nu

ChatGPT to invisibly watermark texts to identify them as AI-generated

ChatGPT to invisibly watermark texts to identify them as AI-generated

OpenAI is introducing an invisible AI watermark called textGrain to comply with European regulations requiring AI-generated text to be identifiable. The watermark embeds a specific word-choice pattern that can be detected by specialized software, though the tool is currently limited to researchers due to ongoing accuracy challenges.

OpenAI is implementing an invisible digital watermark, known as textGrain, to label content generated by its AI models, including ChatGPT and Codex. This initiative aligns with the European Union's AI Act, which requires AI-produced texts to be identifiable. Unlike visual watermarks, textGrain functions by embedding a secret, detectable pattern of word choices within the text itself. Currently, the detection tool is restricted to select researchers and expert organizations because the technology is not yet foolproof and remains susceptible to errors.

The effectiveness of the watermark is limited, particularly in shorter texts or passages that have undergone human editing. OpenAI acknowledges that altering even a small percentage of words can significantly reduce the detection rate. Furthermore, the presence of a watermark only suggests AI involvement rather than confirming the exact extent of its use. Consequently, the tool is not currently available for general public use by educators or employers, as results remain unreliable and easily circumvented through manual revisions or alternative AI platforms.

Based on reporting by nu. Translated and condensed by LocalHeadlines.

Related articles

Gall & Gall and Etos websites back online, apps remain unreachable

Gall & Gall and Etos websites back online, apps remain unreachable

The websites for Gall & Gall, Etos, and Albert Heijn are back online following a recent cybersecurity incident at Ahold Delhaize, though some app features and delivery services remain restricted. While the investigation into the cause is ongoing, the company has confirmed no current evidence of data theft and normal supply chain operations have resumed.

Number of cybersecurity companies more than quadrupled in five years

Number of cybersecurity companies more than quadrupled in five years

The number of registered cybersecurity companies in the Netherlands has quadrupled since 2021, reaching 962 by October 2024. This rapid growth is driven by increasing cyber threats and regulatory changes, highlighting the ongoing challenge for organizations to protect their digital systems.

Looking, looking, and fake buying: the dopamine hit without the costs

Looking, looking, and fake buying: the dopamine hit without the costs

New online shopping platforms are gaining popularity by allowing users to curate collections of products without the ability to actually purchase them. Experts suggest these sites serve as a psychological tool to help curb impulsive consumerism by providing the satisfaction of browsing without the financial cost.

Albert Heijn site and app back online, grocery deliveries resume Friday

Albert Heijn site and app back online, grocery deliveries resume Friday

Albert Heijn has restored its website and app following a cyberattack that caused a week-long outage across Ahold Delhaize platforms. Customers can access account features immediately, while the ability to place new grocery orders will resume this Friday.

Long wait times at Schiphol due to electronic passport control failure

Long wait times at Schiphol due to electronic passport control failure

Schiphol Airport is facing significant passenger delays after a technical failure incapacitated its automated eGate passport control system on Sunday morning. Royal Netherlands Marechaussee officers are currently conducting manual checks as they investigate the cause of the malfunction, which is not believed to be a cyberattack.