the cognitive load of figuring out which parts the author actually stands behind.

Why this matters beyond one blog

Norms like this tend to form exactly the way Bedor's did, through small moments of friction that accumulate until someone writes them down. Plenty of teams are independently arriving at similar conventions: tagging AI-drafted sections in design docs, requiring a human summary on top of generated analysis, treating an unreviewed model patch as not-yet-submitted rather than ready for eyes.

There's a market dimension here too, even if the post doesn't make it. The wave of AI coding tools and internal assistants has optimized hard for generation throughput, the speed at which output appears. Far less attention has gone to the receiving end, where someone has to evaluate all that output. The teams that figure out healthy consumption norms, and eventually the tools that bake those norms in through provenance labels and required human annotation, will have a real edge over the ones drowning in unread machine text.

Bedor closes with a footnote promising he wrote every word with, in his phrasing, his meat fingers. It reads as a joke, but it doubles as the thesis. The signal that something was worth a human's time is, increasingly, evidence that a human spent their own time on it first. That's a low bar. The interesting part is how quickly it became one worth stating out loud.