Thou who dares to plagiarize with ChatGPT, take heed! OpenAI hath crafted a formidable tool capable of discerning GPT-4’s written output with nigh 99.99% accuracy. Yet, alas, the company hath languished in indecision for over a year, wavering betwixt releasing this tool unto the public realm.
‘Tis reported that a deliberate approach is being taken by the company, pondering the complexities and broader impact upon the realm beyond OpenAI. According to the scribes at TechCrunch, “The text watermarking method we’re forging shows great promise, but holds risks we must balance as we explore alternatives.” An emissary from OpenAI hath spoken thus, citing susceptibilities to circumvention by malevolent forces and potential disproportionate effects upon non-English speakers.
The text-watermarking contrivance hath an enchanting charm, weaving a covert pattern into the model’s output detectable solely by the OpenAI tool, unbeknownst to the end user. Whilst this tool may discern the words wrought by GPT-4, ’tis powerless against other models like Gemini and Claude. Furthermore, the watermark itself is easily erased by transmuting the text through the eldritch Google Translate, shifting it betwixt tongues and back again.
Nay, ’tis not the first time OpenAI hath endeavored to craft a text-detection tool. A previous attempt foundered in the turbulent seas of inaccuracy, meeting an untimely end due to its woeful detection rate and penchant for false positives. Launched in the year of our Lord 2023, this detector required a compendium of text no less than 1,000 characters ere rendering a judgment. ‘Twas able to distinguish AI-generated content with a mere 26% accuracy, branding human-created content as AI-made 9% of the time. One sagacious professor at Texas A&M was thus led astray, failing an entire class in grave error.
In light of potential uproar from users, OpenAI treads cautiously upon the precipice of releasing this tool. The Wall Street Journal reports that 69% of ChatGPT users deem such a tool unreliable, fearing it shall birth false accusations of deceit. An additional 30% vow to bid adieu to the chatbot if this feature were to come to pass. The specter of developers unraveling the watermark, rendering it null, haunts the company’s decision-making.
Whilst OpenAI contemplates the fate of its watermarking system, other AI entities race to unleash their own text detectors. GPTZero, ZeroGPT, Scribbr, and Writer AI Content Detector join the fray, yet their veracity pales in comparison. The human eye, that ancient tool of discernment, stands as the most steadfast guardian against AI subterfuge—a scant comfort in these tempestuous times.
