Provenance, Not Detection

In the last few weeks, much of the conversation in the AI sphere has been about AI writing. Pangram has become the de facto standard tool for identifying AI writing in general. GPTZero and Grammarly are also often used inside universities and classrooms. People have started collecting “Claudeisms” and other expressions that sound like ChatGPT. Feldar thinks the resulting digital witch hunt, combined with new EU transparency rules and machine-readable watermarks from large AI labs, is a recipe for disaster.

In the early stages of widespread AI adoption, the model acted as the executive branch of human writing. The user prompted the model to write a desired text, the model produced it, and the result was sent to its destination. Feldar believes that this kind of writing will soon cease to exist, especially for larger projects.

Human writing was never a straight line between the mind and the output. As soon as the tooling changed, it became an interplay between all the instruments available at the time. Humans also write in unique styles, so the process is different for everybody. We believe that the typical process now involves going back and forth between model and person many times. One person might use AI more extensively, rewriting whole passages, while another changes a single sentence in an entire essay. No matter what the process looks like, current detectors and watermarks demonise this behaviour in public, even though it is fundamentally no different from the process humans have followed for hundreds, if not thousands, of years.

AI-written articles are considered by some to be on par with human-written ones, while studies have shown that AI is already used extensively across assessed work. There is no going back to the old world. We understand why detectors are used in academia and other industries to ensure standards. But a detector, no matter how technologically capable, only meets the writing after the evidence of how it was made is gone. This is not a flaw in the detectors themselves, but the result of our decision to value the result in this way. Provenance in Feldar starts from the other end. We want our writing environment to preserve an accurate report of how a document was made and exactly which tools produced the final result, so there is no need to reconstruct it later with an external tool.

What worries us is that the witch hunt visible online devolves into a cat-and-mouse game. Students will try to find alternative ways around detectors so they can use the power of AI without being caught by their professors. Projects that aim to penetrate Pangram’s detection are being made available, but more often than not, the resulting prose becomes too disassociated to pass as human work. The Samsaran wheel of AI prose turns again when the next detector update arrives. Attached to this is social shame: an AI-written paragraph, no matter how comprehensive, is not considered on its merits but used to discredit the content of the text itself. The model took the patterns of human writing and reproduced them at scale, and the consequences are suffered by individuals unable to get past Turnitin.

In our last blog post, What Feldar Believes About Writing, we described linguistic centres of gravity. Models fundamentally compress human output and form structures. When a writer wants to refine a text, prompting a model means choosing a direction from its basin of possible languages.

No matter how open the request is, a simple prompt gives providers like ChatGPT and Anthropic room to apply whichever improvement costs least. Feldar’s concern is that, given enough time, this collapses into a small number of very large language areas. Product designers and others must preserve the distinctiveness of written language by pushing against these developments. This is why model providers want to implement watermarks: knowing when generation occurs and preserving that evidence downstream is a coherent goal. Detectors will never be able to establish this on their own. None of this is new; see Scott Aaronson’s work at OpenAI in 2022 or Google’s SynthID. No matter how transparent providers become, written language will move through these systems, and provenance needs to live somewhere else.

Feldar is a writing application that accompanies the user through the entire writing process, which makes it well suited for provenance. Every time a user applies a tool in Feldar to refine a manuscript, we record the contribution. We also distinguish between language written directly by the author and material from somewhere else. The resulting record gives us the ability to show exactly how a piece of work was produced.

Instead of being fearful of social pressure, the writer can use the system as a collaborator, much as they would work with a human editor. This goes beyond small personal projects. Universities and other institutions can use the resulting provenance report to set their own measurements for how much AI assistance is allowed. Classes and workshops that specifically focus on composition will most likely allow no or very little generated text. Other courses, where the purpose of the text is to convey information separately from stylistic interpretation, will most likely choose differently. These thresholds help examiners prepare for the coming influx of AI-written prose without remaining stuck in the past.

Feldar hopes this can make discourse around AI writing healthier. On Reddit, communities have formed in parallel, covering the same topics while differing mainly in their willingness to use AI. We want to break these dynamics. Where disclosure matters, the user should be able to show something that reflects the actual process. This is also why the provenance record in Feldar is private and owned by the author. A shared attestation can then help institutions verify the output.

Undoubtedly, there will always be gaps in the system. We do not claim that Feldar can reconstruct the process of a writer who generates a paragraph in a different window, then types it into a Feldar document, and flag it correctly. We want to make definite statements about what happens in our environment.

Once provenance outgrows the core Feldar app, we hope the protocol can be used for other interfaces as well.

The goal is to record different insertion points from model to manuscript alongside everyday writing, rather than relabel the work with a classifier after it has been completed. Provenance earns credibility through the limits of what it can claim, and this is why writing diversity matters. Once origin is clear, a writer using em dashes will no longer be desperate to find the acceptable course for their personal style. We leave language to the human, as it is meant to be. We help diversify language, opposing the underlying standardisation created by reliance on a few frontier models.

Writing should remain hard where the work calls for it. Much of writing is more cost than craft, and models can be genuine partners in that process. As models improve, writers must still choose between several viable directions. The machines should create more of those directions instead of constraining the person to the default.

The alternative is AI detection, where the score becomes the goal itself, detached from the underlying process. Writers change their prose to outplay classifiers rather than for themselves, with no visible cooperation between person and model.