Minister for Recovery, Infrastructure and Transport Mykola Kalachnyk announced agreements with the Romanian side on expanding logistics routes, railway repairs and increasing the capacity of the Sulina Canal.
The political force ended its joint work with Oleksiy Honcharenko, citing fundamental differences in views on ending the war and state security.
Saudi Aramco chief warns global oil stocks have reached critically low levels. He says market tension will persist until shipping normalises through the strategically vital waterway.
The interior ministers of Lithuania and Poland discussed plans to relocate at least one million Baltic residents to Poland in the event of an escalation of the conflict with Russia.
The AI content labelling technology will appear in ChatGPT and Codex for EU users, while becoming globally available via API.
The Dnipropetrovsk District Administrative Court ruled that PlayCity's decision to annul Private Joint-Stock Company PATRIOT's licence was unlawful, restoring the company's status as a state lottery operator.
Oleksandr Kostorny, whom Ukrainian investigators considered an FSB agent involved in passing coordinates of the Yavoriv training ground, has died. He died in June 2026, but the information was only confirmed recently.
On 7 October, parliament will consider bills on customs, tax and criminal legislation. Deputy Yaroslav Zheleznyak stated that the vote would proceed without complications.
OpenAI is preparing the launch of a technology for invisibly labelling texts created by artificial intelligence. The system, named textGrain, will be embedded in the ChatGPT and Codex interfaces for users from the European Union. The company plans to begin labelling within the coming weeks.
As OpenAI announced, the decision to implement this is linked to the implementation of the European AI Act. For developers using the OpenAI API, the option to add such watermarks is already available. This feature will work for API clients worldwide, although the possibility of disabling it is provided.
The tool for detecting the labelling will not yet be open to the general public. Approved researchers and expert organisations will be the first to gain access to it. The company expects their assistance in testing the system and further refining the algorithms.
The textGrain technology identifies AI-generated texts by analysing specific patterns in the model's word choice during the generation process. If the text is of sufficient length, a hidden signal, imperceptible to the human eye, forms in the sequence of chosen words. The detector scans a large number of such choices to find this characteristic marker.
Meanwhile, OpenAI emphasises that the watermark will not make it possible to identify a specific user, their account, a particular query, or the chat history with ChatGPT. Furthermore, the system does not indicate which part of the text was written by a human and which by artificial intelligence.
Developers acknowledge that the current version of the system does not guarantee 100% detection accuracy. According to OpenAI, for texts of 200 tokens, the watermark is detected in approximately 80% of cases, while for 400 tokens, this figure is around 95% (assuming a target false positive rate of 1%).
The effectiveness of verification is significantly reduced if the text is edited. Testing established that replacing 10% of words with synonyms lowers the probability of detecting the watermark from approximately 92% to 66%. If a quarter of the words are replaced, the figure drops to 17%.
Therefore, the absence of a detected watermark is not proof that the text was written by a human. Previously, the company Anthropic announced its intention to use a similar technology, also planning to mark content generated by its Claude model with invisible watermarks following the adoption of the European AI Act.
Comments
No comments yet. Be the first!