Following a report from The Wall Avenue Journal that claims OpenAI has been sitting on a instrument that may spot essays written by ChatGPT with a excessive diploma of accuracy, the corporate has shared a little bit of details about its analysis into textual content watermarking — and why it hasn’t launched its detection technique. Based on The Wall Avenue Journal’s report, debate over whether or not the instrument must be launched has stored it from seeing the sunshine of day, regardless of it being “prepared.” In an replace printed on Sunday to a Might weblog publish, noticed by TechCrunch, OpenAI mentioned, “Our groups have developed a textual content watermarking technique that we proceed to contemplate as we analysis options.”
The corporate mentioned watermarking is considered one of a number of options, together with classifiers and metadata, that it has appeared into as a part of “in depth analysis on the world of textual content provenance.” Based on OpenAI, it “has been extremely correct” in some conditions, however doesn’t carry out as properly when confronted with sure types of tampering, “like utilizing translation techniques, rewording with one other generative mannequin, or asking the mannequin to insert a particular character in between each phrase after which deleting that character.” And textual content watermarking might “disproportionately influence some teams,” OpenAI wrote. “For instance, it might stigmatize use of AI as a helpful writing instrument for non-native English audio system.”
Per the weblog publish, OpenAI has been weighing these dangers. The corporate additionally wrote that it has prioritized the discharge of authentication instruments for audiovisual content material. In a press release to TechCrunch, an OpenAI spokesperson mentioned the corporate is taking a “deliberate method” to textual content provenance due to “the complexities concerned and its probably influence on the broader ecosystem past OpenAI.”
