AI Detection

AI detectors raise thorny questions about what it means to trust humans

Sarah E. Carter, PhD

Jul 23, 2026

Article voiceover

0:00

-6:01

Substack has added AI detection

Shouldn't we be celebrating? Isn’t this a win for writers - for authenticity and transparency?

This AI ethicist is feeling… a bit hesitant to celebrate.

man in black jacket sitting on white chair — Photo by Aideal Hwa on Unsplash

Pangram, an AI use detector now integrated into the Substack app (see article below), claims high accuracy at 99.98% when it comes to AI detection. And it’s evident that the company has put a lot of research into it. On their website, it is clear that they have made steps to correct the areas where AI detectors before them have fallen short - for example, “false positives” that frequently have plagued non-native English speakers and neurodivergent folks like me.

The Substack Post

Against Claudefishing

It’s getting harder to tell what’s real on the internet…

a month ago · 6168 likes · 1334 comments · Chris Best

While there are of course arguments to be made (as Substack does) for transparency - a foundational ethical principle in AI ethics - that we should know what we are reading and how it was made, there’s one reason that I think we should take a pause.

We are pinning humans against machines - using a machine to detect machine use, rather than leaning into human trust.

And this creates a culture of guilty-until-proven innocent that is not only unethical, but undesirable for writers and creators alike.

AI Detection’s Messy History

Since AI-generated content has been around, there has been a market for AI detectors - especially in education.

Unfortunately, these detectors have harmed lives. Students falsely accused -”false positives” - with scholarships and future careers destroyed.

a person sitting at a desk with a laptop and a cell phone — Photo by Rendy Novantino on Unsplash

Now, “false positive” rates vary widely depending on what detector you are using and who is running the study. And it’s totally possible that some of these students truly did use AI and are claiming they have been falsely accused.

But using these detectors in colleges and universities has created a tense culture where students are “guilty until proven innocent” when it comes to AI use. Students have started saving rough drafts of their work - a literal paper trail - to prove to their professors that their work is theirs if they are accused of AI use. There are countless posts on social media of students seeking support, guidance, and advice on how best to protect themselves in the event of a false accusation.

And sometimes, even these precautions are not enough, and they have to spend thousands seeking legal recourse. There’s growth in the legal sector in this arena, with entire law firms devoted to representing students falsely accused of cheating with or using AI.

We’ve created a world in education where it is their word - the human word - against the machine.

And the machine, unfortunately, usually wins.

Is this what we want when it comes to written and creative content - on Substack?

In Defense of Self-Disclosure: Trust Humans, not Machines

Substack says that we should know what we are getting - how content was generated. And I agree that reader mismatches between how they think what they are reading was written and the reality are important. That’s the ethical value of transparency - we deserve to know how something is made so that we can choose what we consume, what we read, where to spend our time, in accordance with who we are and what we value. Transparency supports our human autonomy in this regard.

But I wonder if this might be better accomplished by self-disclosure statements from the writers themselves rather than further pinning machine against human. And indeed, Substack is working on “how I make this work” statements for writers to utilize.

Of course, people will lie. We are relying on trust. Trust is messy. People will break it.

But if we offload the messy process of trusting humans to a machine - essentially relying on AI to detect AI - what kind of world are we creating? A world where one’s trustworthiness is defined by a machine rather than their own word and character.

In fairness, Substack will give authors the opportunity to run their drafts through Pangram beforehand and report anything they believe is false. This is good, but -

We will still be guilty until proven innocent - forced to run our drafts anxiously through Pangram, challenging false accusations, rather than being trusted as the writers and creators that we are. Just like university students. A culture of tension, saving our drafts, preemptively making our arguments in defense of our authenticity.

Our word against the machine’s.

So I will take this new evolution with a grain of salt. And I will prioritize AI use statements from the author themselves over Pangram’s detection.

I will also start adding statements at the end of my pieces for how AI was used. 99% of my content uses absolutely zero AI, as preserving my authentic voice and the way I write is important to me and my audience. But I have been playing around using the AI Lumo to help organize my scattered thoughts - especially as someone who is neurodivergent - occasionally. And while I never use Lumo to generate thoughts or ideas, I will now explicitly disclose it when it is used as organizational support.

I’m committed to including all of this at the end of my work and encourage all other authors to do the same so that there is always a human counterpoint to a machine judging their work. So that readers can choose whether to trust the machine or the human.