Stop Rewriting AI Output to Sound Human

The trendiest sentence right now is “this tastes too much like AI, we need to de-AI it.”

Author: Koutian Wu; GitHub: ktwu01

So a ridiculous kind of labor has appeared: AI produces a slide deck, an email, a proposal in two minutes, and humans spend two hours editing it to look like another human stayed up all night making it. Shorten the sentences a bit, dumb down the titles a bit, make the layout a little less tidy, swap “clear” for “professional,” then pull ten people into a meeting to confirm the page looks like a proper company.

Most of the time this does not improve the content. It just pays protection money to the most conservative people in the office.

The word “professional” is often an old guard’s polite term for familiarity. They are used to dense PowerPoints, to pages paved with filler, to humans deliberately splitting one sayable sentence into three levels, five color blocks, and ten minutes of talking. When they see material that AI made clean, complete, and direct, they first ask “does this look like what we used to make,” and put “is it useful” second.

Of course it does not look like it. The old way was inefficient to begin with.

AI slop does have real problems. It likes to pile up words, to wrap one judgment in a paragraph of steaming correct nonsense, to manufacture the illusion of information density. To grab attention, it writes every sentence like a conclusion and every conclusion like a product launch. Those defects deserve deletion, so delete them.

But many people merge “cutting filler” and “reshaping into human habits” into one job. The first compresses information; the second often restores inefficiency. AI wrote one page a reader can understand, and humans insist on adding background, vision, slogans, org charts, and acknowledgments before it looks like a proper report. AI wrote one explicit request, and humans insist on padding it with pleasantries, boilerplate, and strategic vagueness before it feels “tactful.”

The core of this is obedience.

The Musk types are right about at least one thing: reasoning down from first principles on many tasks never required asking office tradition for permission first. A slide’s job is to help people understand and decide, not to prove its maker knows PowerPoint. An email’s job is to get something done, not to make the recipient feel the sender received fine corporate-format training. A proposal’s job is to expose judgments and costs, not to let everyone find their familiar jargon inside it.

Human language was never efficient. It is full of probing, politeness, identity checks, in-group codes, and historical baggage. Many occasions need all that, because people are not APIs and collaboration is not data transfer. When you quote a client, comfort a friend, negotiate, lead a team, or handle conflict, of course you cannot treat people as tokens. Tone, white space, and relationships all have prices.

But once inside goal-directed work, clinging to that baggage turns absurd. Internal syncs are not literature, project weeklies are not palace memorials, and technical decisions should not become cosplay contests over who looks more professional. Expression that helps people understand fast has already done the most important job. Not sounding human may be exactly its advantage.

More bluntly: many people say they hate AI slop, while what stings is AI devaluing the craft they used to look professional with. Turning an empty slide into something that looks busy used to be a skill; now AI does it better in thirty seconds, and the skill suddenly costs nothing. So they invent new standards: it must carry human traces, human warmth, visible hours you spent.

Why? Hours spent produce no value by themselves.

Humans should of course edit AI output, but only two kinds of edits: facts are wrong, or it gets in people’s way. For the rest — “looks too AI,” “not professional enough,” “off our usual style” — first ask: whom does this requirement serve? The reader, the decision, or some people’s nostalgia for the old order?

If the answer is the last one, leave it alone.

Don’t train AI to act like a cautious, verbose, approval-seeking human. There are already enough humans.