Identifying AI Content Is A Fool's Errand

This article rotted in my drafts for three years before initial publication and at no point before or since has there been a way to detect if media is generated by AI with 100% certainty, nor will there ever be in the future. The ‘vibe’ or stylistic signature of content is not a way to identify if it is AI made; especially as it becomes more common, that ‘vibe’ is driven out, and people learn to replicate it.

Generative AI is trained on the entire accessible and applicable corpus of human output. It is also being created and refined by humans trying to minimise any difference from what is human-creatable with almost limitless budget and resources. AI models are also actively trained to minimise distinguishable patterns, and any detectable patterns that are distinguishable become targets to eliminate.

We are beyond the point of being able to identify AI-generated content, and there is no way to reliably mark content as being AI-created in such a way that it cannot be circumvented.

Detecting Text and Speech

In response to the rise of large language models, a lot of ‘detectors’ launched – especially ones targeting the education sphere.

GPTZero is one of the more well-known and scrupulous detectors, though a single glance at their page explaining how their system works shows technobabble and vague diagrams. Throughout their marketing material they make claims of being the leading offering and the most accurate. In their FAQ, under the section ‘What are the limitations of AI Detectors?’, they write:

No AI detector is 100% accurate, and AI itself is changing constantly. Results should not be used to punish or as the final verdict. Accuracy improves with longer inputs (document-level results are stronger than paragraph or sentence-level) and is strongest on English prose.

This raises a fundamental question: what is the purpose of a ‘detector’ so unreliable that even the creators advise not trusting it for important use cases? What is the purpose of a tool like this at all if one mustn’t trust it and instead vet the material themselves?

The truth is, no matter what, students will cheat. Before AI, the equivalent was paying an essay writer or using plagiarism converters. For educators to tackle students using AI, they need to take another approach. Trying to offload the effort by using so-called AI detectors, which we have already established are flawed, isn’t the way to go about it.

Plagiarism is a gross case of academic misconduct, so levying accusations without just cause can be extremely damaging and isn’t something to be taken lightly. Automated AI detectors should not be considered admissible evidence, and many shouldn’t have their evaluations considered at all.

While it is true that people will mainly use the popular offerings, there are so many models out there – each with their own distinctions – such that it is impossible to identify patterns for each of them. Even if you could, it would be an endless task as new models are releasing continuously, and one could feasibly train their own or gaolbreak or fine-tune an existing model to evade detection.

People often point to writing features such as the em dash (—), tricolon, or emoji list as proof of text’s AI origin, forgetting that AI models are trained on human output and that all of the aforementioned have long been features of content in AI’s training corpora. Some people take the presence of these writing features to the extreme that they vocally and entirely unnecessarily attack writers despite them not using AI.

When I first tried GPT-3 in 2022 via the OpenAI Playground, I found that it wrote in a style almost identical to my own. My writing at the time had a certain quality of rigid verbosity, and GPT-3 mirrored that. It was bad enough that detectors, especially in their extremely rudimentary forms, often flagged it. As a student at the time, I was worried enough that I took conscious effort to change my writing style away from that previous form.

Beyond people who already have a style similar to popular AI offerings, there is also the case of the proliferation of AI content everywhere influencing how people are writing and speaking. The paper Empirical evidence of Large Language Model’s influence on human spoken communication has looked into this. To quote it:

This research explores whether the loop of human-machine cultural transmission is indeed closing, with humans adopting the language patterns generated by AI and vice versa.

To address this, we analyzed approximately 280,000 video transcriptions from over 20,000 academic YouTube channels, focusing on (i) changing trends in word frequency following the introduction of ChatGPT, and (ii) the correlation between the trend changes and word usage by ChatGPT. We show that, following the release of ChatGPT, word frequencies in human spoken communication began to shift, exemplifying the transformative impact of AI systems on human culture.

Even if not directly interacting with AI models, their output is so prevalent that it’ll be picked up second-hand. AI models are fundamentally warping language around them, and this will continue for as long as they exist.

Good AI voices can be a little uncanny if you actively listen for it but are largely beyond the hump now and indistinguishable at the cutting edge.

Detecting Imagery

Images from the most capable models have gotten really good. Good enough that I cannot distinguish them from real in passing, ¹ and even when pixel peeping find it difficult or impossible. The earliest models were blatantly AI, with nothing seeming cohesive and an ethereal property. As time progressed though, they’ve become more and more coherent.

Certain tells remained, such as an unexpected number of fingers or things interacting in odd ways. Certain details that weren’t the focus of the shot were deformed upon inspection, and details such as textures, clothing, hair, etc, would subtly fade and disappear in a non-Euclidean way.

Still, certain models have specific tells. For example, OpenAI’s 4o Image Generation has a tendency to apply a warm, sepia-tone filter of sorts over everything it generates, and this becomes more exaggerated as you request further refinements and it further yellows images.

Conversely, many models don’t have specific tells at all. There is no specific ‘look’ attributable to imagery generated by them, and some advanced models such as Google’s Nano Banana maintain consistency across revisions in such a way that they can output AI-altered versions of genuine photography (though with some very minor changes), further muddying the waters.

Video

Video is a collection of images, so much of the same is applicable. However, the introduction of motion comes with further tells. In early AI videos, continuity between frames was atrocious due to each image being generated as more-or-less individuals.

As the technology has progressed however, continuity has improved hugely, though physics and interactions remain objectionable. As of writing in September 2025, video can be generated with matching audio at the quality of individual images with reasonable object permanence. Still identifiable in most cases to the discerning viewer, but quickly joining the ranks of still image generation.

Possible Solutions

The obvious way to have things identified as AI-generated or altered is with a watermark. This is also completely futile.

How do you implement this watermark? As a visual indicator? What stops someone from just cropping it or erasing it? As metadata? What stops someone from just scrubbing it out?

One might argue that the presence of metadata could be evidence of something being AI, and its absence means it is either AI or not, but that falls apart if someone decides to apply AI identification metadata to non-AI content.

The Coalition for Content Provenance and Authenticity (C2PA) does exist and has members including OpenAI, Meta, Google, Microsoft, and plenty of other AI companies but is more or less useless for acting as a definitive source for if something is AI generated.

OpenAI note in their policy that:

Metadata like C2PA is not a silver bullet to address issues of provenance. It can easily be removed either accidentally or intentionally. For example, most social media platforms today remove metadata from uploaded images, and actions like taking a screenshot can also remove it. Therefore, an image lacking this metadata may or may not have been generated with ChatGPT or our API.

It is worth noting that most existing image formats lack any secure way to store metadata without the ability to tamper with it. Even if you somehow could make the metadata itself untamperable, one could always just screenshot or take a picture of it, creating a new image devoid of that metadata. Further, if at any point that format of storing metadata was broken or cracked, the validity of all previous images would be immediately thrown into question.

There is also the case that all existing generative AI content before the introduction of the metadata lacks any distinction, which allows it to slip through, and that anyone could train their own model that sidesteps these precautions entirely.

OpenAI have also dabbled with visual watermarks, with small coloured blocks in the bottom corner of their DALL-E output, but they were easily removed or avoided. OpenAI didn’t bring this visual watermark feature back for future image models either. Images generated with Gemini have a small watermark in the bottom left, but that is also easily removed.

A better approach is Google’s SynthID, which uses stylometry, but that is not a perfect solution and can be worked around by editing. One can also just use AI tooling without such watermarking built in.

Text can be watermarked by altering the probabilities of certain words, but this isn’t sure-fire and can be circumvented by back-translating the text, paraphrasing it, or employing one of a number of other transformations. It also won’t work forever, given language’s fluidity in adopting AI styling as referenced earlier. It is possible for the file containing the text to have its own metadata, but it is all too easy to just copy or transcribe text out of one file to move it into another without the metadata ².

Incentives Against Detection

There is also the case that there are parties who do not wish for AI-generated content to be identifiable. Detection companies have financial incentives to claim effectiveness even when limited, and AI companies have incentives to make their output indistinguishable.

People using AI output in many cases do not wish to have it identifiable as AI output – especially given the negative reputation AI output has for being slop and the potential backlash that brings with it. There are many con artists making substantial financial gains peddling AI output who I’d suggest would be willing to go to lengths to avoid people taking note ³.

Companies building generative AI offerings may not wish to have their output identifiable as AI, especially if their competitors don’t identify their output, which could prove a selling point in some cases.

I’ve also seen the argument that content that is modified by image editing software like Adobe Photoshop is not marked in any meaningful way, so AI output shouldn’t need to be either.

---

There is no reasonable way to identify AI-generated content with 100% certainty, and as time goes on, people who say, ‘I can always tell,’ are only sounding more and more disconnected from reality.

I wish to make clear that we have absolutely no reliable means to identify AI-generated content and that there is no reasonable way to implement an identification system in the future. Indistinguishable AI-generated content is here and upon us.

If you don’t think you’re exposed to AI-generated content, perhaps question if that is because you aren’t exposed to any or because it has reached a point that you don’t notice any.

Footnotes

1. Keeping in mind that I’ve written fairly extensively about AI, read a huge amount of papers, and spent countless hours experimenting with generative offerings. I am more able to identify generative AI output than the average person. ↩

2. I’ll include that phrases like ‘As an AI model…’ are not proof that text is AI generated. A human is capable of writing that and may even be likely to in the case of satire. ↩

3. Then again, the targets of these scams are often people who wouldn’t be able to identify even poor AI output. Much like many scams are intentionally flawed to filter out discerning individuals so they can prey upon the obtuse, perhaps content being identifiable as AI would come as a benefit. ↩