Insights
How AI Writing Tools Hide Workplace Dissent
An organisation cannot act on a disagreement it never hears. This is the part of the AI transition I think about most, and it is rarely discussed, because it produces no visible failure.
Consider what happens to a sentence on its way through a modern company. Someone writes a note that is uneven, slightly too blunt, a little uncertain in the middle. Then it is polished — not by a manager or a colleague, but by a model, in four seconds, before anyone else sees it. What arrives is calmer, better structured, and easier to agree with.
Multiply that by every memo, review and proposal in an institution, and something happens that nobody chose.
A large study published this month in Nature Human Behaviour found that when large language models polish and rewrite text, the core content survives — but writing-complexity variance falls by roughly 21 to 50 percent across datasets and models. The organisation still has the information. It has lost the range.
A separate line of work is more uncomfortable. Researchers examining LLM-edited argumentative essays found the rewritten versions frequently failed to convey the writer’s actual opinion. The edit did not soften the position; it relocated it. The writer rarely noticed, because the text sounded like something they might have said.
The obvious reading of this is wrong
It is not an argument that AI writing tools are bad, or that people should write everything themselves out of some romantic attachment to friction. Most organisational writing deserves to be compressed.
What the friction was carrying
For most of the history of management, tone was the cheapest diagnostic instrument an organisation had. Nobody designed it that way. But when someone wrote a memo that was strangely tense, or hedged three times in a paragraph, or used a sharper word than the situation required, that texture told a leader what the content did not: this person is not convinced, or is frightened, or has seen something they cannot yet argue. Learning to read that is a large part of what experienced judgement actually is.
Polish removes exactly that layer, and removes it selectively: the unevenness goes, the conclusion stays. The organisation keeps receiving well-formed agreement and loses the ability to tell a person who agrees from a person who has stopped arguing.
A third finding completes the picture. Across three experiments published this year in Scientific Reports, employees were consistently more willing to voice concerns to an algorithmic leader than to a human one on cognitive tasks, mediated by fairness perception and psychological safety. The dissent has not disappeared. It is simply being expressed somewhere the organisation cannot learn from it.
So the problem is not silence. It is misrouting. The candour goes into a private window; the smoothed version goes into the institution. A company can look more aligned every quarter while knowing less every quarter about what its people actually think.
If that is right, a few things follow for
Decide which artefacts are allowed to be rough. Not all of them — a few. Pre-read notes, dissent memos, early proposals, post-mortems. Say plainly that these are not to be polished, and that unevenness in them is not a lapse in professionalism.
Stop reading fluency as competence. It was never a reliable signal; it is now nearly meaningless, and treating it as evidence penalises the people still thinking in public.
And ask, once a quarter, a question most leaders never ask: where in this organisation is disagreement currently going? If the honest answer is “into a chat window,” that is not a technology problem. It is a report on the safety of the human channels.
We spent two decades teaching people to communicate more clearly. We may be about to discover that clarity, mass-produced, is a way of hiding.