OpenAI debuts a free web-based tool to help determine if text was written by a machine, rated as “very unlikely”, “unlikely”, “unclear”, “possible”, or “likely”
AxiosIna Fried
Context & Ripple Effects
OpenAI’s classifier arrived after GPT-3 had been framed as a meaningful advance in handling the ambiguities of human language, making attribution a more practical concern alongside generation. The tool’s later fate is central to its arc: OpenAI withdrew the classifier over its low accuracy, rather than treating its ratings as a durable detection product.
First-order effects
People assessing submitted text gain a free OpenAI tool that converts its judgment into five machine-authorship likelihood bands.
OpenAI adds a public-facing detection product alongside its text-generation technology, making the reliability of those ratings part of its own product responsibility.
Second-order effects
Organizations considering the ratings for screening or review lack a dependable basis for automated decisions after OpenAI shut the tool down for low accuracy.
The withdrawal shifts OpenAI’s detection work away from this classifier and toward the more effective provenance techniques it said it would research.
Third-order effects
The later report of a ChatGPT watermarking method suggests AI-text assurance may move from inferring authorship from prose toward embedding provenance at generation time, though OpenAI’s internal debate shows deployment is not automatic.
The trend: AI assurance is moving from post-hoc text classification toward provenance mechanisms built into generative systems.
This had to happen, if not for the obvious ethical reasons then at least for legal ones: OpenAI launches AI classifier for indicating AI-written text https://openai.com/...
Translation: they created another SOTA benchmark that will be used to measure “improvements” in generative large language models like GPT based on how “undetectable” their outputs are https://twitter.com/...
OpenAI is responding to widespread concern about the technology it's put out into the world by releasing an AI text detector that is accurate 26% of the time and yields false positives 9% of the time https://openai.com/...
Like an AI tool that can tweak AI text to make it undetectable as having been written by AI isn't going to be out within the week. https://twitter.com/...