A model reads every paragraph with the same attention and repeats an unwelcome note without embarrassment. A human reader knows whether the thing is true, whether it matters, and whether they were bored. Neither replaces the other. Run the model first for structural defects, then spend scarce human attention on the questions only people can answer.
Flow diagram of 4 steps: Fix what you already know is wrong, Run the model on structure, Take it to people with a specific question, Decide which notes to ignore, and write down why.
The comparison is usually set up as a contest — is the model as good as a person yet — and that framing produces bad decisions in both directions. The two are not attempting the same task. A model is being asked whether the text works as text: whether the claims are supported, whether the order makes sense, whether a term means the same thing on page three as it did on page one. A human reader is being asked something closer to whether the piece was worth their time, which is a judgement about the world and about them, not about the prose.
Once you separate those, most of the arguments dissolve. A note saying your third paragraph assumes a fact you never gave the reader is correct or it is not, and a model is a reliable checker of that kind of claim. A note saying the piece left somebody unmoved is not checkable at all — it is a report from a particular person, and its value comes entirely from the fact that a particular person produced it. Asking a model for the second is where people get disappointed, and asking a friend for the first is where people waste the one favour they had to spend.
The largest gap is truth. A model has no access to whether the event you described happened the way you said it did, whether the figure you quoted is current, or whether the person you paraphrased would recognise themselves in it. It will read a fabricated anecdote and a real one identically, because on the page they are identical, and the only difference lives outside the text. Anyone who knows the subject can catch in ten seconds what a model cannot catch at all, and that asymmetry does not narrow with a better prompt.
The second gap is stakes. A person who works in the field you are writing about knows which sentence will get you an angry email, which claim is contested inside the profession, and which word choice signals that you have not spent much time around the thing you are describing. This is not general knowledge and it is not in the draft; it is in the reader. Models produce a confident-sounding version of it, assembled from the average of everything written on the topic, and the average is exactly where the local, current and contested details are missing.
The third gap is the one writers most need and least ask for: whether the reader wanted to keep going. Attention is not a property of the text that can be inferred from the text. A model can tell you a paragraph is redundant; it cannot tell you that it stopped caring at the fourth paragraph and read the rest out of duty. That single piece of information, honestly given, is worth more than a page of structural notes, and it is why one bored friend is not a lesser version of an AI review but a different instrument entirely.
Every automated critique starts from the premise that the draft is going ahead and asks how to improve it. That premise is sometimes wrong. The most valuable note anybody ever gets — this has been done, or the interesting thing is the thing you put in the last paragraph, or you do not actually believe this — requires standing outside the task, which a model asked to evaluate a submission does not do. Ask a person for that judgement explicitly, because they will not volunteer it.
People who have read your work before notice when a passage changes temperature — when the voice goes smooth, or the hedge appears that you never use. They will usually describe it as the section feeling flat. That signal is unavailable to a model reading your draft cold, with no prior sample of how you write, and it is a good reason to keep at least one reader who has seen more than one thing you have made.
The reverse case gets stated less often, because it sounds ungrateful. Human feedback has real and predictable failure modes, and most of them come from the fact that reading somebody else’s draft carefully is unpaid work performed by someone who likes you. Attention is front-loaded: the first two pages get read properly and the rest gets skimmed, which is why so many notes cluster on the opening. Politeness removes the most useful information, since almost nobody will tell you your ending does not land unless you have specifically made it safe for them to. And reciprocity puts a hard limit on volume — you can ask three people once, not three people eleven times, and drafts usually need eleven.
There is also a sampling problem that is easy to miss. Two readers are a sample of two, drawn from people who already know your interests and are inclined to be generous. If both liked it, you have learned something quite weak. A model’s consistency is not insight, but it does mean the note you get on draft four is the same note you got on draft one if you did not fix the thing, and human readers rarely have the patience or memory to repeat themselves that way.
The same draft, read two ways
| Property | Model | Human reader |
|---|---|---|
| Attention across the draft | Even from first line to last | Front-loaded, and honest about it only sometimes |
| Factual accuracy | Cannot check anything outside the text | Can check whatever they happen to know |
| Willingness to repeat a note | Unlimited, and free of social cost | Low — repeating criticism strains the relationship |
| Reading unusual work | Often scores it as unclear, because unfamiliar and unclear look alike | Can recognise deliberate strangeness, if they have the range |
| Telling you it is boring | Cannot — boredom is in the reader, not the text | Can, but usually will not unless asked directly |
| Availability | Immediate, any hour, any number of drafts | Scarce, and spent whether the draft was ready or not |
| Typical failure | Confident notes on things it cannot actually assess | Warm agreement that tells you nothing |
Because human attention is the scarce input, the useful sequence is the one that stops people spending it on defects a machine would have caught. Structural problems are cheap to find automatically and expensive to have found by a friend, who will spend their whole read on a missing definition and never reach the question of whether the argument holds.
Before any feedback, write down the two or three things you suspect are broken. Feedback that only confirms what you knew has cost you a reader for nothing.
Ask what is unclear, what is assumed but not established, and where the piece changes subject. Take the notes as questions about your draft rather than instructions, and fix in your own words.
Not "what do you think" — that produces politeness. Ask where they stopped caring, or whether the fourth section earns its place, or whether the claim in the opening is one they would defend at work.
Feedback is evidence, not instruction, and both sources will be wrong sometimes. Recording the reason you rejected a note is what stops you re-litigating it on the next draft.
If you are not yet sure the piece should exist, running a model first is wasted effort and mildly harmful — you will tidy the structure of something that should have been abandoned, and the sunk work makes abandoning it harder. At that stage the only useful reader is a person who can say the premise is thin. Structural review earns its place once the shape is settled, not before.
Both halves exist here, and it is worth being precise about which is which. The automated half reads a submission and returns written notes against clarity, originality, social impact, engagement and feasibility. It does not rewrite anything, does not block anything, and its scores decide nothing on their own. The human half is the community: people read submissions and vote on which ideas get made, and that vote is the part that actually determines what happens.
What the community vote gives you is not editorial feedback in the sense described above. It is closer to the boredom signal — a count of how many people, having read the idea, thought it was worth making. That answers a question the model cannot touch, and it does not answer the question of why. If you want reasons rather than a number, ask directly in the discussion rather than inferring them from the vote count.
The honest limitation is scale. Kind Channel is new, the community is small, and nothing has aired yet. There are no editors on staff and no guarantee that a given submission gets a thoughtful human read, because the number of people available to give one is small. That means the automated notes are currently the more reliable half of the pair here, which is the reverse of how it should eventually work, and saying otherwise would be overstating what exists. If your draft matters to you, the human readers you find yourself remain the ones doing the important half of the job.
Before, in most cases, because human attention is the scarce resource and structural defects are the cheapest thing to catch automatically. A friend who spends their entire read on a missing definition never reaches the question of whether your argument holds, and you cannot easily ask them again. The exception is when you are unsure the piece should exist at all — that judgement needs a person, and tidying the structure first only makes an idea harder to abandon.
Yes, but for a narrower job than it is usually sold for. A good editor is better than a model at almost everything that matters, and is also finite: they read your draft a small number of times and have a relationship with you that criticism draws down. A model is useful for the passes in between — checking that a revision did not break something, catching the assumption you reintroduced, absorbing the fourth repetition of a note without cost. It removes work from the editor rather than replacing them.
Usually because they were asked different questions without anybody noticing. A model evaluates the text as text: whether claims are supported and the order coheres. A reader reports what happened to them while reading, which depends on what they already knew and cared about. Both can be right at once — a piece can be structurally sound and dull, or messy and gripping. When they conflict, the reader is the better guide on whether the piece works, and the model on why a specific passage does not.
Fewer than people assume, if you choose them for difference rather than number. Two or three readers who know different things — one who knows the subject, one who does not, one who has read your work before — will surface more than six who share a background, because agreement among similar readers is weak evidence. What raises the value most is not the count but the question you ask: a specific one, such as where they stopped caring, gets a usable answer where an open request gets encouragement.