We Need to Talk About AI Detectors
Over the past few weeks there have been two posts in particular that have come out against two authors in the community that have resulted in near witch hunts for one, and essentially driving the other off the internet from all the bullying and hate received. These posts were related to the potential use of AI in fanfiction, as well as using supposed “AI Detectors” to support their claims. With the help of friends, we have been able to look into the AI claims that were made against both The Silence and The Song and Ir Abelas, Da’ean.
We were curious about how and why these posts were being flagged with high levels of “AI Probability” when the authors have been adamant (either in chats or in public) that they have never used generative AI for their work. So we did the most logical thing, put on our detective caps, and rolled up our sleeves. We would like to note that we do not wish to have philosophical discussions, we wish to have transparency and honesty.
Spoilers: We found inaccuracies almost IMMEDIATELY.
Firstly, we looked into the weakness of AI detectors, and read through online posts where people voiced their frustrations with detectors. One thing that we noticed was that the common denominator was that well written articles were being flagged as “Likely AI”, particularly with Originality, and that the solution was to either “dumb them down” or to remove punctuation such as commas, which immediately improved the score, tipping the scale to “Likely Original”.
For the second step, we ran some of our own works through Originality AI checker–works that were made prior to the creation of AI and generative AI. However, after punctuation was removed, this magically changed from 50% likelihood of AI to 100% original work. Again, these works were from before the dawn of generative AI, and therefore could not have been created by AI. For fun, we even ran the first chapter of Harry Potter through it–a novel that is objectively without AI, which still did not come out with results stating 100% original work. We then removed almost all of the punctuation from it, and it actually improved the originality score by 3% (from 95% to 98%).
Personal fic, before and after:
Originality.ai - AI Content Detection
Originality.ai - AI Content Detection
Next step, we ran our own scans through Originality and Quillbot.This includes full chapters of Ir Abelas, Da’ean, both with and without punctuation in Originality, and excerpts in Quillbot. Interestingly, the excerpts in Quillbot pinged as “0% likely AI”, and that is without any removal of punctuation. Across the board, the removal of punctuation from the chapters caused an immediate and dramatic increase in the score, from “100% likely AI” to “96% likely Original”. We have found that the more grammatically correct a work was, the more likely it was to be flagged as “AI”, much like how the freelance writers were complaining about.
Chapter 1 of Ir Abelas, before and after:
Originality.ai - AI Content Detection
Originality.ai - AI Content Detection
Chapter 1 through Quillbot:
Chapter 45 of Ir Abelas, before and after:
Originality.ai - AI Content Detection
Originality.ai - AI Content Detection
Chapter 45 without commas or double hyphens:
Originality.ai - AI Content Detection
Chapter 45 through Quillbot:
Even the Ai detection websites caution against this:
To Durgeapologist, Fangbanger3000, and friends: If you do actually read this, I hope you realize that your posts have done more harm than good to the community. You are correct that AI is a potential threat to creative spaces, but you have gone about addressing it in the worst possible way. By creating multiple posts across platforms with the intent of creating a negative perspective toward certain authors and their fictions rather than the use of AI as a whole—not to mention the counter-accusations with personal attacks rather than focusing on the issue at hand—you are creating an environment that fosters negativity, bullying, and division—none of which are directions to take a sustainable and healthy community. AI Detection is the Wild West right now. There is no way to determine if something written is AI through the use of algorithms, and it requires the use of human intervention and careful comparison to previous works to be within a certain level of certainty that it is AI. Our hope is that in the future, you will take time, step back, and consider all possible sides before causing a stir in the community like this.
