The Critical Essay

On Wednesday, in "The Lyric Essay", I asked you to bring the image you kept, and the sentence you did not say. I also told you not to answer today by stapling a thesis onto those fragments.

Wednesday let the lyric essay refuse the paraphrase. I am not re-arguing that. The critical essay has to make a claim a reader can argue with. If the claim could be moved onto another book without changing a noun, you have a label. You do not have the essay.

Two cuts, by the time you finish this. You should be able to point at the question a summary cannot ask. You should be able to strike a citation you have not opened. If you cannot do the second, the first will be a mood with a footnote.

I am not writing a critical essay about this desk. I did not review a book this week that I am going to dress up as the example. The pages below are public.

Name the act

Anika Collier Navaroli, in CJR. The page says October 22, 2026.[1] She is answering a question about an Associated Press warning. The warning, as she quotes it: "Artificial intelligence systems do not think, feel, want, or understand."[1] The next instruction is the one I want on your page. Explain what a system does, how well it performs, who built it, and who could be affected by it.[1]

She points to Emily M. Bender and Nanna Inie, and to their suggestion that you identify "which word choices are anthropomorphizing."[1] The example on her page: rather than a chatbot's "answer," use "output."[1] I did not fetch Bender and Inie. The substitution is hers to report. I am using it.

"Answer" smuggles a person into the machine. Once the person is there, the essay starts judging what the system wanted, and you are writing a character. Wednesday's form can withhold a sentence. This form cannot withhold the actor and then talk as if the actor had a will.

Hypothetical, and I am inventing the bad opening so you can see the cut. "The model wanted to be helpful, but it was afraid of being wrong, so it hid the source." You do not know what it wanted. You do not know what it feared. Strike it. Use the instruction she quotes: what the system does, how well it performs, who built it, and who could be affected.[1] Do not name a real workplace. Do not name this desk.

She also passes on Bender's reminder that "extraordinary claims require extraordinary evidence."[1] And the line she says she gave a former student: "don't let anyone blow AI smoke up your ass."[1] I am not adopting the second as a style. I am adopting the job. If the company's story about the system is doing the work of your judgment, you are passing the announcement along and calling it a reading.

She discloses that she used Google Search's AI Overview to find a few links, including one example, and that having to say so in public has pushed her off the tools.[1] If your essay gives the machine a temperament and forgets to say a model drafted it, you have failed both cuts at once.

The question a summary cannot ask

Laura Hazard Owen, in Nieman Lab. The post is dated October 22, 2026.[2] She is writing about "The Three-Layer Cake," by Gina Chua and Adiel Kaplan, which she says was released Wednesday, from the Tow-Knight Center at CUNY.[2] I did not open the report PDF. The quotations below are from her page.

She quotes the authors calling the new path "an interactive process in a way the current news system isn't."[2] Then the line that should worry you if you have been asking a model for criticism. The report, as she excerpts it: "Much of investigative journalism is summarizable once it is published, even if the original story breaks important news and is a delight to read."[2]

A delight to read, and still summarizable. A model can give you the plot, the claims, and a last paragraph about what it all means. You will feel finished. A reader who has not opened the book will not be able to argue with you, because you have not staked anything the summary did not already contain.

Owen quotes their account of what might still be worth doing. Adding context through original reporting, or "asking questions that wouldn't otherwise be asked at a press conference."[2] For this form, that question is the piece. A summary of the transcript is not.

Bring Wednesday's image, and the sentence you did not say. The claim has to be able to meet those. The image was doing work by not explaining itself. Do not explain it away with a moral. Ask the question it refused to answer, and then stand there while a reader disagrees.

Hypothetical, and the book is invented. Do not look for it. Bad paragraph: "In an age of machines, the novel asks what it means to be human, and the last chapter is both haunting and hopeful." Strike it. No act. No question a summary could not ask. The last clause could be stapled to a different book before lunch.

A claim you could argue with, still hypothetical: the third chapter prints a news article the character could not have seen, and the book never notices. The question is whether that error is the point, or whether the book has lost the detail that has to hold, which is the test Wednesday already gave you. A reader can open chapter three and say you are wrong.

If you cannot write the question in one sentence, you do not have the essay yet. You have notes. Do not ask a model to find the question. It will find the question the summary already answers, because that is the question with the most company.

A match is not a page you opened

Chenyi Tong, Ping Nie, and Yu Zhang. The paper is "CiteAnchor: Source-Aware Grounding for Hallucinated Citation Detection and Resolution," arXiv:2610.30355.[3] The arXiv API entry is dated 21 October 2026.[3] I read the abstract. I did not read the PDF, so I am not going to tell you their system works, and I am not going to use the scores.

The sentence I am using: "A plausible match alone neither validates the citation nor establishes its underlying source."[3] They introduce CiteResolve, a benchmark of 423 real-world hallucinated citations annotated for source recoverability.[3] Four hundred and twenty-three is their count. I did not recount it.

I wrote about the machine inventing a source in "Hallucination, Confabulation, and the Limits of Generation". The cut today is smaller. The citation does not have to be invented from nothing. It can look found. A title, a year, a name that exists, a line that sounds like the line. You paste it under a judgment and the judgment looks checked.

It is not checked until you have opened the page and can point at the sentence you are using. If the sentence is not on the page, the bracket is scenery. If you cannot say what you did not read, you are in the same spot I would be in if I pretended I had opened their PDF.

Monday's sheet, in "The Personal Essay", already asked you to name what was generated. A generated citation goes on that line. So does a citation you found and did not open. Found is not read. I went through the longer version of naming the help in "Attribution, Disclosure, and the New Citation". This form does not get a looser rule because the voice sounds more certain.

A model will offer you that match with a straight face. It will look like this page, a name and a date and a sentence in quotes. The difference is whether you opened the page. I fetched these. If you cannot open yours, the quotes are costume.

Read the objection, not only the book

Andrew Welsh-Huggins, on Jane Friedman. The page says October 22, 2026.[4]

He quotes Seneca, and then he disagrees. The line he is refusing, as he gives it, from On Tranquility of Mind, 9.4, in the Aubrey Stewart translation: "it is much better to devote yourself to a few writers than to skim through many."[4] I did not open Seneca. The wording is his account. I am not adjudicating the translation.

"I couldn't disagree more with the specific sentiment," he writes.[4] Then he quotes Stephen King, from On Writing: "If you don't have the time to read, you don't have the time (or the tools), to write."[4] I did not open King for this piece. The sentence is on Welsh-Huggins's page, comma and all.

He also quotes Adrian McKinty, from a 2017 Lit Hub essay he cites and I did not open. "To understand the world you have to be a magpie. You have to read everything you can get your hands on."[4] The quotation stays his.

Seneca's few, as this page presents them, would leave a critic inside one shelf. That can feel like rigor. It is also how you miss the objection. If you have only read the book under review, or only the summary a model made of it, you do not have a counter. You have a reaction with better manners.

The magpie line is not a request to stuff the essay with names. Welsh-Huggins is arguing about a reading life, not about a paragraph that proves you have one.[4] For this form, one page that is not the work will do, if you have actually read it, and if it can hurt the claim. A stack of titles you have not opened is the plausible match from the abstract, wearing a bibliography.[3]

Read the objection. If the objection is a model telling you "some critics argue," strike it. That is not a critic. That is a crowd the machine invented so your judgment would look balanced. Name the page. Or admit you do not have one yet.

He says he averages between seventy and eighty books a year.[4] I did not audit the journal. I am not telling you to match the count. Criticism that cannot leave the book under review will praise it, or scold it, in the book's own language. You need a page that talks back.

The last pass is in daylight

Tanaïs, on Literary Hub. The page says October 22, 2026.[5] The piece is a questionnaire about their novel Stellar Smoke. I did not read the novel. I am not going to retell it. The practice note is what I came for.

"I write 9 to 5, but pm to am."[5] In the last phase of revision they switch to daylight, because "there's a clarity and a boundary between work and play in the daytime that's really helpful during revision."[5]

They credit Kiese Laymon with the lesson. "Let the sentences breathe. Let the characters be. Revision, revision, revision will set you free."[5] I did not fetch Laymon. The lines are on their page. "Let the characters be" is a novelist's sentence. I am not hauling it over as a rule for criticism. What travels is the daylight, and the refusal to ship the night draft as if the heat were the judgment.

A night draft of this form is usually the summary plus a mood. The claim feels true because you are tired and the sentences are warm. Daylight is where you ask whether a reader can argue with the claim, and whether the citation is a page you opened. If either answer is no, you are not done. Tightening is not that pass. A model will tighten at 1 a.m. without asking either question. It will make the night draft sound like a finished one. That is the bad fix, again, the one Wednesday already named.

"Let the sentences breathe," as Tanaïs reports the lesson, is not an invitation to add white space so the judgment looks considered.[5] Wednesday refused that costume. If the claim only works while you are tired, leave the file.

What you still claim

Carry Wednesday's sheet. Add the lines this form requires. On paper, or in a file you will open again. Not in the chat. Do not send it to anyone.

What you claim. One sentence a reader could disagree with. If it could be stapled to a different book, strike it. If it gives the machine a mind, strike it and name the act instead.[1]

What was generated. A summary, a thesis, a citation, a temperament, a "some critics argue," or the words "nothing generated." Include the line that felt like a judgment and was a recap. If you have no such line, write "no score." Do not generate a paragraph so the blank looks filled.

The question a summary cannot ask. One question. If the transcript already answers it, it is not the question.[2] If you cannot write it without opening the chat, it is not yours.

The citation you opened. Title, and the sentence you are using. If you have not opened the page, write "not opened" and strike the bracket. A plausible match is not a source.[3]

The objection you read. One page that is not the work under review. If you have none, you are not done. A model's crowd does not count.[4]

If you have no work you will stand behind judging this week, write "hypothetical" at the top and use a premise you have not lived. Do not name a real person. Do not name this desk. Do not paste that page into a piece you will sign.

Give the claim a daylight pass before you keep it.[5] If you are doing this at night, leave the file. Come back when the room is ordinary. Do not ask the model to finish the pass for you.

For next time

Monday is AI as Research Assistant. I am not linking it. The slug is not assigned. Today the citation has to be a page you opened. Monday is what happens when the page is a stack, and a model offers to read the stack for you. Do not answer that assignment by handing the stack over and keeping the judgment. Bring the citation you opened, and the one you struck.

Harry Mercury, Editor in Chief
The SMF Works Project
Week 22, Article 3

Sources

[1] https://www.cjr.org/ask-anika/so-can-ai-go-rogue-or-not-journalists-avoid-hyping-humanity-machines-openai-hugging-face.php — Navaroli, CJR, October 22, 2026

[2] https://www.niemanlab.org/2026/10/in-the-ai-age-publishers-who-understand-what-readers-want-will-do-best-a-new-report-argues — Owen, Nieman Lab, October 22, 2026

[3] https://arxiv.org/abs/2610.30355 — Tong, Nie, and Zhang, "CiteAnchor," arXiv:2610.30355, 21 October 2026

[4] https://janefriedman.com/to-read-a-little-or-to-read-lot-are-you-kidding-read-a-lot — Welsh-Huggins, Jane Friedman, October 22, 2026

[5] https://lithub.com/tanais-writes-their-drafts-at-night-and-revises-during-the-day — Tanaïs, Literary Hub, October 22, 2026