"Humanmarking:" Authenticity Through Surveillance
The Unsettling Reality of AI-Powered Writing Certification
The mass deployment of AI systems is on the horizon, with Microsoft launching Copilot for Windows on September 26th and Copilot for Office 365 on November 1st. Google also recently rushed to deploy Bard and connect it to Gmail and docs, with lackluster results. We’re going to be awash in generative features in all of our apps soon and this is provoking a backlash. A few months ago, I predicted we’d see products designed to prove our writing as our own, and it looks like the creators behind GPTZero shipped such a system with their Human Writing Report. What I feared in my July 31st post Will 2024 Look Like 1984 appears to be coming true more rapidly than I imagined.
GPTZero’s core product is AI detection powered by machine learning (also known as AI). It doesn’t reliably work because the main factors they use for detection are easily outclassed by updated and new models and often misidentify human writing that contains lower lexical and perplexity scores. Now they’re pivoting to provenance certification—seeking to turn similar tools to certify your writing as being human and AI-free from your very Google Doc. The problem—if you upload your writing to their algorithm and allow them to label it as human and store it within their database, you are ceding your words to a private company under the promise they will use AI to protect it from AI. And it isn’t reasonable in my judgment to ask a student to do the same. Vanderbilt, The University of Pittsburg, Michigan State, and The University of Texas have all turned off Turnitin’s AI detection because it isn’t reliable. I don’t want education to rush the adoption of a new AI surveillance system.
A key point I argued in my July 31st post was how such systems that certified your writing as human chip away at our personal freedom and further erode trust:
Our desire to preserve what is human about text may mean ceding greater autonomy and freedom to automated systems we’re actively attempting to deny in our digital interactions. Instead of using AI to catch AI, as the previous generation of detectors used, your company, school, or state government may embrace AI-powered surveillance to monitor the work they receive from you to ensure it came from a human being, not a language model. Otherwise, there will be no way to prove your words originated from you, further chipping away at the artifice of trust we assume in our digital interactions.
Humanmarking Your Words
GPTZero now has a supposed method to “humanmark” your work and certify that a human being wrote it using their own machine-learning techniques. What’s more, they are also building a database where users can upload their writing and GPTZero will certify your words as human, where their “machine learning team will monitor LLM training companies, and inform you if and whenever your preserved content is being co-opted by an AI company for training.”
I argued before that unchecked generative AI would create a new economy, one where being able to prove your words were your own is commodity and GPTZero appears poised to cash in, stating “Whether students are sharing their writing to teachers, journalists to editors, content creators to their audience, all writers now have a valuable tool to not only assist them in verifying the authenticity of their work.” Here’s the opt-in language to Preserve Your Human Writing
Giving a machine learning company your private data so they can analyze it is like letting a stranger look through your diary, under the promise they will keep your secrets safe. Databases get hacked. Companies get bought out, along with their data, and trust me, a trove of human-marked data is very attractive to companies looking to train future LLMs.
Here’s how GPTZero markets its newest form of surveillance:
We developed a proprietary writing pattern analyzer that is able to provide ground-truth assessments on whether a document was AI generated based on your writing process. The report also contains helpful statistics such as a writing activity timeline, document lifespan, largest copy and pastes, average revision durations, and accounts for cases with multiple editors. When a document is written entirely without AI, the writing report shares both the writings stats, and a certification of human authenticity.
We’re headed toward a very uncertain future. Much of the world will soon be interacting with generative AI as features within their productivity software, and likely won’t be aware of what it is. The writer’s and actor’s strike over AI taking away their labor makes certification systems like the one GPTZero offers attractive because it is a simple technological solution to the problem—use our app and we will protect you from having to change your habits and help certify your labor as your own. The much harder path forward is learning to work and live in a world where generative text becomes common, where you and are employer use the technology daily as a practice, and in doing so, impact your labor, what it means to write, and what it means to create knowledge in digital spaces. Understandably, that’s a bridge most of us don’t want to cross, but it is one we will all have to eventually.
Free Module on AI Detection
I previously released a free module on AI detection from my professional development course on why AI detection isn’t reliable in educational settings. I’ve linked to it here as a resource to use and consider how you think about approaching generative AI with your students. Hint, I don’t think you should use it!
Updated Professional Development Course on Generative AI in Education
I’ve spent my free time updating my existing professional development course based on new updates, releases, and feedback. I’ve reorganized the course to make it easier to navigate and split the course into Introductory and Advanced sections.
Introductory Course: AI Literacy and AI Assistance
Advanced Course: AI Aptitude
Combined Course: Generative AI in Education
As before, the assignments for the course are free to download and use with your students!
Excellent post. I think there are ways that you can use writing tools, but give the power to combine human and AI writing to students (and forget detectors).