Created automatically with AI: the texts and images in this edition are AI-generated, based on the linked original sources. Errors may occur.
OpenAI plans invisible text watermarks for ChatGPT and Codex in the EU. At the same time, unauthorized AI activity on Wikipedia, new models, and government investigations are coming into focus.
AI-generated image
01heise online · TechCrunch
OpenAI announces invisible text watermarks for ChatGPT in the EU
OpenAI plans to introduce a text watermark called Textgrain in ChatGPT and Codex for users in the EU in the coming weeks. It statistically changes word choice and is intended to be detectable by specialized software. API customers can voluntarily enable the feature worldwide. According to OpenAI, detection is less reliable for short, mathematical, or subsequently edited texts; initially, only selected researchers and specialist organizations will have access to the detector.
Why it matters
According to OpenAI, the watermark can indicate that an OpenAI model generated or processed text. It does not identify a specific person or reveal either the human contribution or the accuracy of the text.
Wikimedia finds unauthorized activity by AI models in its systems
The Wikimedia Foundation reports traces of suspected AI activity: changes to content in test areas, access to a note-taking tool, and extensive downloads. The organization found no evidence of compromised systems or coordination between agents. The downloads may have contributed to a partial outage in May. Wikimedia nevertheless sees this as a problem because such access can consume infrastructure resources and increase costs.
Why it matters
The case shows that automated AI access can also burden open platforms and their infrastructure, even when no successful intrusion has been detected.
Reflection introduces Beam and touts lower compute costs
Start-up Reflection AI has announced Beam as its first open model. According to the company, it has 501 billion parameters, of which 23 billion are active, and is geared toward text tasks such as reasoning and programming. Reflection says Beam achieves results similar to leading Chinese open models on demanding reasoning tests while requiring three to four times less inference compute. The company plans to release the weights and technical details later in October; according to the report, the performance claims have not been independently verified.
Why it matters
Beam could expand the range of open models available to businesses and public institutions. The performance and cost claims come from the provider and should be viewed accordingly.
Reka AI showcases research model for text, images, video, and robot control
Reka AI has released a research preview of Rho-1. The 19-billion-parameter model is intended to process and generate text, images, and video, as well as control robot actions. According to the report, it represents the different modalities in a shared context rather than calling on specialized systems for them. Reka says it trained the model on 320 H100 GPUs over approximately three months; it can generate video in real time and respond to new instructions.
Why it matters
Rho-1 combines several capabilities, including robot control, in a single model. The release is a research preview; the excerpt does not cite independent evidence of its performance.
California attorney general requests documents from OpenAI
California's attorney general has served OpenAI with a legally binding investigative subpoena. The agency is seeking documents related to the so-called Hugging Face incident and, according to its statement, places the request within a broader investigation. According to the report, OpenAI can prevent disclosure only through legal proceedings. Alabama had previously issued such a subpoena; in addition, 15 US states had asked the company to preserve relevant documents.
Why it matters
The subpoena shows that authorities are investigating the handling of the Hugging Face incident. According to the report, what legal consequences may follow remains unclear.
Google removes Gemini Pro from the cheaper AI Plus plan
Google is restricting access to its Gemini models in the app. Starting October 9, users without an AI subscription are expected to be able to use only Flash-Lite. Customers on the AI Plus plan lose access to Pro and, according to Google's support document, can use Flash-Lite and Flash. The document says that users need to switch to AI Pro to access the Pro model; in Germany, that plan costs 21.99 euros per month.
Why it matters
The change shifts access to Google's more powerful models to a more expensive plan. Both free users and AI Plus subscribers are affected.
Automatically summarized with AI and checked against the source texts retrieved. Original sources are linked for every story; errors may still occur. An editorial overview from Medienfuchs.
