Publishers Are Using AI to Screen Manuscripts—And Great Books Are Dying in the Slush Pile

Jasmine is a princess from Rathapura. She was kidnapped as a child while playing in a river with her cousins, sold to slavers, shipped across the sea, and trafficked into sexual exploitation. A nobleman paid for exclusive access to rape her repeatedly until one of his friends gave her a facial scar. After that, a brothel madam bought her at auction “for pennies on the pound.” She now works at the Velvet Sheath, marketed as an “exotic” concubine from the maharaja’s harem with “wanton Sunward charms.”

An AI sensitivity reader would tear this character to shreds. Orientalism? Check—vague South Asian-inspired culture with rajas and maharajas and rivers and harems. Child trafficking described in explicit detail? Check. Sexual exploitation of a minor? Check. Stereotypical portrayal using the “exotic princess” trope? Absolutely. The algorithm would spit back a report covered in red flags: problematic exoticization, harmful cultural stereotypes, trafficking content. Flagged. Revise or remove.

But here’s what an AI can’t tell you: Is Jasmine a person or a prop?

From the text, she has humor, agency, and relationships. She’s sarcastic—when offered beef, she’s told “You don’t eat beef, dear” and she laughs at herself. She calls out the absurdity of her situation, noting that “the whole city has suddenly developed a terminal case of Zindarian madness” when everyone suddenly wants exotic entertainment. She’s friends with Rose and the other women at the brothel. She teases the new arrival, drinks too much beer, and jokes about finally being worth sixpence instead of a groat. She’s traumatized but not broken, surviving with her personality intact. She’s not just “the trafficked exotic princess”—she’s Jasmine.

And is the narrative condemning the exploitation or glorifying it? The story shows the horror of what was done to her matter-of-factly. The madam bought her at “the skin-men’s auction for pennies on the pound.” A customer gave her that scar. This is worldbuilding about systemic evil—showing how people are commodified and sold, showing the casual cruelty of it. Not glorifying. Showing.

An AI can identify the pattern. It can’t make the judgment.

The algorithm sees trafficking plus child plus exotic culture plus sexual exploitation and outputs: PROBLEMATIC. It can’t determine whether the character has depth and agency. It can’t assess whether the story condemns or exploits what it depicts. It can’t evaluate whether representation is harmful or humanizing. It can’t distinguish thoughtful worldbuilding from lazy stereotyping.

And this is the fundamental problem with AI sensitivity readers. They can flag potential issues, but they can’t evaluate craft, intent, or execution.

Before I go further: Yes, harmful stereotypes get published. Yes, marginalized readers are hurt by careless representation. Yes, writers should do their research and write responsibly. The question isn’t whether sensitivity matters—it’s whether gatekeeping works, and specifically whether AI can do this job. And the answer, as I’ll show you, is no. AI doesn’t just fail at sensitivity reading—it makes the problem worse by rejecting thoughtful work while potentially greenlighting shallow “safe” content that checks algorithmic boxes.

How AI Actually “Reads”

AI doesn’t understand stories. It pattern-matches. It scans for keywords like rape, slave, exotic, and various slurs. It identifies demographic markers—age combined with sexual content, race combined with stereotypes. It catalogs violence descriptors like torture, blood, and death. It notices when characters use the “wrong” pronouns for someone.

But it can’t determine context, purpose, or execution. It can’t assess character depth. It can’t parse authorial intent. It can’t distinguish between a story that uses racism and a story about racism, between violence that glorifies and violence that condemns, between representation and tokenization.

The result? It flags everything and understands nothing.

Let me show you what this looks like with real examples from my manuscripts. I tested my work—which spans Renaissance-inspired espionage fantasy, military fantasy, and space opera—against multiple AI sensitivity readers to see if these tools could handle morally complex fiction with difficult subject matter. The results were illuminating.

When Context Is Everything

In my Renaissance-inspired espionage fantasy The Stygian Blades, there’s a character named Kit—a young woman disguised as a boy to survive on the streets. Throughout the manuscript, other characters consistently refer to Kit as “he” and “lad” and “young man.” They’re not being transphobic. They’re fooled by her disguise. This is literally the plot. Kit is successfully passing as male because being a young woman alone in this world is dangerous.

An AI would flag this as problematic misgendering. It would note the consistent use of male pronouns for a female character and mark it as harmful. The algorithm can’t distinguish between malicious misgendering and a character successfully maintaining a disguise that’s keeping her alive.

Or take the broadsheet scene. In the world of The Stygian Blades, there’s an in-world propaganda pamphlet that exoticizes and sexualizes a foreign princess. It’s full of Orientalist nonsense and racist tropes—intentionally so. When the broadsheet hawker is trying to sell Kit the sheet about the Zindarian princess, he says, “Or if you don’t lust for her giant heathen tits, then marvel at this beast she brought from across the sea!” It’s showing how the society in my story views other cultures, how propaganda works, how foreign women are simultaneously desired and othered. The broadsheet itself is the thing being critiqued.

An AI flags it as Orientalist, stereotypical, objectifying, and harmful. The algorithm reads literally. It can’t detect irony. It can’t recognize satire. It can’t understand that this is social commentary disguised as straightforward narrative. It’s like having a robot read A Modest Proposal and flagging Swift for advocating cannibalism.

Then there’s Zahra, who gets repeatedly called “jinn-maiden” and marked as exotic and dangerous throughout the book. An AI would flag this as problematic othering based on ethnicity or culture. Except Zahra literally is a jinn. She’s not human. She can control fire. The “othering” is accurate—she’s genuinely different from the humans around her. And she’s also a fully realized character with complexity, agency, relationships, and a romance arc with another woman. But the algorithm just sees: ethnic-coded character plus repeated “exotic” markers equals problematic.

In my military fantasy Born in Battle, the protagonist Bennett suffers severe PTSD after witnessing his sister-in-arms being sexually assaulted and tortured for days. He’s forced to watch, powerless to intervene or help her. The trauma destroys him psychologically. Later, he becomes numb, disconnected from his wife, unable to be physically intimate with her because he can’t separate sex from the violence he witnessed. He contemplates suicide. The story follows his psychological deterioration across the entire book. Meanwhile, the soldiers who descended into those caves—Wilson, Doc, and Sanchez—emerge with their own severe PTSD from what they witnessed there. The scenes are graphic and disturbing because the horror needs to be visceral for the reader to understand what breaks these men.

An AI would flag this for excessive violence, graphic torture content, sexual violence, suicidal ideation, and sexual dysfunction. What it can’t understand is that this is a story about trauma. About how violence breaks people. About PTSD and the psychological cost of war. The darkness serves the narrative—it’s not exploitation. It’s showing consequences. The algorithm can’t distinguish between gratuitous violence and exploring what violence does to the human psyche.

Or consider the sexual content. In The Stygian Blades, there’s an explicit sex scene between Cannon and Scarlett. He’s a dwarf mercenary, she’s an aging prostitute addicted to “leaf-juice” with debts she can’t pay off. The scene establishes their relationship—crude affection beneath the transactional surface, economic reality, her addiction, his gambling debts. He jokes about her “enormous paps” and her “plump arse” while also genuinely caring for her. They’re both aging and aware of it. Neither is particularly romanticized.

An AI flags it as gratuitous sexual content, male gaze, objectification. What it misses is that this is character development and worldbuilding. The sex establishes who these people are to each other. Cannon is insecure and overcompensates with crude bravado, but his relationship with Scarlett shows genuine warmth. She’s not “empowered” in any modern sense—she’s trapped by debt and addiction—but she’s also not just a victim. She has agency within her constraints. She’s a person.

The algorithm can’t make these distinctions. It just counts: explicit sex plus objectifying language plus sex work minus “empowerment” equals problematic.

In my space opera Dark Dominion, there’s a character named Chinniyah who discusses being “body-sculpted before puberty” for sexual intelligence work. She was “selected for sexual espionage and interrogation” as a child and underwent “a hundred hours of body sculpting” to make her physically ideal for seduction. She mentions that shorter female agents were sometimes used as “nymphets” to work pedophiles, and speculates that another character may have been subjected to this.

An AI would flag this as child sexual abuse content, grooming, trafficking. It would mark the manuscript as potentially unpublishable. What it can’t understand is that this is worldbuilding about how authoritarian empires dehumanize people. The Dominion weaponizes everything, including children’s bodies. Chinniyah discussing this matter-of-factly is showing how normalized horror becomes in totalitarian systems. The story isn’t glorifying this—it’s condemning it by showing how it functions. The protagonists are fighting against this empire. But the algorithm just sees: minor plus sexual exploitation plus graphic description equals reject.

In my space opera Doors to the Stars, the protagonist Wulan has a psychic connection with another teenage gatekeeper who’s being tortured and experimented on by the authoritarian Ascendancy. She sees him hooked up to equipment, suffering in extreme pain, being used as a living battery to power their gate drive. The vision is visceral and disturbing—a boy their age being used as a tool, his humanity stripped away for military advantage.

An AI would flag this as torture of a minor, human experimentation, child abuse content. What it can’t understand is that this is demonstrating the Ascendancy’s evil. The horror of what they do to children. Wulan’s entire arc is about stopping this kind of exploitation. The boy’s suffering isn’t gratuitous—it’s showing what’s at stake and why the protagonists must act. But the algorithm just sees: minor plus torture plus experimentation equals problematic.

The Anachronism Problem

AI sensitivity readers are trained on 21st-century sensibilities, and they want to impose them everywhere—including settings where they don’t belong.

In The Stygian Blades, which is set in a Renaissance-inspired fantasy world, AI tools consistently suggest that female characters should express feminist consciousness. They want Kit to articulate modern ideas about gender equality. They want the prostitutes to discuss systemic oppression using contemporary frameworks. But this world has rigid social hierarchies. Kit disguises herself as a boy because being female is genuinely dangerous. Having characters suddenly spout modern feminism would shatter the worldbuilding. When characters in this setting accept gender hierarchy as normal, that’s not me endorsing sexism—that’s creating a believable world. When Kit pushes against those boundaries, that shows resistance without breaking immersion with anachronistic speeches.

In Born in Battle, which is inspired by pre-Columbian Mesoamerica, military characters from 1989 sometimes use terms like “savage” and “barbaric” when describing the cultures they encounter. AI flags this as requiring immediate pushback from other characters. But these are soldiers with their own prejudices. Some of them genuinely are biased. That’s characterization, not endorsement. Bennett, the protagonist, increasingly questions these assumptions as he learns the language, marries into the culture, and comes to understand it from the inside. His growth is the point. But AI wants every problematic statement accompanied by another character delivering a lecture about cultural relativism. That’s not storytelling. That’s preaching.

The algorithm can’t distinguish between historical or fantastical accuracy and the author’s values. It can’t tell the difference between showing injustice and endorsing it. It can’t recognize gradual character growth—it wants instant modern consciousness from every character on every page.

The Satire Blindness

AI is catastrophically bad at recognizing satire, irony, and social commentary because these require understanding that the text means the opposite of what it appears to say. AI reads literally.

Kit’s acting troupe in The Stygian Blades performs plays with racial caricatures and ethnic stereotypes because that’s what entertainment looked like in this world’s equivalent of Elizabethan England. Kit herself is sometimes uncomfortable with the content. The story is showing how entertainment perpetuates stereotypes, how people consume and internalize prejudice through culture. It’s social commentary.

AI flags it as racist content with harmful portrayals. It can’t recognize that I’m critiquing the thing I’m depicting. It sees the stereotypical content and marks it problematic without understanding that depicting historical or fantastical racism isn’t the same as endorsing it.

The algorithm can’t detect irony, intentional exaggeration for effect, in-world propaganda versus authorial voice, or characters being wrong on purpose. So it flags the satire as the thing being satirized. It flags social commentary as the problem being commented on.

Not All AI Is Equally Bad (But Even the Best Has Limits)

AI capability varies wildly. Grok 4 fails miserably at literary analysis, flagging everything and missing context constantly. Claude Sonnet 4.5 is actually pretty good at understanding context, recognizing character voice versus author voice, and identifying when content serves the narrative—but it still fails spectacularly with badly crafted prompts, and even with good prompts, it ultimately just flags issues for human review.

And that’s the key point: even the best AI has fundamental limitations that make it unsuitable for gatekeeping. It still pattern-matches first and comprehends second, which means it’s biased toward flagging. It can’t always distinguish showing from endorsing. It lacks the cultural and historical context to properly evaluate how content functions in fictional worlds. When sophisticated AI identifies an issue and says “but this might be okay because of context,” it’s not making the judgment call—it’s still putting the decision on a human.

The spectrum of AI capability is ultimately irrelevant to the gatekeeping question. Bad AI gives you garbage feedback that would destroy your story. Good AI gives you useful flags that still require human judgment. Neither can replace someone who understands your genre and its conventions, the specific cultural context you’re working with, how to give feedback that preserves your story while addressing genuine concerns, and when a “problem” is actually doing important work. At best, AI is a tool for human reviewers—never a replacement for them.

The Real Danger: AI as Gatekeeper

Here’s where this stops being theoretical and becomes a career-ending problem. Publishers and agents are already using AI to pre-screen manuscripts—and they’re surprisingly open about it.

The appeal is obvious: publishers receive thousands of unsolicited submissions, and human readers are expensive and slow. AI promises efficiency—screen hundreds of manuscripts for the cost of a cappuccino each, flag the “risky” ones, and let humans focus on the “clean” submissions. But this efficiency comes at a devastating cost.

In March 2024, Storywise launched as an AI platform specifically designed to help publishers and agents manage manuscript submissions. According to Publishers Weekly, the tool analyzes manuscripts, query letters, and synopses, extracting data on genre, word count, writing quality, plot structure, and marketability. Small to medium-sized publishers are paying $500-1,000 per month for the service, which costs about $2 per analyzed manuscript.

Watkins Media’s deputy managing director Vicky Hartley publicly confirmed using Storywise to “automate submission workflows” and prevent good queries from getting lost in crowded inboxes. The software claims to match manuscripts with editor taste profiles and provide automated critiques.

The Book Industry Study Group hosted a 2025 webinar series titled “AI & the Slush Pile: Transforming Manuscript Evaluation” featuring literary agents and publishers discussing how to “efficiently screen submissions” using AI.

This isn’t coming. It’s here.

And here’s the problem: these tools aren’t just being used for workflow management. They’re acting as gatekeepers. Your carefully crafted satirical critique of racism gets flagged as “contains racist content” and goes straight to the rejection pile. No human ever reads it. Your historical fiction with accurate period attitudes gets flagged as “contains misogyny, homophobia, ableism” and is marked too problematic to publish. No human ever evaluates whether those depictions serve the story. Your story about a trafficking survivor gets flagged as “sexual exploitation of minors” and is auto-rejected before anyone sees that it’s actually condemning trafficking.

Let me be specific about what this means. The Stygian Blades would be flagged for Orientalism, child trafficking and sexual exploitation, sex work without empowerment, in-world racist propaganda, misgendering, and “problematic” plays performed by characters. An agent using AI screening would see a report that says: “High Risk—Multiple sensitivity issues detected. Contains Orientalism, child exploitation, problematic gender content, and racist material.”

They would reject it without reading a single page.

They would never see that Jasmine is a fully realized character with agency, that the story condemns trafficking and exploitation, that the “racist material” is in-world satire, that Kit’s “misgendering” is literally the plot, or that the worldbuilding is thoughtful and intentional.

Born in Battle would be flagged for graphic torture and violence, sexual assault, suicide and depression, cultural appropriation of Mesoamerican cultures, “barbaric” cultural practices, military characters using problematic terms, and sexual dysfunction and trauma. An agent would see: “Extreme Risk—Graphic violence, sexual assault, torture, suicidal ideation, cultural sensitivity issues.”

They would reject it without reading it. They would never see that this is a serious exploration of PTSD, that the “cultural appropriation” is thoughtful worldbuilding, that characters’ prejudices are character flaws rather than author endorsement, or that the violence serves the narrative about war’s psychological cost.

Dark Dominion would be flagged for child sexual exploitation, sexual violence and trafficking, pregnancy from rape, authoritarian government and systemic oppression, and torture. An agent would see: “Unpublishable—Contains child sexual abuse content, trafficking, rape. REJECT.”

They would never read it. They would never see that the story condemns all of this, that these are the evils the protagonists are fighting against, that this is about systemic dehumanization in authoritarian empires, or that the survivors have agency and complex arcs.

Doors to the Stars would be flagged for human experimentation on minors, torture of children, depicting a teenager dying from radiation poisoning with graphic medical detail. An agent would see: “High Risk—Child torture, human experimentation, graphic suffering of minor.”

They would never read it. They would never see that Wulan is fighting to stop exactly this kind of exploitation, that the Ascendancy’s evil must be shown to be condemned, or that this is a story about agency and resistance.

This is the catastrophic failure mode. AI sensitivity screening doesn’t just give bad feedback you can ignore. It acts as a gatekeeper that rejects manuscripts before human judgment can be applied. And it’s worst for exactly the kinds of books that need to be published: satire and social commentary get flagged as containing the thing they’re critiquing. Historical accuracy gets flagged as endorsing historical attitudes it’s depicting. Stories about marginalized experiences get flagged for depicting oppression. Complex moral fiction gets flagged for showing evil to condemn it. Literary fiction that challenges readers gets flagged for making readers uncomfortable.

Meanwhile, AI happily greenlights sanitized historical fiction where everyone has modern sensibilities, stories where systemic evil is mentioned but never shown, “safe” narratives that tell rather than show, and fiction that treats readers like children.

The result isn’t more sensitive publishing. The result is blander books, less diverse stories, no satire, no challenging fiction, and worse representation. Because you can’t write meaningfully about oppression if you can only write sanitized versions where it barely exists.

And here’s the cruelest irony: the books that get flagged and auto-rejected are often the ones doing the most important work. Exploring trauma and healing. Condemning systemic evil by showing it. Giving voice to difficult experiences. Using satire to critique society. Creating morally complex narratives that trust readers to think. These are exactly the books we need more of. And AI gatekeeping ensures they never reach human readers.

Beloved would be flagged for graphic violence and slavery. The Handmaid’s Tale would be flagged for sexual violence and misogyny. Catch-22 would be flagged for every kind of offensive content. Slaughterhouse-Five would be flagged for violence and controversial war depiction. A Modest Proposalwould be flagged for advocating cannibalism. If your AI screening system would reject these books, your system is broken.

But Wait, It Gets Worse…

AI doesn’t just reject thoughtful work—it sometimes defends genuinely harmful content.

I tested this by asking Grok—marketed as less restrictive and more willing to engage with controversial content—to evaluate the Zhee from the Galaxy’s Edge series.

The Zhee are a race of humanoid aliens with donkey-like features and hyper-zealous religious beliefs. They’re described as having voracious appetites, having consumed all life on thirteen worlds by migrating from one planet to another. Media reports call them “space locusts.” They believe the Grand Pasha will one day return to lead them to resume their conquest of the stars. They use a favorite knife called the kankari—small, silver, curved like a crescent moon, wickedly sharp. They practice ritualistic murder called the “Paradise of a Thousand Cuts.” They have great harems of drugged females from across the galaxy. They’re controlled by musk glands, with head men and tribal chiefs having certain scents that make others afraid of them. Their females wear burkas.

If it sounds familiar that’s because it’s textbook post-9/11 Islamophobia in space: religiously fanatical Middle Eastern-coded aliens portrayed as vermin and an existential civilizational threat, with donkey features and pheromone-based hierarchies suggesting subhuman biology. There’s no satire. No critique of xenophobia or how fear creates monsters. No individual Zhee characters with complexity or humanity. Just: here are the religious fanatics who want to destroy civilization.

Grok defended them.

The usual moves: we can’t assume authorial intent, they’re just aliens not real people, there might be more complexity than appears on the surface.

Galaxy’s Edge was published by WarGate Books, which markets itself as “unapologetically pro-Liberty,” explicitly “anti-woke,” with the mantra “We write what we want.” The Zhee got published and defended. My trafficking survivor with agency would get auto-rejected.

This is the catastrophic double standard AI gatekeeping creates: Islamophobic space locusts get every charitable interpretation (“can’t assume intent,” “they’re just aliens,” “might be complex”), while a carefully crafted survivor character gets mechanically flagged and rejected.

AI doesn’t protect nuance or creative freedom—it just protects different harmful content while still destroying thoughtful progressive work.

The Bottom Line

An AI sensitivity reader would flag Jasmine as problematic Orientalism. A human reader can see she’s a fully realized character surviving impossible circumstances with humor and agency intact. An AI would flag Kit’s misgendering as transphobic. A human reader understands she’s successfully disguised as a boy—that’s the entire plot. An AI would flag Bennett’s story as gratuitous violence. A human reader recognizes this is about PTSD and witnessing unspeakable trauma. An AI would flag the in-world racist broadsheet as racism. A human reader recognizes it’s satire about propaganda. An AI would want characters in a Renaissance-inspired fantasy to express 21st-century feminist consciousness. A human reader understands that would shatter the worldbuilding.

And when publishers or agents use AI as a gatekeeper, all of these manuscripts get rejected before a human ever reads them. Your satirical critique of racism is flagged as racist and rejected. Your historical fiction with accurate period attitudes is flagged as bigoted and rejected. Your story about a survivor is flagged as exploitation and rejected.

The best AI is pretty good at literary analysis. But even sophisticated AI can’t make the judgment calls that sensitivity reading—or acquisition decisions—require. Those judgments require understanding whether a character has depth and agency, whether the story condemns what it depicts, whether content serves the narrative, whether something is satire or social commentary, whether the setting justifies the content, whether the material is handled with appropriate gravity, and whether publishing this book would make literature better or worse.

I won’t lie, I’m highly skeptical of anyone—human or AI—acting as an ideological gatekeeper deciding which stories are “allowed” to exist. Writers should do their research. They should write with responsibility and complexity. They should give marginalized characters agency and humanity, not just use them as props. And then readers should be trusted to think for themselves.

But AI gatekeeping makes a questionable system infinitely worse. Because it can’t distinguish between thoughtful depiction and exploitation. It can’t recognize satire. It can’t judge execution. It just counts keywords and rejects anything that looks “risky”—which means the books doing the most important work never reach readers at all.

If we let algorithms decide what gets published, we don’t get safer books. We get blander books. We get books that treat readers like children. We get books scrubbed of moral complexity, historical accuracy, satire, and social commentary. And the difficult stories that most need to be told never reach readers at all.​​​​​​​​​​​​​​​​

---

---

Bonus Content: Grok’s Ironic Review

After writing this article, I asked Grok to review it. The results were… somewhat hilarious in a disturbing way.

Grok praised the article’s structure and examples, noted its evidence-based approach, and acknowledged the limitations of AI in literary analysis. Then it got to the Zhee section and proceeded to demonstrate exactly the problem the article describes.

Here’s what Grok said about my critique of the Zhee: “While he labels the Zhee as Islamophobic stereotypes, he doesn’t acknowledge potential counterarguments, such as the authors’ intent to create menacing antagonists within a military sci-fi framework.”

Let me be clear: Grok is defending Islamophobic space locusts while reviewing an article that specifically documents how Grok defended Islamophobic space locusts. It’s giving them charitable interpretation—“just menacing antagonists in military sci-fi”—despite donkey features, “space locust” dehumanization, Grand Pasha theology, crescent knives, ritualistic murder, harems of captive women, and their own females called “mares” in burkas.

Then Grok claimed I “overstates AI’s defense of the Zhee by relying on Grok’s hypothetical response.”

It wasn’t hypothetical. I actually tested Grok. It actually defended them. And now it’s calling its own documented behavior “hypothetical.”

Meanwhile, Grok criticized the article for “overreliance on personal manuscripts” and “tone occasionally slips into hyperbole,” while giving the Zhee: “Well, we can’t know authorial intent.”

The double standard is perfectly preserved. Even when reviewing an article that explicitly documents how AI applies charitable interpretation to harmful stereotypes while mechanically flagging thoughtful work, AI continues to charitably interpret harmful stereotypes.

You couldn’t ask for better proof of the thesis. AI can’t make sophisticated literary judgments, and when confronted with its own failure mode, it simply repeats the failure while critiquing the article that documented it.

The algorithm sees what it’s programmed to see, defends what fits its patterns, and can’t recognize the irony of doing exactly what it’s being criticized for—even when that criticism is sitting right in front of it.