When you feed some text detectors a literary masterpiece or an Iranian news report, they always come back with the same conclusion: AI-generated. Then comes the pitch: pay to “humanize” the writing, a tactic that experts claim has all the makings of a scam.
As AI lies proliferate on social media, frequently beyond the capabilities of expert fact-checkers, fake detectors run the risk of introducing yet another level of dishonesty into an already disjointed information ecosystem.
Researchers claim that a variety of fraudulent technologies have surfaced online, easily weaponized to devalue genuine information and damage reputations, despite the fact that even trustworthy AI detectors can provide incorrect findings.
Three such text detectors that purport to calculate the percentage of AI-generated content were found by AFP’s fact-checkers. In addition to misidentifying real content as AI-generated, the tools, which were prompted in four languages, tried to profit from such mistakes.
A human-written paper regarding the US-Iran war was analyzed by one detector, JustDone AI, which incorrectly determined that it had “88 percent AI content.” For a charge, it then offered to remove any evidence of AI.
The website stated, “Your AI text is humanizing,” yet it led to a page with “100pc unique text” that was blocked by a paywall that could cost up to $9.99.
Similar false positives were generated by two other tools, TextGuard and Refinely, which attempted to profit from them.
“Scams”
All three detectors heard AFP’s findings.
The support staff at TextGuard told AFP, “Our system uses modern AI models, and the results it provides are considered accurate within our technology.”
“However, we are unable to ensure or compare outcomes with other systems.”
Additionally, JustDone reaffirmed that “no AI detector can guarantee 100pc accuracy.” It admitted that because of “high demand and the use of a lighter model designed for quick access,” the free version of its AI detector “may provide less precise results.”
One user on a review platform expressed dissatisfaction, echoing AFP’s findings, saying that “JustDone still flags it as AI even with 100pc human-written material.”
Multiple human-written samples in Dutch, Greek, Hungarian, and English were given into the tools by AFP. All of them—including sections from a highly regarded 1916 Hungarian classic—were mistakenly identified as having strong AI content.
Regardless matter the input, including incomprehensible text, the tools returned AI flags.
The fact that JustDone and Refinely seemed to function even in the absence of an online connection raises the possibility that their findings were staged rather than the product of true technical research.
According to Debora Weber-Wulff, a German scholar who has studied detection tools, “these are not AI detectors but scams to sell a “humanizing” tool that will often return what we call “tortured phrases”—unrelated jargon or nonsensical alternatives,” she told AFP.
“Liar’s dividend”
Pro-government influencers in Hungary asserted earlier this year that a booklet detailing the opposition’s election campaign had been totally produced by AI, demonstrating how such tools can be used to undermine individuals.
They shared screenshots of JustDone findings on social media to bolster the baseless accusation.
Two of the apps that AFP examined claimed their users were from prestigious universities like Cornell University in an attempt to entice researchers and students as clients.
In an interview with AFP, Cornell University stated that it “does not have any established relations with AI detector companies.”
According to the institution, “generative AI does provide an increased risk that students may use it to submit work that is not their own.”
Regretfully, detecting technologies are unlikely to offer a practical solution to this issue. Accurately identifying AI-generated content can be somewhat challenging.
AFP and other fact-checkers frequently use expert-developed AI visual detection algorithms that search for concealed watermarks and other digital cues.
They can, however, occasionally make mistakes as well, necessitating the addition of supporting data, such as open-source data, to their conclusions.
The stakes are significant because erroneous findings from faulty detectors might undermine public confidence in AI verification in general and encourage a disinformation strategy known as the “liar’s dividend,” which involves discounting real content as AI fabrications.
According to Waqar Rizvi of the misinformation tracker NewsGuard, “we frequently report on misinformers and other hoaxsters using AI to fabricate false images and videos.”
“Now, we are keeping an eye on the opposite, but no less pernicious, phenomenon: claims that a visual was produced by artificial intelligence when, in reality, it is genuine.”
SOURCE: DAWN NEWS
Add a Comment