AI Slop Is Becoming a Workplace Tax

The cost of noise

Photo by Swello on Unsplash

A teammate asks a simple question on Slack — “Redis or Memcached for this service?” — and five seconds later, the thread receives a polished essay with headings, trade-offs, caveats, and a conclusion so balanced it refuses to conclude anything. Nobody asked for a literature review. We needed a decision.

This is one of the stranger frictions of the AI era. The problem is not people using AI to write. I use AI — including for this piece, which I drafted with AI assistance and then edited for voice and judgment. Most of us will. The problem begins when someone uses AI to outsource the part of communication that matters most — deciding what is worth saying.

On May 22, 2026, a small page about this behavior hit Hacker News and drew a lively discussion — over 400 points and several hundred comments by the time I read the thread. The page was tiny, almost aggressively so, and the reaction was not about one website. It was about a workplace habit many people already recognized: pasting a huge AI-generated response into a conversation where a short human judgment would have done more good.

The original announcement does not specify measured productivity loss, adoption rates, or a controlled study. So we should not pretend it does. But the pain is real enough to connect with existing research on information overload at work. Microsoft’s 2023 Work Trend Index found 64% of surveyed workers struggled to find time and energy to do their jobs, while 68% said they lacked enough uninterrupted focus time. Asana’s 2023 Anatomy of Work report said workers spent 58% of the day on coordination work rather than skilled work.

Now add cheap machine-generated verbosity to that pile. Congratulations, we have invented a new office chore: reading someone else’s prompt output.

1. The new tax is paid in attention

Workplace communication already had too much ceremony. We had status meetings, pre-meetings, recap emails, alignment docs, decision logs, ticket comments, Slack threads, and the occasional calendar invite with no agenda, which should be treated as a small act of organizational vandalism. AI did not create this mess. It removed the speed limit.

Before AI, writing a 900-word answer to a simple question required effort. Effort acted as friction. You had to care enough, or be confused enough, to spend the time. Now the cost of producing the text is near zero, while the cost of reading it remains very much not zero.

That asymmetry matters. If I spend 20 seconds generating a response which takes five people four minutes each to parse, I have moved work from myself to the group. I may feel productive. The team just paid 20 minutes of attention tax.

And attention is not a fluffy concern. In engineering teams, attention is how incidents get debugged, architecture gets reviewed, security assumptions get challenged, and product trade-offs get noticed before they become expensive. When every conversation becomes a scrollable swamp of synthetic caution, people stop reading closely. Then the one important caveat gets buried next to seven paragraphs of generic “it depends.”

This is why I’m skeptical of the harmless version of the story — “more information is always better.” In a team setting, more text can mean less shared understanding.

2. AI makes weak judgment look expensive

Here is the part that feels socially awkward: AI-generated overcommunication often looks impressive at first glance. It has structure. It has bullets. It has diplomatic phrasing. It sounds like someone did work.

But a lot of workplace writing is not valuable because it is long. It is valuable because it is situated. It knows the current system, the deadline, the team’s appetite for risk, the person asking, and the decision needed next. A generic answer can describe Redis and Memcached. A useful answer says which one we should use in this project and why.

For example, this is not helpful in a team chat:

[embed]

This is helpful:

[embed]

The second version is shorter, but it contains more judgment. It names the decision, the reason, the constraint, and the future implication. It gives teammates something to challenge. The first version gives everyone homework.

This distinction matters because AI is very good at creating the appearance of diligence. It can produce a balanced comparison in seconds. But workplace value usually comes from a narrower move: turning context into a recommendation.

3. The hidden bug is protocol mismatch

As an engineer, I think about this as a protocol problem. Every communication channel has an expected payload size and response shape. A pull request review comment is not an architecture decision record. A production incident channel is not a brainstorming document. A one-line Slack question is not a request for a vendor-neutral white paper.

When we paste a giant AI answer into a small conversational frame, we violate the protocol. The receiver has to do extra parsing:

[embed]

This is why AI verbosity feels worse than ordinary verbosity. A human rambler usually leaks intent. You can hear uncertainty, obsession, urgency, or confusion. AI prose often arrives smoothed flat. It is confident in tone and noncommittal in substance. That combination is exhausting.

A useful engineering checklist is simple:

Before sending AI-assisted text, ask:

• What decision or action should this message enable?
• Can the first sentence stand alone?
• Which parts depend on local context only I know?
• What can I delete without changing the outcome?
• Would I still send this if I had typed every word myself?

That last question is brutal, which is why it works.

4. The before-and-after test

Here is a realistic workplace example.

Bad version:

[embed]

This sounds responsible. It is mostly fog.

Better version:

[embed]

The second message is not anti-AI. Use AI to help draft it if you want. Ask it to compress your notes. Ask it to find missing risks. Ask it to rewrite the final message for clarity.

Just do not paste the thinking foam into the room and call it collaboration.

The responsible workflow is:

[embed]

The lazy workflow is:

[embed]

One of these helps a team move. The other creates a group editing task that nobody agreed to accept.

5. Sometimes the long answer is the right answer

There is a counter-argument worth taking seriously. Some questions deserve depth. “Should we use Redis or Memcached?” may be simple in one system and deeply consequential in another. A migration plan, privacy review, incident postmortem, or database consistency decision can absolutely require detailed writing.

The rule is not “short always wins.” That becomes its own form of stupidity.

The rule is match the answer to the decision surface.

If the decision is small, give the answer first, and add a link or appendix. If the decision is large, write a structured document with owners, constraints, options, and a deadline for feedback. If the conversation is exploratory, say so explicitly. A long message with a clear purpose is not the problem. A long message pretending to be an answer is.

I also think we should leave room for people who write differently, especially across cultures, languages, seniority levels, and neurotypes. Not every verbose message is lazy. Sometimes someone is trying hard to be precise. Sometimes they lack confidence. Sometimes they are translating twice — once between languages and once between workplace norms.

But AI changes the default assumption because volume no longer proves effort. A long response used to signal time spent. Now it may signal one button clicked. So the sender has a new responsibility: make the human part visible.

A good phrase is enough:

[embed]

That sentence does two useful things. It discloses the tool without turning the message into theater, and it makes ownership clear.

6. Product teams should design against verbosity

This is not only an etiquette issue. It is a product design issue.

If workplace tools add AI buttons everywhere, those buttons should not all say “generate.” The default should often be “summarize,” “extract decision,” “list open questions,” or “draft a 3-sentence reply.” Products shape norms. A big shiny compose button teaches people to produce. A good collaboration tool should help people reduce.

I would love to see communication tools add small guardrails:

• A “decision first” prompt for long replies in active threads.
• A warning when a generated message is much longer than the question.
• A one-click “compress to recommendation” action.
• A field for “my actual ask” before posting to a busy channel.
• Thread summaries that separate facts, opinions, decisions, and unresolved questions.

The trade-off is real. Too many guardrails become patronizing. People will hate a tool which scolds them for writing. But subtle defaults can help. The best interface does not shame the user; it nudges the message toward usefulness.

AI vendors often sell writing volume as productivity. In collaboration, volume is often the wrong metric. The real metric is how quickly the right people reach a shared next step with enough context to trust it.

7. Say the thing

The workplace does not need more perfect paragraphs. It needs clearer ownership.

If I ask a teammate for a recommendation, I am not asking them to simulate neutrality. I am asking them to spend their context. Tell me what you think. Give me the one caveat I should not miss. Show me where you are uncertain. Then stop.

AI can help us communicate better, but only if we use it to sharpen intent rather than inflate output. The future I want is not one where every chat message becomes a miniature consulting report. It is one where we spend less time excavating meaning from polished filler and more time doing the work we were trying to talk about in the first place.

The next time we reach for an AI-generated answer, the standard should be simple: do not make your reader pay for words you were unwilling to judge.

References

• No Slop Grenade — the small page that started the May 2026 Hacker News discussion
• Microsoft Work Trend Index 2023 — Will AI Fix Work?
• Asana Anatomy of Work Global Index 2023
• Atlassian — Meeting Overload Is Real
• Hacker News discussion, May 22, 2026