When people talk about AI writing, the focus is often on the fluff, the jargon, the important-sounding yet empty phrases, the em dashes, the overuse of vacuous words.
The negative impact of AI writing is more than style annoyances. Writing has always involved a contract between the author and the reader. A writer spends time writing to make it easy for the reader to consume. Communicating effectively with words takes effort. To ensure the reader only has to work to understand the topic, the author must put care into delivering the message.
AI writing flips this. It’s so easy to generate writing. Pages of text appear in seconds, with polish—even diagrams!—ready to send. The cognitive load is removed from the author. The load doesn’t disappear though, it’s transferred to the reader, with interest. The contract is broken.
Writing sends more signals than just the subject matter. AI doesn’t just strip these, it replaces them with signals of its own, unintended, but sent all the same. They confuse, change meaning, and waste the reader’s time.
6 ways AI writing puts a cognitive tax on the reader
1. AI doesn’t think of the reader
We should write with purpose and a goal. We need to identify the reader’s needs, and have them in mind when we write. What do we want the reader to get from this? This heavily impacts what we say and how we say it.
Knowing who the reader is takes more than a job title. What is their perspective? What are they worried about? What will it take to land a message with them and change their thinking? Without supplying all of these judgements, the AI is just guessing.
2. The power of the author’s personality
Writing conveys a lot about the author. Directing AI to write for you removes this – the conversion is lossy.
Tone
Tone carries subtext: sincerity, irony, scepticism, urgency, humour. When AI chooses its tone, the reader may mistake its choices for the author’s. A tentative suggestion can sound like an instruction and routine feedback come across like a formal warning.
A familiar tone keeps us engaged. When somebody’s writing stops sounding like them, that pull is gone.
AI writing feels disembodied. Reading feels sterile and transactional.
Emphasis
An author will deliberately put weight and emphasis on different parts of their writing, even varying it in a sentence. This conveys meaning to the reader about what the author deems important and what they should focus on.
Without this, everything and nothing is important. The reader can’t tell if the emphasis reflects what the author considers important.
Mind-to-mind connection
Reading is inherently a social act between two minds.
When you know somebody, you can read their words in their voice. You know their status, their role, their perspective. There’s no cognitive overhead in deciphering their stance, it’s implicit in who they are. This familiarity provides a cognitive shortcut.
AI obscures that familiar voice, removing the connection. It introduces a trap: the reader is constantly searching for the human and coming up short. This results in cognitive friction, reducing their ability to focus on the core message of the text.
Conviction
AI writing can sound authoritative even when its claims are poorly supported. How confident it sounds is not a reliable indicator of the strength of the evidence.
Research on hallucination argues that training and evaluation can reward models for guessing rather than acknowledging their uncertainty.
When we send those words in our name, that confidence becomes ours in the eyes of the reader. We may be exploring an idea, they see certainty and made decisions.
A message from a CEO with a high level of misplaced conviction will not result in the desired behaviours from their team. A message on a topic from a novice with the conviction of an expert will reduce the reader’s trust. Humility at the right time goes a long way to creating a connection with the reader.
3. The illusion of polish
The medium of communication plays a key role. A phone DM is less formal than an email, and less formal still than a letter in the post. How the message is sent signals something to the recipient.
A polished artefact conveys a finality, the result of thinking, editing, collaboration, and the end of a long process. Reviewing such an artefact would mean taking an idea apart, not building one up. A sketch doesn’t tell, it suggests, it invites exploration.
The reader’s role in this case is anchored on final proofreading. They see their role to wave it through. They skip the deep thinking and the author won’t get the feedback they may be seeking.
Even worse, a reader may spot a flaw in the reasoning of the author, but feel disinclined to raise it due to the perception that course correction will waste a lot of the author’s time. The author sees no feedback as a positive – the chance for improvement is lost. Their belief that AI can produce great output is reinforced.
Compare that to a simple Google Doc with bullet points and some unfinished stubs. Commenting and discussing on the ideas inside it seems natural. You’ve been invited to the thinking part, not the proofreading part, and collaboration happens by default.
4. Proof of work
Costly signalling theory says that the amount of work somebody puts into something is in direct proportion to the meaning and significance we attach to it. Spending time drafting and editing a piece of writing is telling the reader that it’s worthy of their time. It’s a signal we care about them.
Costliness carries meaning. Handmade presents mean more even if they’re less expensive or less useful. We deliver important messages in face-to-face meetings, not over text messages.
With AI writing we’re signalling a lack of value in the output. There may have been a lot of work discussing with an AI model to get to that output, but the reader still sees it as the output from a single prompt: “write this because I don’t want to put the time in”.
Even worse, AI writing mimics costly output via its polish and verbosity. Humans are wired to avoid manipulation through dishonest signals, and will naturally demote the importance of such output even further. The reader will skim-read, matching their effort to their perception of the author’s.
5. Context leakage
It’s common that writing is edited in the same AI chat used to create it. The AI will over-index on adjustments that were made, giving them outsized importance or even including negations explicitly where unnecessary.
The AI has poisoned the output with information from its context that isn’t relevant to the writing. AI has trouble separating the scaffolding from the building itself, and merges it all together. The reader then has to untangle things which adds more cognitive load. Worse, it can change the intended meaning.
6. Defensive reading
When a human encounters AI writing they’re primed to approach it as a detective, aiming to understand the real motive of the human involved. They try and work out where the emphasis should be and what the extraneous fluff and jargon actually means. Their goal is to combat everything we’ve described above.
Only then can they actually try to understand the message intended. Again, the author’s goal is for the reader to understand with as little friction as possible. Defensive reading hinders understanding.
Why people use AI to write
There are lots of reasons people see the shortcut of using AI to write for them as desirable. And it’s understandable.
For some, writing well is a huge undertaking. They may not be writing in their first language. They may be neurodiverse. The effort to simply put words down leaves little room for converting their thoughts into clear writing.
Writing is thinking. It’s taxing. For some they need to be in a certain frame of mind. It can’t be done easily when tired.
You need a dedicated chunk of time, which for a lot of people is precious. Would you rather spend a morning on writing, or a morning on getting other things done and have AI do the heavy lifting?
Using AI as a thinking partner is an increasingly common pattern. You work with an AI to get to a level of understanding (although the actual level of understanding is debatable). Spending more time synthesising that understanding into a format for sharing with others feels like duplicated work so letting the AI do it as they have the context feels natural.
AI can provide a convenient shield for the human from the reader. Writing is putting yourself out there, opening yourself up to criticism. You expose your raw thinking to somebody, introducing a feeling of threat. AI’s words and takes on your thinking move you one step away from the reader. More cynically it can provide a convenient excuse in the face of feedback that the meaning wasn’t really their intention, the AI wrote it.
Proposed norms
Because of all these concerns, I believe the only fix for AI writing is to not use AI to create writing for human consumption. If you use AI to help draft a message, you still owe the reader the care that writing it yourself would have required.
- If you’re sending something that pre-AI would be authored by you, your name is on it and you have to be able to justify every word (see below for more on justifying output).
This prevents output being shared without editing. It forces the author to put themselves in the shoes of the reader.
- Explicitly label provenance of documents (Hand-written, AI-written and human-edited, AI-written)
Forewarned is forearmed. The reader won’t have to spend time decoding, and can choose how much effort to put into reading it.
- Cover note for all AI artefacts if they’re not meant to be in the human’s name
Explaining what the artefact is for and how it was generated enables the reader to make a decision about how they consume that artefact.
- Deliver AI thinking as structured data, bullet points, key ideas, rather than complete prose
Select the points you actually want to share, remove repetition, and make it clear what you endorse. This reduces a lot of the noise in AI writing, and focuses the output on the message to be conveyed. The reader has less to wade through to understand what’s being shared.
Justifying what you share
I had a long conversation with Jonathan Tanner about what it means to justify what you share. We discussed brainstorming and first drafts as potential counter examples. In this case, the author may well believe the output isn’t correct, so must they “justify” it?
I believe that they still can. It’s a first draft of their thinking. It’s an output of a brainstorming session and list of their ideas. If the AI has written it, and has changed the meaning or introduced terminology that doesn’t convey what the human intended, then they cannot justify it – it’s no longer their draft or brainstorm.
Indeed, as I wrote in Work in progress, sharing incomplete and draft documents sends important signals in itself. The sharing at this point is part of a process, an invite to collaborate. Framing it as polished output cuts out that opportunity for the recipient.
Summary
The rise in AI writing has seen an increase in the number of words put in front of others. This hasn’t led to an increase in knowledge sharing, it’s arguably reduced our capability to transfer thoughts and ideas.
As William Zinsser said, if we write, we’re in the storytelling business. And to do this well, “all you have to do is tell a story, using the simple tools of the English language and never losing your own humanity”.
AI makes it easy to feel that the work is finished because the words are there. But somebody still has to work out what you’re trying to say. If you haven’t done that before you press send, you’re asking the reader to do it afterwards. You’re transferring the work to them, with interest.
Thanks
Thanks to Jonathan Tanner for being a sparring partner throughout the writing of this and forcing me to go deeper on the concerns I have with AI writing.
Thanks to Maya Burnand and Daisy McCorgray for reviewing drafts.