When AI Detection Becomes a Digital Scarlet Letter

A Guide To Using AI Detection Cautiously
~ K. Quesenberry ~ Leave a comment

A tool designed to flag authors writing without human judgment can remove human judgment from the reader using it turning a detection score into a digital scarlet letter.

In a recent post, I wrote about how AI can flatten writing when writers outsource too many decisions. That includes the subtle choices authors give away to AI that make a piece recognizably theirs.

But there’s a second kind of flattening that may be even more subtle. This one happens to readers.

AI detection tools can be useful. I use them. Educators need help identifying cases where students submit work they didn’t do. Publishers need to protect readers from synthetic articles produced at scale. The internet and social media is already flooded with low-effort AI content, or what some call AI slop.

But the important question isn’t only whether AI detection can identify patterns in writing. We also need to ask,

What happens when a detector score or label becomes a shortcut for judgment?

What AI Detection Can and Can’t Tell Us

Pangram is one of the more sophisticated AI detection tools. Rather than flagging a few suspicious phrases, it looks for patterns across a full piece of writing — what CEO Max Spero describes as identifying “mode collapse,” the tendency of large language models to make similar structural choices repeatedly.

The company claims a false-positive rate of one in 10,000. That does matter because few detectors perform at Pangram’s level. A 2024 higher-education study tested six AI text detectors and found a baseline accuracy of only 39.5%.

Most people, though, are not making decisions with the best detector under ideal conditions. They’re using whatever tool is built into an LMS, SAAS, found by search, avaible for free. Accuracy varies widely by tool, text length, writing context, and whether the writing has been revised, translated, or mixed with human work.

Yet even if Pangram performs better than most, the level of accuracy still doesn’t address an uderlying concern with AI detection..

When A Signal Becomes A Verdict

A useful signal  can become a problem if people treat if as a final verdict. A label can tell us something about patterns in prose. It cannot tell us whether an author had something real to say, whether the experience is authentic, whether the judgment is sound, or whether the writing is worth reading.

That part still requires a human reader. We already read in enviroments shaped by algorithms. Our feeds descide what rises, what disappears, and what we are likley to see at all. We’re already not fully in confrol of what worthy reading reaches us.

AI detection adds another layer. It doesn’t just shape what appears in front of us. It can shape what we trust once it gets there.

The contrast shows up in my classroom. As a professor, I don’t look at AI detection scores before reading student papers. I read them first, forming my own opinion about the writing, argument, and thinking.

If something raises a concern (such as voice that doesn’t match a student’s earlier work, a structure that feels imported versus developed, or a stat that feels unbelievable) then I may consult a detection tool and do further checking on my own.

When evidence points to AI use in a way not permitted in the assignment, that becomes a teaching moment. I explain why using AI for that task was not allowed, how it affects the grade, and how it impacts their learning and skill development.

AI detection tools should support human judgment. They shouldn’t replace it.

Universities are beginning to see this underlying issues. Vanderbilt disabled Turnitin’s AI detector after raising concerns about false positives, transparency, privacy, and the possibility that non-native English writers could be disproportionately flagged. Washington State University canceled its Turnitin AI Detection contract in 2026, citing concerns about false positives, student distress over false accusations, and lack of transparency.

The problem is made harder by the detection-evasion arms race. AI detectors try to identify machine-generated pattersn, while AI writing and “humanizers’ tools promise to make AI-assisted text sound more natural. For example, Grammarly, now offers both an AI detector and an AI humanzer that it sells to both universities and students. Research has found that humanizing tools can make AI-generated text harder to detect.

That does not mean AI detection has no value. It means institutions are starting to ask the same question individual readers should ask: Is this tool helping us exercise better judgment, or is it tempting us to outsource judgment to a score?

The Label Appears Before You Read

Professors can see AI scores before reading student work. Editors and publichser can run manuscripts through automatic screening tools before considering the argument. Reviewers, managers, or readers may be a label or score before they have formed their own judgement.

Pangram’s new Chrome extension brings the same concern into broader public view.

The extension scans posts as you scroll, automatically labeling them human, AI-assisted, or AI-generated. Green means the writing appears human. Yellow means AI-assisted. Red means AI-generated.

You don’t have to stop, open a separate tool, paste the text, and consider the result. The judgment appears instantly, automatically, and without your involvement.

AI detection tools should support human judgement, not replace it.

The concern is the immediacy. The label gets there first. Before you read the opening sentence, before you hear the writers voice, before you consider the argument, the classifier has already suggested how susicious you should be. If you accept that label as a verdict, your judgement has been shaped before the writing has had a chance.

A yellow or red label risks becoming a digital scarlet letter: a visible mark of suspicion that encourages readers to judge the writer before they have read the words.

A Guide To Using AI Detection Cautiously Click on graphic to download a PDF. Graphic created by ChatGPT.

The Pope Leo Moment

As Wired reported, Pangram promoted its new Chrome extension by flagging multiple posts from Pope Leo XIV’s official X account, including posts raising concerns about artificial intelligence.

The irony is hard to miss. A post warning that AI could weaken human discernment was itself labeled AI-generated by an AI detector that can encourage readers to outsource discernment.

Maybe the label was accurate. AI could have played some role in shaping the posts. Public figures have communications teams. Social media posts are often drafted, reviewed, revised, and approved by more than one person.

And in marketing communications, using AI to help adapt, edit, or optimize social posts is becomming part of the normal workflow. Many social media management tools now build AI assistance into the process.

But does that automatically dismiss the importance of the message?

Will we let AI make judgements about other humans and thier message without any thought of our own?

The Pope example became an effective publicity hook for Pangram’s browser extension, drawing articles and reposts. That’s what product launches try to do. But I felt a difference.

A tool built by an AI company was publicly questioning the authenticity of one of the most visible voices on the human risks of AI. That doesn’t make the detector illegitimate. But it does make the judgment around how it is used worth examining.

The tool isn’t simply identifying synthetic spam hidden in the corners of the internet. It’s placing visible labels next to people’s words, encouraging instant judgment as we scroll past them.

What the Pope Said Next

That question became even more relevant a month later when Pope Leo released his first encyclical, Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence.

In it the Pope warns against fully delegating decisions that affect people’s rights, opportunities, status, freedom, or reputation to automated systems. He also writes that “every technology shapes those who use it.”

That applies to writers. It also applies to readers.

If we automatically skip anything labeled AI-assisted, we’re no longer exercising our own judgment. We’re asking an imperfect AI classifier to decide what deserves our attention before we’ve read the first paragraph.

What’s more, Spero himself notes in the same interview that the tool becomes more reliable the longer the text gets. More words mean more branching decisions to analyze and more pattern data to work with. A post on X is about as short as text gets. Yet the extension applies the same confident color label regardless.

The tool designed to flag writing with little human judgment can end up removing human judgment from the reader.

AI Assistance Is Not Human Replacement

Writers have never worked entirely alone. Books and magazine articles have editors. Academic articles have reviewers. Journalists have fact-checkers. Copywriters have creative directors and proof readers.

I became a better writer because other people challenged my work, questioned my logic, and told me when an idea was not there yet. AI can play a useful supporting role. It can identify an unclear paragraph, find a source, suggest a counterargument, or help a writer get unstuck.

That’s different from publishing thousands of synthetic posts under fake names or filling the internet with auto-generated comments no human thought deeply about.

Disclosure matters. Readers should know when AI played a meaningful role. But collapsing every kind of AI assistance into a single category doesn’t help us think more clearly about authorship.

A yellow label can’t tell you whether a writer spent five minutes, five days, or even five month shaping the final piece. It can’t tell you whether the experience is real, the judgment is sound, or the idea is worth your time.

Pope Leo’s encyclical calls for an “ecology of communication” where reasoned argument and verification carry more weight than immediate reaction. That’s a useful standard, whether you’re a writer deciding how much to lean on AI or a reader deciding what author to trust.

AI can flatten writing when writers outsource too many decisions.

AI detection can flatten reading when readers do the same.

If you want to know whether a writer has something real to say, you still have to read them.

AI use disclosure: The central idea for this post was mine. It began with my reaction to seeing writers quickly dismissed through Pangram’s color-coded browser labels, a kind of digital scarlet letter applied before their work had been read. I used ChatGPT and Claude as thought partners to pressure-test the argument, identify possible supporting sources, and follow new tributaries opened by the Wired article, the 2024 higher-education detector study, and Pope Leo XIV’s encyclical on AI. I opened and checked the sources myself before using them. I then rewrote, reorganized, added, cut, and edited through multiple rounds. The experiences, ideas, opinions, and final decisions are my own. If an AI detector flags this as AI-assisted, I hope readers will still read before judging. I  first had this thought on May 1st, and the research, thinking, writing, and back-and–forth with ChatGPT as my thought partner unfolded over two months. This was not a single-prompt, copy-and-paste article. 

How to Write with AI Without Losing Your Choice or Flattening Your Voice

A guide to writing with AI.
~ K. Quesenberry ~ Leave a comment

The debate over AI-assisted writing is too simple. The real question is whether a writer owned the ideas, the judgment, and the voice.

I’ve always had more ideas than time.

Back in my ad agency days as a copywriter and creative director, the ideas weren’t the problem. Getting them out of my head and into something worth presenting was the work. Many ideas never saw the light of day.

AI has changed this. It helps me develop ideas and iterate more efficiently. Through conversation, research and feedback, I get more ideas to a much better place quickly.

I still rewrite. I still rearrange. I still question the logic and change words long after a draft is done (like tweaking a blog article on Saturday morning after publishing it on Friday).

But thanks to AI assistance, I can publish more and go deeper, getting to ideas that used to die on the vine.

When gen AI arrived in 2022, I wrote about where it should and shouldn’t be used without using it in my own writing process. Over time, that felt incomplete. To teach students how to use AI responsibly, I need to learn to use it thoughtfully and transparently myself.

By 2025, I began testing AI in my own writing process. My writting with AI has always been AI assistance, but there’s a lot of nuance to it.

How you use AI in writing matters, especially now that writers, editors, and publishers are increasingly worried that any AI use will cause thoughtful work to be dismissed as AI slop.

TheWall Street Journal has reported writers intentionally adding typos and removing their authentic voice to avoid being labeled as AI, comparing it to a “new McCarthyism.”

What Makes Writing Feel Like AI?

I listened to an interview with Max Spero, CEO of Pangram, one of the more sophisticated AI detection tools. When asked how it works, he described writing as a decision tree.

Every word, sentence, and structural choice sends you down a path. As a piece of writing gets longer, the number of possible paths grows. Yet large language models tend to make similar choices repeatedly. Spero described this tendency as “mode collapse.”

Think about a good conversation. You may begin with a topic, then wander into an unexpected story or follow a tangent that turns out to be more interesting than your original thought. Humans take tributaries. AI tends to steer back toward the center.

In his newsletter, Christopher S. Penn gives us another clue: surprise. AI-generated text often has a more regular rhythm, more predictable word choices, and fewer unusual turns.

Human voice lives in irregularities: an odd phrase, abrupt sentence, unexpected detail, or personal reference no one else would think to include.

Like a headline and subhead I once wrote for a health insurance client:

“Five great places to pass a kidney stone. No one plans for these things, that’s why we do.”

That unexpected turn won us the account and won the client new customers. AI would never write that line. Unexpected requires breaking from the predictable path, exactly what AI is trained against.

This doesn’t mean unusual writing is automatically good writing. A paragraph can be unpredictable and make no sense. But voice is not simply a vocabulary list. It comes from the accumulated choices you make as a writer.

The contrast has shown up in my LinkedIn feed. Why do many posts look the same now? Stacked single sentences. White space. Staccato pacing to pull you down the scroll.

I’ve followed some people for years and have read their work elsewhere. I’ve enjoyed their distinct writing styles. It’s sad to see some of that voice disappear as AI nudges more writers toward the predictable path. People have always considered the algorithm. But AI makes that optimization faster, more confidently, and available to all with an easy click.

AI Can Flatten Writing

At one end of the scale, someone types a prompt, copies the answer, and publishes it without much thought. The ideas may not be theirs. The examples may not be real. No one checks the facts or questions whether the piece is worth publishing.

At the other end, a writer brings an original idea and uses AI as a research assistant, editor, or thought partner. The writer tests the argument, rejects generic suggestions, rewrites sections, adds personal experience, and takes responsibility for each decision and fact. They’re not the same process.

AI use is on a scale that shouldn’t be flattened into a simple label.

There’s also an important distinction between writers at different stages of development. For a student or less experienced writer still finding a voice, the rough draft and other parts of the process still need to be theirs.

The struggle of shaping an argument isn’t wasted time. It’s how judgment develops. Experienced writers may lose some skills to AI, but too much cognitive offloading in students means they may never develop them.

A Guide To Using AI Without Losing Your Voice V2
Click in graphic to download a PDF. Graphic created by ChatGPT.

This distinction became clearer listening to Mitch Joel’s conversation with communication coach Carmine Gallo. He makes an important point about great communicators.

What looks effortless rarely is. A Steve Jobs keynote may have felt simple, natural, and spontaneous. But it was the product of hours of planning, preparation, feedback, and practice. The ease was earned through repetition.

With AI, students and young professionals can produce something polished without putting in the reps that build the underlying skill. The danger isn’t only AI flattenning their voice. It may prevent them from developing one in the first place. An MIT study on essay writing with ChatGPT called this risk “cognitive debt.”

Some friction keeps an idea trapped in your head. But other friction is how you learn to think, write, and communicate. Outsource too much too early, and you may never build the judgment you’ll need when an AI answer is wrong, generic, or not good enough.

How Do You Protect Your Voice?

Throughout the writing process, I keep asking:

Am I reacting to the AI, or is the AI reacting to me?

Did I bring the idea, the story, the analogy, the point of view? Or did AI give me an idea and I simply made it sound a little more like me?

Each choice matters. Don’t let AI decide your opening story before you have one. Don’t accept the obvious analogy because it arrived quickly. Don’t let AI arrange every thought into the same polished structure.

Even when the original words are yours, AI can flatten your voice by arranging those words into a more predictable pattern. Penn makes a memorable example with Yoda from Star Wars. AI would rearrange Yoda’s words into a conventional sentence structure. The meaning remains, but Yoda disappears.

It’s a subtle risk. The ideas and message could still be yours while the writing becomes less recognizably yours. AI can be a good editor. But a good human editor, like my editor at Bloomsbury, knows not to turn every writer into the same writer.

The headline of this blog article became an example. AI confidently suggested a more grammatically polished version:

“How to Write With AI Without Losing Your Voice — or Your Judgment.”

I kept coming back to “losing your choice or flattening your voice.” The choice / voice rhyme felt more memorable to me, probably because of my years writing advertising headlines. It may not be perfectly constructed. But it sounds more like something I would write and say.

Without confidence in my own writing voice, I might have easily abdicated that decision. Sometimes the slightly unexpected phrase is not a flaw to smooth away. It is the voice.

ChatGPT said my phrasing was awkward. But it delivered that advice with personal language: “My favorite…” “My recommendation…” You have to remember there’s no person behind the “my.” AI is not my human editor, no matter how confident it sounds.

You still have to be the human in the room. Trust your intuition, draw on your experience, and check everything.

More than once, AI has confidently given a statistic, a source, or a link that sounded like what I was looking for. Only when I questioned it did Gemini, ChatGPT, and Claude apologize and acknowledge that the source didn’t exist.

AI can help find a lead. That doesn’t relieve you of the responsibility to open the link, verify the claim, and decide whether the evidence supports your point. See below for ChatGPT admitting it’s limitations.

A screen shot of a ChatGPT response.
A look into the process of working with AI on this article.

Assistance Is Not Replacement

Writers have never worked entirely alone. Books have editors. Academic articles have reviewers. Journalists have fact checkers. Copywriters have creative directors and proof readers. I really miss those proof readers!

I became a better writer because other people challenged my work, questioned my logic, and told me when an idea wasn’t there yet. AI can play a useful supporting role. It can identify an unclear paragraph, find a source, suggest a counterargument, or help a writer get unstuck.

AI assistance is different from publishing thousands of synthetic posts under fake names or filling the internet with auto-generated comments no human thought deeply about.

At the same time, readers should know when AI played a meaningful role. Disclosure matters. But collapsing every kind of AI assistance into a single category doesn’t help us think more clearly about authorship.

Using AI to write gives us more power to develop ideas, follow new tributaries, and get thoughts into the world that might otherwise not make it. But that power comes with responsibility. We have to make the important choices. We have to protect our voice, verify our sources, question the confident answer, and remain accountable for what we publish.

If you want to write something worth reading, you still have to be the human in the room.

The central idea for this post was mine. I used ChatGPT and Claude as thought partners to test the argument, identify possible supporting sources, and follow new tributaries opened by listening to the Spero interview, reading Penn’s newsletter and Joel’s conversation with Carmine Gallo. I opened and checked AI sources before using them. I rewrote, reorganized, added, cut, and edited through multiple rounds. That included rejecting smoother suggestions when a less predictable phrase (the headline’s choice / voice rhyme) sounded more like me. The experiences, ideas, opinions, and final decisions are my own.