While models have become extremely powerful, and today’s agents are terrifyingly capable, AI can make it easy to pass along something that looks complete while leaving the real work unfinished.
In the world of bug reports, security, and general issue escalations, someone else usually has to validate the claims, recover the context, and find the useful facts.
I’ve started calling this the cognitive AI tax.
cognitive AI tax noun
The avoidable time and mental effort transferred to others when AI-generated material is passed along before the sender has validated, edited, and understood it.
The recipient pays the tax by reconstructing the issue, separating facts from claims, repeating skipped work, and removing irrelevant material.
Taxation without validation
My firsthand experience is in tech, so that’s where I’ll focus today. The cognitive AI tax isn’t limited to technical work, though. It can show up anywhere people use AI.
AI has been able to produce long, technical-looking explanations in seconds for years now. In my experience, something shifted around November 2025: the models became much more capable, and their output became credible enough that humans in technical review roles had to take AI-written reports seriously.
Anyone can now point a model at something, find legitimate issues, and report them to the right people.
These reports land with people and teams that increasingly use AI themselves. But much of the technical analysis and problem-solving still requires a human. When someone forwards AI-generated analysis without validating or editing it, the recipient has to find the actual problem, separate claims from facts, repeat troubleshooting the reporter could have done, and cut everything that doesn’t matter.
In practice, that usually means I have to retrace the entire issue, validate every claim, and hunt for the few useful facts buried in a verbose, bloated, jargon-filled short story. We often get far too much of what doesn’t matter and not enough of what does.
In the worst cases, AI still hallucinates. It regularly overhypes, miscategorizes, makes dangerous assumptions, is flat-out wrong, or completely fabricates things.
After boiling things down, sorting fact from fiction is still half the battle.
That transfer of work is the cognitive AI tax.
The silicon tea party
I’ve begun using AI to battle this tax. I use it to boil reports down to their core issues and start validating their claims.
Unfortunately, there are often gaps in the reporting. The AI didn’t think to look for, or ask the user for, specifics. Sometimes the missing detail is as simple as a timestamp, a backtrace, or an audit of whether there was any actual impact.
Then, instead of condensing the details it judged relevant, the model scatters tiny needles across several haystacks.
To aid my work at Automattic, I’ve been working on AI skills to help people use these models effectively. The goal is to help the user gather the necessary information and write plainly, without bloated and often irrelevant verbosity. To push them to verify and learn. Both the AI agent and the human behind it.
These skills are usually based on four simple principles.
- Treat confidence as presentation, not evidence. Verify factual claims against the source, logs, dashboards, or tools. Mark an unverified claim as reported or suspected, or remove it.
- Make every detail earn its reading cost. Dense, nested analysis is harder to audit and gives unchecked claims more places to hide. Keep details that change validation, scope, report routing, risk, or the requested action.
- Own everything you send. Do not send a draft you have not read. You should be able to explain its claims and answer follow-up questions. AI can produce a first pass, but the sender remains responsible for the result.
- Write for the next person to act. Lead with the request, then summarize the verified work and the remaining blocker. Correct errors as soon as you discover them.
My hope is that these skills help people work more efficiently and effectively, learn along the way, and reduce the cognitive AI tax.
Revolt against the tax
Out of respect for others, I ask that you not impose this tax. Eliminate it completely, or consider paying it yourself before passing the work on to another.
Tax audit
Where have you encountered some form of this tax? It’s not limited to just tech.
Do you impose it on others?
Have you had success with reducing it?
