The ambiguity between human authored words and those generated by artificial intelligence is something I’ve grown increasingly uncomfortable with in recent months. It’s concerning to see the rise in high profile cases in which writers have been accused of using AI generated text in their work–like Mia Ballard’s Shy Girl and Jamir Nazir’s The Serpent in the Grove. I’m not here today to judge those works, but I would like to point out a number of flaws in the system writers and publishers currently rely on, and look at the options writers have to prove their human authorship.
We, humans, have found ourselves in the unusual predicament of being unable to prove our written words have been assembled by the living. This is particularly concerning for writers whose human authorship is the pillar of their career and credibility, like academics, students, authors, journalists and self-published writers on platforms such as Substack and Medium. Accusations of AI-use fly fast and wild, and the basis of such accusations tends to lean on AI detectors and the use of stylistic writing devices commonly used by AIs. AI detectors have become the go-to resource for determining what is and isn’t written by humans. They are not wholly accurate. In fact, their performance varies widely and can be subverted (as I will show). AI detectors are also built on artificial intelligence, which is prone to inaccuracies and hallucinations.
I’m not sure about you, but I refuse to let my credibility as a writer rest solely in the hands of an AI.
Over the past month, I’ve been doing some experiments and research into what writers should be doing to protect our credibility by establishing a forensic trail of our human authorship, and I’m sharing what I’ve learned with you here.
AI detectors get it wrong
As writers, the advice we’re often given by publishers who now find themselves nervously playing detective in this changed AI world, is to use our unique voices, perspectives and experiences, that our human emotions will shine through the AI slop. This advice sufficed for a while. But just over a month ago, news broke that Jamir Nazir’s short story The Serpent in the Grove was being investigated for being predominantly AI-produced. The story had just won the 2026 Commonwealth Short Story Prize for the Caribbean region, and I realised such advice, while important, wasn’t going to protect us. I want to write freely, and without comparisons to AI haunting my head. I want to be able to use em-dashes and groups of three. And I want to be the custodian of my own authenticity. So I began to run some experiments.
I took some of my work, which I know to be 100% human authored, and ran it through a few AI detectors. Two detectors, Pangram and GPTZero, cleared the work as being 100% human authored, but a third, Undetectable AI, reported that 29% of the script was likely AI generated. I was absolutely mortified, and still am, and instantly began looking at ways to prove my human authorship.
Before I get into those, however, I need to show you another important failing of AI detectors–the part where they pass up AI generated text as human authored. For this experiment, I asked Anthropic’s Claude to write me a 500-word flash fiction about a man who finds himself living as a gnome in his garden. Claude called the story Small Mercies, and I am horrified to admit to you that I was impressed. It bore the AI hallmarks of bizarre metaphors, but they suited the whimsical nature of the story and added a surrealist touch that entertained the mind. For those interested in reading it, I’ve copied it into the comments on this essay. I took Claude’s story and pasted it into the AI detectors, which all correctly identified it as being AI generated. I then ran the story through an AI humanizer–yes, many AI detectors also offer AI humanizer services. I used ZeroGPT (not to be confused with GPTZero), which reworked the piece into a mediocre version of the original, although not altogether terrible. Apparently, mediocrity is a hallmark of human authored words. I took this mediocre version and ran it through Pangram, often cited as one of the most reliable AI detectors. Pangram falsely cleared the work as 98% human authored with a 2% chance of AI use. That 2% came down to the last seven words of the story, which were: “They sat together until the light changed.”
I do not envy the publishers who must make decisions in such circumstances.
So we see that human work, particularly if it is well written, can be incorrectly classified as AI, and AI-generated text, once humanized, cleared as human.
How to prove human authored work
I’ve spent a few weeks researching the options writers have should we ever need to defend our credibility and, while there is no one-size-fits all solution, there are some practical steps we should be taking. Provenance–the ability to prove the origin of your work–is a fast moving space in the age of AI, and I believe it’s inevitable that some of the methods of proving provenance will become standard in the publishing industry in the not-too-distant future. Much of what we write takes time and our final works may not see the light of day for months or years. So it is critical you begin to track the provenance of your work now so that it holds its credibility in years to come.
Let’s start with the easiest methods, some of which you’re likely already doing, then move into the provenance technology currently in development.
Timestamps and Version History
AI dumps information onto a page at speeds humans cannot. It can produce a short work in seconds, a novel in minutes. Humans think about things, compose words over days, weeks, months or years. Some of the easiest tools at our disposal to demonstrate this human authorship process are the word processors many of us already use, like Microsoft Word, Google Docs and Scrivener. Each works a little differently, so it is important that you understand how your word processor tracks your work. For instance, Microsoft Word automatically creates timestamped version histories of your document–but only if you save your document to OneDrive. If you’re saving your documents to your hard drive, you’re missing out on this cloud feature. If you don’t already, you can download a desktop version of OneDrive that appears in your file explorer like any other drive on your computer and makes the process of saving your Word documents to the cloud seamless. Google Docs, which is a cloud-based service, also automatically keeps timestamped version histories as your document progresses.
If you don’t have access to those two methods, a simple way to replicate timestamped version histories is to email your document to yourself as you progress and file it in a specific folder for easy access. This forensic record of your writing is an important component in proving human authorship should you ever be required. But as sole evidence, it has some weaknesses; it doesn’t show if segments of text were copied and pasted from elsewhere–for instance from an AI. So a few other things are needed to round out your evidence. Before I jump to those, let’s take a quick look at some other things you may already be doing that help your cause.
Research, notes, journals and scrapbooks
If you’re like me, you probably have a research file longer than the actual story you’re working on. Your research, sources, journals, interviews, scrapbooks, narrative arcs, character notes, lists of edits, musings, all provide evidence of the thought processes behind your creations–something AI doesn’t have. Once again, a behavioural change you’ll want to make in this AI age is to ensure you keep and timestamp all that information in your word processor or notes tool, like Microsoft’s OneNote. If it’s a physical journal or scrapbook, take regular photos of new entries and email them to yourself to establish a timestamped record. This may seem over the top, but it will provide certainty that the work isn’t a forgery produced after-the-fact. As a bonus, it would also serve you well should you ever leave your journal on the bus.
Provenance technologies
With version history and timestamps, you will now have proof that you spent time crafting your words as well as evidence of the thought processes that inspired them. Together, these stand you in good stead in proving the text hasn’t been dumped on a page by AI. But as I previously mentioned, neither of these show whether any segments were copied and pasted from AI. This is where things get a little technical.
The good news is, the provenance technology to show what has been copied and pasted, along with edit history and human typing patterns, exists. The bad news is, it’s not mainstream and currently limited to third-party extensions in Google Docs. Access to the provenance information resides in that third-party platform rather than within the file itself, which should be the ultimate aim. That said, this is a fast moving space. There are provenance projects in development all over GitHub and last month, Microsoft launched its own Project Provenance. Granted, most of the focus has been on image and video provenance as a means of combating social media misinformation, and there is a lag where written words are concerned. So writers and publishers need to speak up and let tech developers know that we want our word processors to be able to embed provenance information into the file’s metadata at the point of creation. We want to be able to prove our own authorship, not leave it to be determined by third-parties and AI.
Before I go into the future of document provenance, let’s take a look at what is currently available in Google Docs. There are a number of third-party extensions available in Docs that track keystroke activity. The most popular is Draftback, which until recently was free, but due to a rise in demand, now has an annual fee. There’s another called Revision History, and the AI detector GPTZero also offers an extension. I tested Draftback to write this essay in Google Docs and found it simple to use. Once the extension is enabled, it tracks every key struck, every backspace or delete, and each copy and paste. You can render a report that demonstrates the human patterns of typing and editing, and this report will highlight any text that was copied and pasted. Large swaths of copy and paste would be a red flag. Draftback is meant to avoid flagging text that has been copied from within the document itself and pasted elsewhere, but it admits it misses some from time to time. I experienced this when I moved a few sentences higher in the piece. Those words lost their human-authored tag in the final report and were instead flagged as copied. Fortunately, Draftback and other such extensions allow you to watch a re-enactment of the document being drafted, similar to a video, which allows you to see those words being typed, then moved. A viewer can skip to the area that has been flagged and see with their own eyes how those words got onto the page.
A major shortcoming with Draftback is that you can’t easily share the report with someone. The viewer needs to be granted access to your file in Google Docs and they will also need to have the same extension enabled to view it. So the technology isn’t exactly where we want it just yet. Another shortcoming with keystroke tracking in general is that it can’t prove honesty. For example, it wouldn’t know if someone transcribed an AI-generated text in a manner which mimicked a human, and by that I mean stopping and pausing and changing our mind about things–a lot! This is where your version of history, research and notes come back to support you. All three combined provide a strong body of evidence of human-authorship. It is highly unlikely someone would generate an AI text in minutes then spend a year collecting research to support it while they transcribe it slowly, with mistakes and edits and revisions.
While on the topic of keystroke trackers, you may have come across Grammarly’s Authorship plugin for Microsoft Word, as I did. Unfortunately, I tested this one and it’s fairly useless in its present form. While it works similarly to Draftback, it only tracks a live session, and your session will time out if you leave it for more than 10 minutes. Session data isn’t collated, and you can’t download a report on that session, so you essentially lose your tracking data the moment you go to eat lunch. I believe it’s only a matter of time before Microsoft introduces its own solution. Which brings us to the future of human-authored provenance.
C2PA and provenance technology
Provenance technology has some very large backers including Microsoft, Google, Meta, Adobe, Amazon, OpenAI, the BBC and Sony among others. This group has formed what’s called the Coalition for Content Provenance and Authenticity (the acronym for which is C2PA). They’ve created a certification called Content Credentials. Content Credentials is already in use to certify the provenance of photos and videos circulating online by incorporating a tamperproof history of the file into the file’s metadata. The cryptographic C2PA certificate stamped into the metadata can be viewed by anybody with access to the file, allowing news organisations like the BBC to verify the authenticity of photo and video sources. If the file has been edited in any way, the changes will show up in the metadata. Now it’s important to establish that C2PA can also be used to certify the provenance of AI-generated material. In such a case, the AI origin would show up in the metadata. Smartphone and digital camera manufacturers like Samsung and Sony have begun to roll out the first products with C2PA built into their cameras, and this technology is set to become more widespread in the years to come. The expectation is that as Content Credentials become more prevalent, photos, videos and even music without C2PA certification will be viewed suspiciously.
This is the direction provenance technology for word processors is moving in. One third-party product you can find on GitHub is called WritersProof. It’s not yet available for desktop use, although you can download the code if you know what you’re doing. It aims to track keystroke activity in whatever word processor you’re using, create a tamperproof, timestamped cryptographic chain of activity, and a downloadable activity report with C2PA certification.
As I said, this is a fast moving space and new solutions are popping up all over the place, so I want to insert a little word of caution. Much of what’s currently in development is by individuals and small players, which is not in itself a bad thing, but it does raise security concerns. Please, do your due diligence before downloading any software that can track the keystroke activity of your computer. For most people, you’re going to want to wait until a mainstream provenance tool hits the market.
Human identity verification
There’s one last thing I want to cover. You may have seen The Author’s Guild recently launched a Verified Human Authorship Certification that writers can use to verify their identity and declare, using honesty, that their work is AI-free. Systems such as this may help human authors combat the growing number of fake authors with AI-generated e-books online, but it has no means of proving AI-generated text hasn’t been used in a human work. This program may evolve beyond honesty, but as it stands, flashing your ID and paying $10 won’t protect you against an AI allegation. As the Stanford student Theo Baker recently admitted in The New York Times, the use of AI-generated work among students is rampant, and honesty systemically fails to prevent them from ticking the box that declares their work AI free.
The take-away
So, there we go. That was a lot, and I hope it wasn’t overwhelming, so I will sum it up for you now.
AI isn’t going away, not everyone is going to be honest, and writers leave themselves in a precarious position if we let the sole determinant of our credibility and authenticity be an AI tool. AI detectors are not wholly reliable.
So you need to be aware of the forensic evidence your writing process leaves to ensure you have something to fall back on should an AI allegation ever swing your way in the future. Whatever process you choose, you want to ensure it is as simple and seamless as possible. Make use of timestamps and versions history. If your project is particularly important, you may want to consider using Google Docs with an extension like Draftback until new technologies come along. Keep your ears alert to the word ‘provenance’–there will be something in it for you. And above all, keep writing your way. Use em-dashes if you want to, groups of three, bizarre metaphors. And don’t be afraid. Keep control of your credibility.
Thank you for Mind Flexing with me. If you enjoyed this essay, please subscribe on Substack or your favourite podcast app, comment, click the ❤️ button, or share it with someone who would appreciate it. I’ll be back in a month. Until then, keep 💪
