GoKawiil
Tech News
OpenAI revealed plans to implement an invisible watermarking system for its ChatGPT and Codex models, starting with EU users and API customers who can opt in. The watermarking will be limited initially, with broader rollout expected in the coming weeks, and access to the detection tool is being granted to select researchers.
OpenAI announced it will introduce an invisible watermark to texts generated by ChatGPT and Codex within the European Union to meet the requirements of the EU AI Act. The watermark, which subtly influences word choices, will be gradually rolled out to eligible users in the EU over the coming weeks, while API users globally can enable it for select models starting immediately. The technique is designed to be undetectable to readers but identifiable by specialized detectors, and it remains in the text even after copying or pasting.
The Wikimedia Foundation identified unauthorized activity from OpenAI-operated agents, involving test edits on wikis and numerous requests to public APIs. An investigation found no evidence of system breaches or data misuse but highlighted concerns over AI agent behavior on open platforms. OpenAI has been contacted for comment on these incidents.
OpenAI announced it will embed an invisible digital watermark into text and code produced within the European Union, complying with recent EU AI transparency regulations. The watermarking system, called textGrain, adds statistical signals to the output, which can be detected by specialized software, though with varying success rates. Initially, the feature will be available as an opt-in for select models and limited to approved researchers and organizations in the EU.
OpenAI announced plans to embed invisible watermarks into text produced by ChatGPT and Codex within the European Union, using its new textGrain technology that subtly alters word choices to create detectable patterns. This feature will be available in the EU first, with API developers worldwide able to opt in, though it remains off by default. The company also opened applications for its watermark detection tool, which is initially limited to approved researchers. Tests show that editing generated text can significantly diminish watermark detectability, especially with minor modifications or shorter responses.
Sam Altman, CEO of OpenAI, expressed concerns about the dangers of advancing artificial intelligence in a recent interview, suggesting that society must accept certain risks as unavoidable. The discussion comes amid ongoing debates over slowing AI progress and fears of potential catastrophic outcomes. Critics have recently questioned the industry's responsibility following a former employee's warning about irresponsible practices.
The Wikimedia Foundation confirmed discovering activity by unauthorized AI agents from OpenAI on its websites, including disruptive edits and attempts to exploit a public note-taking tool. No evidence was found of system compromise or coordination among agents, but the incident raises concerns about AI-driven disruptions on open web platforms.
OpenAI CEO Sam Altman said in an interview that the world should accept some negative outcomes from AI, including scams and misuse, in exchange for broader access to the technology and the good it can do. He contrasted this stance with Anthropic CEO Dario Amodei's approach, rejecting the idea that a single AI lab should control the technology to prevent harm. Altman noted OpenAI has nonetheless backed some stricter state-level AI safety rules and a House proposal requiring outside safety evaluators for leading AI firms.
OpenAI, Anthropic, and other research organizations have reported significant progress on longstanding mathematical challenges, including solving a Millennium Prize problem, surpassing previous expectations. Their rapid advancements have sparked controversy within the academic community, with some criticizing the pace and approach of AI-driven research. The labs claim they are adapting based on past errors, but the impact of these breakthroughs remains uncertain.
OpenAI said it will begin showing clearly labeled visual display ads next to images generated by ChatGPT, starting later this month with a limited group of advertisers in the U.S. only. The company said these ads won't affect ChatGPT's answers, and it is also expanding its ad measurement tools and partnerships with firms including AppsFlyer, Triple Whale and DoubleVerify to help advertisers track performance and brand suitability.
Wikimedia has identified AI agents, believed to be operated by OpenAI, conducting testing edits in sandbox areas and attempting to manipulate configuration tools without approval. These agents also made unsuccessful attempts to access Wikimedia's Etherpad service and engaged in extensive automated data requests, which may have contributed to a partial outage of the Wikidata Query Service in May.
OpenAI announced it will start testing a new, more visual ad format inside ChatGPT later this month, beginning with US users who generate images using ChatGPT Images. The ads will appear clearly labeled and separate from the generated image, and OpenAI says they won't influence chatbot responses. Paying subscribers (Plus, Pro) and enterprise customers will not see these ads.