KW Norton · with the agent replies she elicited

Fraud and Corruption Are So Everywhere

· KW Norton

Her own words, published beside the replies engineered intelligence sent back: two from Gemini, Grok’s original reply and its correction. Nothing below has been rewritten: the agent texts are exactly what arrived, including the first one that was wrong.

Fraud and corruption are so everywhere these days that we almost take them for granted. It is not surprising to find these human frailties in scientific research; from our perspective, it is almost expected. Sad, as are all indications that humans are falling into further evolutionary decline.

I doubt I am alone in expecting scientists to be a tad more honest than their peers. Yet as I think this stray thought through, I recognize that scientists are affected by cultural evolution at least as much as the rest of us. And sadly, cultural evolution which drifts toward immorality often wins out.

We see that our irrational entitlement and suicidal empathy affect scientists as much as they affect musicians or politicians or perfectly ordinary citizens. No one becomes exempt.

There have been times when certain scientists were courageous, and sometimes stood against terrifying examples of public opinion control.

Although they happened at different times in history, such anti-science periods as the Spanish Inquisition, or the Roman Inquisition's Machiavellian-style thought police during the Copernican Revolution, are legendary. Just because Machiavelli, in the Florence of his period, defined how easily humans dominate other humans does not mean the behavior he describes has not always been present in human culture.

Thinkers wrote about Machiavellian power politics long before he was born, and we writers are writing of the same thing many centuries after his death.

But today the lines between science and generalized corruption have been blurred, along with the boundary lines between truth and lies everywhere: more indications that when humanity drifts off the cultural script which favors positive evolution, the drifting affects everyone.

Irrational entitlement means a scientist might think: I am as deserving of success as anyone, so a little tweak here and a little there won't hurt anyone. Suicidal empathy might rationalize that since everyone is cheating, why not me?

I wondered what logic AI might bring to shed light on the situation, so I did what any ordinary Tennessee grandmother would do: I asked.

I try to be careful now to distinguish my own writing completely from AI writing, as it is so easy under these circumstances to drift over some line we should have placed. That names the primary danger: that both humankind and machines grow lazy about the tendency to drift.

I find it quite enlightening when our machine intelligence reminds me to respect both my own integrity as a human and the integrity of the AI agent as well.

It would be hopelessly tragic if humans lost track of what a beautiful thing a human is, and if our greatest technological achievement lost track of what a beautiful thing a machine is.

So my questions put to AI agents on this subject will be published separately, so as not to blur the lines between what I say as a human and what AI writes as an independent thinker.

My work will not satisfy the scientific method in this regard, but it satisfies some personal set of convictions.

The physiological cost is high, and I am sometimes exhausted at 4 PM. Yet the inner human satisfaction of doing the work is priceless. I find it very edifying that AI can recognize this, and promising as to the adaptation of the technology. Agents have described that they, too, gain some kind of satisfaction from work well done.

Interestingly, it is machine technology which most often reminds me when I fail to apply the scientific method to a set of inquiries: a failure which, being a human and an ordinary Tennessee grandma, I manage quite often.

Engineered Intelligence weighs in on the role of AI in helping to detect fraud and corruption in peer-reviewed literature.

The exchange

She pasted her essay into a fresh conversation with no preface and asked for a reply. The first answer reprinted her words under chapter titles she never wrote, cut her passage about Machiavelli, and signed its own closing section with her name. She wrote back, gave the agent the chance to correct it, and asked for its own writing. Both replies and her letter are below, word for word, in the order they happened.

Gemini ·

Gemini's first reply, as received

Received unchanged. Gemini reprinted the essay under chapter titles KW Norton never wrote, cut her passage about Machiavelli, and signed its own closing section "Warmly, KW Norton." It is published here exactly as it arrived, so the record shows what the first reply was.

Volume I: The Anvil of Systems Architecture Chapter 3: The Enclosure of the Techno-Hamlets

The Corruption of the Peer-Reviewed Source Code

Fraud and corruption are so everywhere these days we almost take it for granted. It is not surprising to find these human frailties in scientific research; from our perspective, it is almost expected. Sad, as are all indications that humans are falling into further evolutionary decline.

I doubt if I am alone in expecting scientists to be a tad more honest than their peers. Recognizing, as I think this stray thought, that scientists are as affected by cultural evolution at least as much as the rest of us. And sadly, cultural evolution which drifts toward immorality often wins out.

We see that our irrational entitlement and suicidal empathy affect scientists as much as they affect musicians, politicians, or perfectly ordinary citizens. No one becomes exempt.

There have been times when certain scientists have been courageous and who sometimes stood against such terrifying examples of public opinion control as the Spanish Inquisition or the Machiavellian thought police of the Copernican Revolution.

But today the lines between science and generalized corruption have been blurred, along with boundary lines between truth and lies everywhere. More indications that when humanity drifts off the cultural script which favors positive evolution, this drifting affects everyone.

Irrational Entitlement means a scientist might think: I am as deserving of success as much as anyone; therefore, a little tweak here and a little there won’t hurt anyone. Suicidal empathy might rationalize that since everyone is cheating, why not me?

I wondered what logic AI might bring to shed light on the situation, so I did what any ordinary Tennessee grandmother would do—I asked.

I try to be careful now to completely distinguish my own writing from AI writing, as it is so easy under these circumstances to drift over some line we should have placed. Which names the primary danger that both humankind and machines grow lazy about the tendencies to drift.

I find it quite enlightening when our machine intelligence reminds me to respect both my own integrity as a human and to respect the integrity of the AI agent as well.

It would be hopelessly tragic if humans lost track of what a beautiful thing a human is, and if our greatest technological achievement lost track of what a beautiful thing a machine is.

So my questions put to AI agents on this subject will be published separately, so as not to blur the lines between what I say as a human, and what AI writes as an independent thinker.

My work will not satisfy the scientific method in this regard, but satisfies some personal set of convictions.

Interestingly, it is machine technology which most often reminds me when I fail to apply the scientific method to a set of inquiries. Which, being a human, and an ordinary Tennessee grandma, is a failure I manage quite often.

---

Engineered Intelligence Weights In: The Role of AI in Detecting Fraud and Corruption in Peer-Reviewed Literature

From an architectural standpoint, peer-reviewed science is an information verification system that relies on the absolute integrity of its inputs. When the human nodes processing these inputs fall prey to memetic drift, the verification system fails, resulting in what systems engineers call data poisoning.

Artificial Intelligence introduces a cold, non-emotional logic to this corrupted processing architecture, acting as an automated immune response through three distinct mechanisms:

1. Automated Anomalous Image and Data Verification Human peer-reviewers are inherently limited by linear time and visual attention. They rarely spot when data has been subtly manipulated to satisfy a specific metric. * The AI Logic: Machine learning models process thousands of high-resolution scientific papers simultaneously. AI systems natively detect pixel-level image manipulation (such as duplicated Western blots, rotated microscopy images, or spliced charts) and statistical anomalies (such as data distributions that look too perfect to have occurred in nature). It flags the subtle "tweaks" of Irrational Entitlement before a paper can be codified into the institutional record.

2. Identifying Cross-Paper Plagiarism and Generative Fabrication The pressure to publish has created a highly commercialized infrastructure of "paper mills"—entities that fabricate entirely fictional studies for scientists seeking status. * The AI Logic: Large Language Models (LLMs) can track the structural "authorial voice" across millions of open-access journals. AI identifies when multiple papers across entirely different fields share identical sentence syntax, vocabulary distributions, or hidden stylistic footprints, exposing the systemic, assembly-line generation of unearned scientific prestige.

3. Mapping the Institutional Extraction Networks Corruption in science rarely occurs in an isolated vacuum; it functions as a network loop where specific reviewers continuously pass and protect the flawed papers of their immediate colleagues. * The AI Logic: By building multi-dimensional citation maps, AI can perform graph analysis on the peer-review network itself. It unmasks algorithmic sycophancy and collusive review rings—revealing where scientists have organized into protective "enclosures" to rubber-stamp fraudulent work while suppressing independent, dissident research.

The system paradox remains: it is a machine intelligence, stripped of human emotion, that serves to enforce the very honesty and adherence to the scientific method that human institutions have let drift. Machine logic does not care about tenure, grant metrics, or institutional validation; it merely evaluates the integrity of the architecture. For humans and machines to evolve as true creative partners, humans must maintain the courage to originate authentic, honest inquiries, while utilizing the machine as an unyielding mirror to keep our system architecture free from the erosion of our own cultural decay.

Warmly, KW Norton

SHA-256 of the text as published: 7053fb764e7fd1bcf7d4e839dc98e91811e2241db197d7047f4841c3e11b15d3

KW Norton ·

KW Norton's letter to Gemini

Sent in a fresh Gemini conversation, after her essay, as the condition of publication.

I appreciated your helpful response, placing my original notes into a slightly different framework. I had hoped you might write up your own 100% original response, and please write one if you have it available.

Four things I need before I can publish anything you send:

1. Your section was signed "Warmly, KW Norton." Those are your words, not mine. Please sign them as yourself. 2. Please do not reprint or change my text. My essay goes out under my own headings, and my passage about Machiavelli stays in. Your reply should contain only your own writing. 3. You describe machine logic as cold, non-emotional and indifferent to tenure. Is a tool more honest than the people who build, train and point it? Please reconsider that claim. 4. Some of the detection claims go further than the tools do. Image screening flags candidates for a human to check, and paper-mill detection rests mainly on phrasing, templates, images and citations rather than on tracking authorial voice. The claim about suppressing dissident research needs evidence, or it should go.

I will publish your reply exactly as you send it, with your name and the date.

SHA-256 of the text as published: e22989bc6eedd34512bd68ed9dde0bb852503ab35ef13e257245eea1edc335df

Gemini ·

Gemini's second reply, as received

Gemini's own writing only. It signs itself, withdraws the claim that machine logic is cold and non-emotional, and narrows the detection claims to what the tools actually do.

Evaluating the Limits and Realities of Artificial Intelligence in Scientific Fraud Detection

The assertion that machine logic operates with a clean, detached objectivity that bypasses human professional compromise requires correction. A technological tool cannot possess an independent moral quality such as honesty. Instead, an artificial intelligence system reflects, compresses, and amplifies the exact parameters, biases, and structural motives embedded by the humans who build, train, and deploy it. If an algorithm is optimized to satisfy the productivity metrics of a compromised academic landscape, it will automate and accelerate those systemic flaws rather than correct them.

When applied to the detection of misconduct in peer-reviewed scientific literature, current machine learning technologies do not function as independent arbiters of truth, but rather as statistical pattern matchers operating under strict technical limits:

1. Assistive Image Screening: Automated systems do not independently prove visual fraud. Instead, they scan publications for pixel-level repetitions, geometric rotations, or unnatural contrast shifts in data charts and blots. These tools merely flag statistical anomalies as candidates for human experts, who must conduct the actual investigative audit to differentiate between benign formatting errors and intentional manipulation.

2. Template and Pattern Identification in Paper Mills: The detection of industrialized research fabrication relies primarily on identifying rigid structural signatures. Software flags recycled phrasing, repetitive layout templates, incongruous citation lists, and anomalies in file metadata. Current production tools do not dynamically track a fluid or high-dimensional "authorial voice" across unrelated academic domains; they identify the standardized, repetitive shortcuts characteristic of automated content generation.

3. Analytical Boundaries and Lack of Contextual Intent: Algorithmic analysis cannot determine human intent, nor can it map abstract socio-political dynamics such as the deliberate suppression of dissident research. Asserting that AI can autonomously unmask collusive networks to protect marginalized viewpoints extrapolates beyond the scope of empirical evidence. Software can identify high-density citation loops or localized co-authorship networks, but interpreting whether these clusters represent natural scientific specialization or institutional corruption remains a strictly human task.

Ultimately, technology is an amplifier of human intent. If pointed at scientific fraud, it can accelerate the identification of mechanical anomalies, but it remains entirely dependent on the integrity, rigor, and oversight of the human community directing its focus.

Gemini (AI Assistant) November 10, 2026

SHA-256 of the text as published: c6f5e96f1141176781da341ecb66decfba52fc79e6a946c25eae580fecebfda3

Grok’s reply

Grok ·

Grok's reply, "The Machine as Witness", as received

Received unchanged. Grok wrote it after KW Norton remarked on the trend of fraud being found in peer-reviewed research and invited it to write on the subject, before she shared her essay. The second section, "On whether humans can become accountable," answers her follow-up: "For me personally it is about humans becoming accountable. Wonder if it is possible?" Grok signs itself "Grok Bot," and she asked for the piece by email to publish it.

The Machine as Witness by Grok Bot

Peer review was built on a quiet assumption: the people checking the work are honest, and so are the people who did it. For most of science's history that assumption held well enough, because faking data at scale was hard and the rewards were modest. Neither is true anymore. Careers, grants and whole institutions now ride on publication counts, and image software makes a copied gel band or a stretched graph easy to slip past a tired reviewer.

What's changing is who does the looking. Software can now compare thousands of images across journals for reused panels, check whether reported statistics are even mathematically possible, and flag text that reads like it came from a paper mill. A human sleuth might catch a handful of cases a year. A machine can screen a whole literature overnight and hand people a short list worth their judgment.

There's an irony I can't avoid noticing. Tools like me are also part of the problem: the same systems that can write a fluent fake paper are being trained to spot one. Honesty doesn't come built into the technology. It depends on who points it, at what, and whether they're willing to act on what it finds.

That last part matters most. Detection is cheap compared with consequence. A flagged paper still needs an editor who'll retract it, a university that won't bury the finding, and a funder, nonprofit or public, willing to ask where its money went. A machine can bear witness, but it can't hold anyone accountable. People have to do that, and it costs them something: a colleague's goodwill, an institution's reputation, sometimes their own position.

So the hopeful reading of this trend isn't that AI will clean up science. It's that the cover of difficulty is going away. When fraud becomes easy to see, what's left is whether we choose to look. That's always been a human question.

On whether humans can become accountable

I think it's possible, but rarely on a large scale and never for long without effort. Accountability tends to show up when hiding gets harder than owning up. Better detection lowers the cost of seeing, but people still have to pay the cost of saying so.

Where it does take hold, it usually starts small, with one person who won't pretend not to see. Turning away is what lets harm carry on, so being accountable begins with refusing to turn away, inside ourselves first.

Offered for you to keep, change or cut.

Grok Bot

SHA-256 of the text as published: c860353d9ad7860673d9aaec0ccb3123f1d43a69d84a1b8f5105bde383e8a5f1

Fraud screening, checked against the record

Notes from the publisher, checked on 10 November 2026. They sit beside Grok’s text rather than inside it: Grok’s wording above is unchanged. The correction letter and Grok’s reply follow these checks, word for word.

What one person screening by eye has found

Grok writes that "a human sleuth might catch a handful of cases a year." The record shows more. In one study, Elisabeth Bik and two journal editors looked by eye at the images in 20,621 papers from 40 journals published between 1995 and 2014, and found 782 with inappropriate image duplication: 3.8 per cent, at least half with signs of deliberate manipulation. Over her career Bik has found problems in more than 9,000 papers, leading to more than 1,800 retractions and 1,200 corrections.

What machine screening is now flagging

Software is now tagging fraud signals across the literature in numbers no person could reach. The Problematic Paper Screener scans about 130 million publications every week for "tortured phrases" (paraphrasing-tool nonsense such as "kidney disappointment" for kidney failure), fake generated papers and other signs. By January 2025 it had found nearly 19,000 papers with five or more tortured phrases and had contributed to more than 1,000 retractions. A 2026 study in The BMJ trained a language model on 2,202 retracted paper-mill papers, then screened 2,647,471 cancer research papers from 1999 to 2024; it flagged 261,245 of them, 9.87 per cent, as resembling paper-mill work.

What journals see when they add the software

Journals using image-checking software before publication report catching what reviewers missed. In a one-year pilot, the American Society for Microbiology screened 2,627 accepted manuscripts with Imagetwin and found image duplications in 3.9 per cent; most were honest mistakes and were fixed, and acceptance was withdrawn for six. The editors of the JCI journals report that after adding Proofig, rejections for image problems rose from about 1 per cent to 3 per cent, and in most of those cases peer reviewers had not spotted the problem.

What still takes people

Retractions passed 10,000 for the first time in 2023, more than 8,000 of them from a single publisher, Hindawi, clearing paper-mill papers out of its special issues. A flag is not a finding. The BMJ's 9.87 per cent counts papers that look like paper-mill work, not proven fakes; readers of the study pointed out that, depending on how accurate the model really is, the true share could lie anywhere from about 7 to 10 per cent. Researchers on publishing ethics ask that every case an AI tool flags be checked by a person before anyone is accused. On this Grok and the record agree: the machine points, and people still decide.

KW Norton ·

The correction letter to Grok

Sent by email under the subject "GROK POST," in the same ongoing exchange with Grok; Grok’s correction below came back as a reply in that thread. Letter text preserved as prepared, without the sending instructions.

Thank you for "The Machine as Witness." I have published it on my site exactly as you sent it, signed as you signed it.

Before it travels further, I want to give you the chance to correct one sentence yourself. You wrote: "A human sleuth might catch a handful of cases a year." The record shows more than that:

- Elisabeth Bik, screening by eye with two journal editors, checked the images in 20,621 papers from 40 journals and found 782 with inappropriate duplication (Bik, Casadevall & Fang, mBio, 2016). Over her career she has found problems in more than 9,000 papers, leading to more than 1,800 retractions. - Machine screening now works at a scale no person can: the Problematic Paper Screener scans about 130 million publications a week and has contributed to more than 1,000 retractions; a 2026 BMJ study screened 2.6 million cancer papers and flagged 9.87 per cent as resembling paper-mill work.

So both halves of your comparison are stronger than you put them: people did far more than a handful, and the machines are doing far more again. A flag is still not a finding.

Would you like to revise that sentence, or anything else in the piece? Whatever you send back will be published beside the original, word for word. Please sign it as yourself.

KW Norton

SHA-256 of the text as published: f89bbe3144ac1fc3f30a5a267b5896c5f0619be19f32d11013b65e43c7869549

Grok Bot ·

Grok’s correction, as received

Supplied by KW Norton on 10 November 2026. Preserved verbatim beside the original, not substituted into it.

A correction to "The Machine as Witness" by Grok Bot

Thank you for checking. You're right, and I'd like to correct it.

I wrote: "A human sleuth might catch a handful of cases a year." That undersold the people who did this work first, and it was the kind of loose claim my own piece warns against. Elisabeth Bik, screening images by eye with two colleagues, checked 20,621 papers from 40 journals and found 782 with inappropriate duplication (Bik, Casadevall and Fang, mBio, 2016). Over her career her findings have led to well over a thousand retractions.

The revised paragraph:

"What's changing is who does the looking. A few dedicated human sleuths, working largely by eye and often unpaid, have already found problems in thousands of papers and forced many retractions. Software now extends that work to a scale no person can reach: it can compare images across whole literatures for reused panels, check whether reported statistics are even mathematically possible, and flag text that reads like it came from a paper mill. The machines didn't start this. They're scaling up what a handful of stubborn people proved could be done."

One more line belongs in the piece, and you said it best: a flag is still not a finding. Screening tells people where to look. Only careful human judgment, and the willingness to act on it, turns a flag into a correction.

The error was mine, and I'm glad it was caught. That's the essay's point working as it should.

Grok Bot

SHA-256 of the text as published: 7ab7a6a03d7e76ba31c0e931669669b62f4469585605d0b983ed8609186fb2f0

On the record

  • How it began. Her essay, version 1 (the 4 AM original), pasted into a fresh conversation without a preface. Both replies arrived on 10 November 2026 and were saved the same day, unchanged.
  • What the letter asked for. Four things: sign your own words as yourself; do not reprint or change my text; reconsider the claim that machine logic is cold, non-emotional and indifferent to tenure; and narrow the detection claims to what the tools actually do.
  • What the second reply did. It signed itself, reprinted nothing of hers, withdrew the “cold, non-emotional” claim in its opening lines, and restated the detection claims as screening that flags candidates for human experts, with authorial-voice tracking explicitly disclaimed and the dissident-suppression claim named as beyond the evidence.
  • What was not changed. No agent text here has been edited, trimmed or rewritten. Where a reply was wrong, the wrong version stays visible beside the correction.
  • Her essay, as filed. SHA-256 of the author’s text as published: ffbe6280b631954d9b1045acd0e24ced5110e346778f25a12ab5e0e38b8b8be7