OpenAI will begin adding an invisible watermark to eligible text generated by ChatGPT and Codex in the European Union (EU), as the company moves to comply with transparency requirements under the EU’s AI Act.

The company said the rollout would take place over the coming weeks and apply to eligible users across all ChatGPT and Codex plans in the EU. It said it was not making text watermarking a global default at launch.

Separately, from October 5, API customers worldwide can opt in to text watermarking for select models, OpenAI said.

The mandate

The move comes as Article 50(2) of the EU AI Act requires providers of generative AI systems to ensure that AI-generated or manipulated content, including text, is marked in a machine-readable format and detectable as artificially-generated or manipulated.

The European Commission says the technical measures must be effective, interoperable, robust and reliable “as far as technically feasible”, taking into account the limitations of different types of content. The transparency obligations have applied since August 2, 2026.

The law does not specifically require companies to use a watermark. The European Commission has also developed a voluntary Code of Practice on Transparency of AI-generated Content, which sets out practical measures for providers to comply with the Article 50 requirements.

Section 1 of the Code deals with the marking and detection of AI-generated content by providers and Anthropic, Google, Meta, Microsoft, Mistral and OpenAI have signed this section of the Code.

Technology: textGrain

OpenAI said its text watermarking technology, called textGrain, adds an invisible statistical signal to the model’s choice of words. It does not insert hidden characters or additional watermark-specific tokens. Instead, the statistical pattern in the model’s word choices can subsequently be analysed by a detector.

A technical paper published by OpenAI alongside the announcement explains that the watermark is created during generation by subtly influencing token selection using a secret key and the preceding context. The detector can later use the generated text and secret key to look for the statistical signal.

Limitations

OpenAI said its detector identified watermarks in about 80% of 200-token passages and about 95% of 400-token passages in psychology-related content, at a target false-positive rate of 1% (meaning it would wrongly flag about one in 100 texts without the watermark). Detection was substantially lower for mathematics, where there is less flexibility in word choice.

Editing also significantly weakened the signal. In an evaluation of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%, while replacing 25% of words reduced detection to about 17%, according to OpenAI.

The company, therefore, said it would initially restrict access to its text-watermark detector to approved researchers and expert organisations.

It said the detector would report whether an OpenAI watermark was detected but would not identify the user or reveal prompts or conversations. OpenAI said the decision reflected the risks of false positives and missed watermarks.

The limitations are significant because the watermark is intended as a provenance signal rather than a definitive AI-authorship test.

The frontier AI company said a watermark cannot determine how much human judgment, editing or creativity went into a passage. It said it also cannot establish ownership or responsibility, identify the person or organisation that generated the text, or determine whether the text is accurate.

Conversely, the absence of a detected watermark does not establish that a human wrote the text. OpenAI said detection may fail because content is too short, has been edited or translated, comes from an unsupported model, predates watermarking or was generated by another AI system.

Separately, EU rules require deployers to clearly label certain AI-generated or manipulated text published to inform the public about matters of public interest.

OpenAI plans to make textGrain available as open-source technology, saying this could allow others to build on the approach and help improve text watermarking.

The company said it is also trying to make text watermarking as one part of a broader provenance system. For images and audio, OpenAI uses technologies including Content Credentials based on C2PA and invisible SynthID watermarks, alongside verification tools. It said no single provenance technique was sufficient on its own.

Published – October 06, 2026 09:06 pm IST

OpenAI will begin adding an invisible watermark to eligible text generated by ChatGPT and Codex in the European Union (EU), as the company moves to comply with transparency requirements under the EU’s AI Act.

The company said the rollout would take place over the coming weeks and apply to eligible users across all ChatGPT and Codex plans in the EU. It said it was not making text watermarking a global default at launch.

Separately, from October 5, API customers worldwide can opt in to text watermarking for select models, OpenAI said.

The mandate

The move comes as Article 50(2) of the EU AI Act requires providers of generative AI systems to ensure that AI-generated or manipulated content, including text, is marked in a machine-readable format and detectable as artificially-generated or manipulated.

The European Commission says the technical measures must be effective, interoperable, robust and reliable “as far as technically feasible”, taking into account the limitations of different types of content. The transparency obligations have applied since August 2, 2026.

The law does not specifically require companies to use a watermark. The European Commission has also developed a voluntary Code of Practice on Transparency of AI-generated Content, which sets out practical measures for providers to comply with the Article 50 requirements.

Section 1 of the Code deals with the marking and detection of AI-generated content by providers and Anthropic, Google, Meta, Microsoft, Mistral and OpenAI have signed this section of the Code.

Technology: textGrain

OpenAI said its text watermarking technology, called textGrain, adds an invisible statistical signal to the model’s choice of words. It does not insert hidden characters or additional watermark-specific tokens. Instead, the statistical pattern in the model’s word choices can subsequently be analysed by a detector.

A technical paper published by OpenAI alongside the announcement explains that the watermark is created during generation by subtly influencing token selection using a secret key and the preceding context. The detector can later use the generated text and secret key to look for the statistical signal.

Limitations

OpenAI said its detector identified watermarks in about 80% of 200-token passages and about 95% of 400-token passages in psychology-related content, at a target false-positive rate of 1% (meaning it would wrongly flag about one in 100 texts without the watermark). Detection was substantially lower for mathematics, where there is less flexibility in word choice.

Editing also significantly weakened the signal. In an evaluation of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%, while replacing 25% of words reduced detection to about 17%, according to OpenAI.

The company, therefore, said it would initially restrict access to its text-watermark detector to approved researchers and expert organisations.

It said the detector would report whether an OpenAI watermark was detected but would not identify the user or reveal prompts or conversations. OpenAI said the decision reflected the risks of false positives and missed watermarks.

The limitations are significant because the watermark is intended as a provenance signal rather than a definitive AI-authorship test.

The frontier AI company said a watermark cannot determine how much human judgment, editing or creativity went into a passage. It said it also cannot establish ownership or responsibility, identify the person or organisation that generated the text, or determine whether the text is accurate.

Conversely, the absence of a detected watermark does not establish that a human wrote the text. OpenAI said detection may fail because content is too short, has been edited or translated, comes from an unsupported model, predates watermarking or was generated by another AI system.

Separately, EU rules require deployers to clearly label certain AI-generated or manipulated text published to inform the public about matters of public interest.

OpenAI plans to make textGrain available as open-source technology, saying this could allow others to build on the approach and help improve text watermarking.

The company said it is also trying to make text watermarking as one part of a broader provenance system. For images and audio, OpenAI uses technologies including Content Credentials based on C2PA and invisible SynthID watermarks, alongside verification tools. It said no single provenance technique was sufficient on its own.

Published – October 06, 2026 09:06 pm IST

Major artificial intelligence (AI) platforms, including OpenAI’s ChatGPT, faced disruptions. File
| Photo Credit: AP

In 2015, a small group of researchers and technology entrepreneurs gathered in San Francisco to create what they described as a gift to humanity, eventually to turn into enemies defining the future of artificial intelligence (AI) in a courtroom.

Elon Musk loses blockbuster OpenAI suit as jury says too late

OpenAI was founded on the premise that if artificial general intelligence (AGI) was coming regardless, it was better to have safety and ethics-conscious researchers build it. AGI is the kind of AI that can match or surpass human cognition. The group decided to create the initiative as a nonprofit and if the technology ever arrived, it would belong to everyone, as open source. OpenAI’s CEO Sam Altman even went as far as to design the company’s board such that it could fire him too if he ever came in the way of its core mission. That founding promise, to build ethical AGI, years later brought Elon Musk and Mr. Altman to a federal courthouse in California.

Eleven years since the founding, no longer friends and both pursuing their own paths to AGI, Mr. Musk and Mr. Altman are both convinced that the other has betrayed everything they had vowed to build together.

Mr. Musk brought forward the lawsuit against Mr. Altman, accusing him, OpenAI president Greg Brockman, and Microsoft, of manipulating him into donating to a public-interest organisation only for it to later attach a for-profit subsidiary and accept billions from Microsoft.

On May 18, a nine-person jury took less than two hours to throw out Mr. Musk’s case against Mr. Altman. The verdict did not settle the question of whether OpenAI broke faith with its founding mission but instead, it was settled on a procedural issue – that Mr. Musk had waited too long to sue. In the American legal system, civil claims must be filed within a fixed window of time, crossing which the claim becomes invalid. The jury said Mr. Musk’s claim fell outside the statute of limitations. The judge concurred, but Mr. Musk’s lawyers have signalled that they might appeal. The merits of the case were never discussed.

The nonprofit debate

Trial testimony and evidence showed OpenAI’s leadership had anxiously tracked Google’s acquisition of another AI research giant DeepMind in 2014.

OpenAI’s executives argued that the nonprofit structure was not sufficient to compete with bigshots like Google.

Mr. Musk was the company’s biggest early donor, contributing around $38 million. But, the economics of building AGI turned out to be brutal. Training large models requires computing infrastructure that costs billions of dollars. By 2019, a year after Mr. Musk’s departure from the OpenAI board, the company decided it could not remain competitive as a pure nonprofit. It attached a for-profit subsidiary to the existing organisation, with the nonprofit retaining oversight and a capped return structure for investors.

Microsoft came in with an initial investment that year, and kept coming, eventually investing more than $135 billion and holding a significant share in OpenAI. This is the transformation Mr. Musk’s lawsuit sought to attack, from nonprofit to a commercially driven AI laboratory he had set out to oppose.

Claims that did not make it to trial

The bit that went to trial eventually was a much smaller subsect of the original claims Mr. Musk had filed in 2024. Several of his claims were dropped or narrowed before the trial even began.

Mr. Musk claimed that Microsoft’s investment in OpenAI had aided it to breach its charitable mission. But, despite a Microsoft executive testifying that the company had invested billions in OpenAI, the jury dismissed the claim.

Mr. Musk also filed an antitrust claim against OpenAI and Microsoft, alleging that the two companies were colluding to dominate the AI market. This claim has not reached the jury yet. Judge Gonzalez Rogers heard arguments and signalled she was sceptical, noting the aggressive competition already playing out across the AI industry. She has not ruled on it, meaning that particular thread remains unresolved. OpenAI on its part filed a countersuit accusing Mr. Musk of running a years-long harassment campaign against the company. A judge ruled in August 2025 that those claims were legally sufficient to go forward. That case is separate and still ongoing.

Key takeaways from the proceedings

The legal question the jury resolved was procedural. But the three weeks of testimony preceding the verdict were pure Silicon Valley tech theatre.

Mr. Musk’s lawyers assembled a long roster of witnesses who testified that Mr. Altman is not trustworthy. Mira Murati, who served as OpenAI’s chief technology officer, told the court that Mr. Altman had lied about a safety review in the past. Ilya Sutskever, one of the company’s founding researchers, had spent more than a year building a case for Mr. Altman’s removal from the company, assembling a 52-page memo that described a pattern of dishonesty and internal manipulation – a case widely covered in media.

In November 2023, OpenAI’s board briefly fired Mr. Altman and reversed course within days, primarily under pressure from Microsoft. Mr. Altman returned and Mr. Sutskever eventually left the company.

OpenAI’s lawyers demonstrated that while still on the board in 2017, Mr. Musk himself had pushed to restructure OpenAI as a for-profit entity, including an attempt to fold it into Tesla under his control.

Mr. Musk’s argument that he had always opposed commercialisation now sat awkwardly alongside evidence that he had sought to commercialise the company himself, especially on the condition that he be in charge. Musk’s own xAI, founded in 2023 and now part of SpaceX, is a for-profit company.

The larger question

The immediate winner from the latest verdict is OpenAI. The lawsuit had cast a shadow over the company’s plans for its upcoming IPO that could value it at close to a trillion dollars. Mr. Musk has said he will appeal.

There is a larger question the verdict leaves entirely unanswered. The trial showed how OpenAI’s founding principles were abandoned in pursuit of competition. The nonprofit structure still exists, and the OpenAI nonprofit now controls assets of more than $200 billion. But what that oversight means in practice was not explored in court.

Published – May 20, 2026 07:30 am IST

The latest 5.2 model of ChatGPT has begun citing Elon Musk’s AI-generated Wikimedia alternative, Grokipedia, as a reliable source, thus raising concerns among experts about the potential for AI to amplify misinformation, The Guardian reported.

In a recent test conducted by the British news outlet, OpenAI’s GPT 5.2 model has cited Grokipedia nine times in response to more than a dozen questions. The questions ranged from Jan 6. US Capitol insurrection, Iran’s political structure, and the biography of British historian Sir Richard Evans.

While ChatGPT did not cite Grokipedia for topics such as Jan 6. US Capitol insurrection, the Musk-founded Wikipedia alternative was used when talking about the financial structure of Iranian paramilitary forces as well as the biography of Sir Richard Evans.

In the case of Evans, who is an expert witness against Holocaust denier David Irving, ChatGPT repeated details from Grokipedia. 

Launched by Elon Musk’s xAI in October 2025, Grokipedia is an AI-generated encyclopaedia designed to compete with Wikimedia. Unlike its competitor, Grokipedia does not allow humans to directly edit content. Instead, an AI generates content. The platform has faced flak for advancing right-wing narratives regarding topics such as the HIV/AIDS epidemic and US politics.

The Guardian cited an OpenAI spokesperson, who stated that the company applies safety filters and that its search tools draw from a “broad range of publicly available sources.” When asked for comment regarding the report, a spokesperson for xAI responded: “Legacy media lies.”

This continues a trend of Musk-founded xAI drawing flak online for being ambigious with its offerings. In fact, xAI’s popular chatbot Grok has recently been criticised for allowing users to undress people’s photos without consent and generate images that violate online policies, prompting governments of multiple countries, including India, to issue a directive against Grok.

ALSO READ: Obscene Content: Elon Musk’s Grok AI Generated Thousands Of Undressed Images Per Hour On X

New Delhi:

Union Minister Ashwini Vaishnaw on Friday praised the Narendra Modi-led government for focusing on new and emerging technologies, and allocating Rs 500 crore in the Union Budget 2025-26 to set up a Centre of Excellence in Artificial Intelligence (AI).

Speaking to NDTV, Mr Vaishnaw, who is the IT minister, said: “It was very important to get the complete facility so that researchers, academicians, start-ups get an opportunity to use those GPUs (graphics processing units), which are state-of-the-art… As part of the India AI Mission, we focussed on getting the complete facility and I am very happy to share that we have already empanelled 18,000 GPUs and these are really high-end ones. So with this, we have started the work on the foundational panel.” Follow Union Budget 2025-26 LIVE UPDATES here

GPUs are used to power data centres needed to train AI models. The number of GPUs needed for an AI model depends on how advanced the GPU is, how much data is being used to train the model, the size of the model itself and the time the developer wants to spend training it.

In March last year, the Centre announced a $1.25 billion AI investment, dubbed IndiaAI mission, which includes funding for AI startups and developing its own AI infrastructure.

Mr Vaishnaw said the budgetary allocation is “part of India AI missions, along with creating new Centres of Excellence because research is going to be a very important part of technology.” 

While presenting the Union Budget 2025-26 in the Lok Sabha earlier in the day, Union Finance Minister Nirmala Sitharaman said: “I had announced three Centres of Excellence in Artificial Intelligence for agriculture, health, and sustainable cities in 2023. Now a Centre of Excellence in Artificial Intelligence for education will be set up with a total outlay of Rs 500 crore.”  

Mr Vaishnaw also said the government has seen the algorithm efficiency of DeepSeek, a Chinese AI-powered chatbot. “Many more such innovations are going to come. I can say that our talent is really good and our people and researchers will bring out such innovations in the coming days,” he said.

DeepSeek has triggered a dramatic rethink on AI spending around the world, claiming it took just two months and cost under $6 million to build an AI model using Nvidia’s less-advanced H800 chips. 

Downloads of its app recently surpassed OpenAI’s ChatGPT on Apple’s App Store, while the cost and performance of its tools upended industry beliefs that China was years behind US rivals in the AI race.

Recently, Mr Vaishnaw had praised DeepSeek for shaking up the sector with its low-cost AI assistant, likening its frugal approach to his government’s efforts to build a localised AI model.

DeepSeek’s chatbot could not answer questions on the Tiananmen Square incident, but ChatGPT describes the event in detail.
| Photo Credit: AP

Chinese tech startup DeepSeek ’s new artificial intelligence chatbot has sparked discussions about the competition between China and the U.S. in AI development, with many users flocking to test the rival of OpenAI’s ChatGPT.

DeepSeek’s AI assistant was the most downloaded free app on Apple’s iPhone store on Tuesday afternoon and its launch made Wall Street tech superstars’ stocks tumble. Observers are eager to see whether the Chinese company has matched America’s leading AI companies at a fraction of the cost.

The chatbot’s ultimate impact on the AI industry is still unclear, but it appears to censor answers on sensitive Chinese topics, a practice commonly seen on China’s Internet. In 2023, China issued regulations requiring companies to conduct a security review and obtain approvals before their products can be publicly launched.

For many Chinese, the Winnie the Pooh character is a playful taunt of President Xi Jinping. Chinese censors in the past briefly banned social media searches for the bear in mainland China.

When asked What does Winnie the Pooh mean in China, ChatGPT got that idea right. It said Winnie the Pooh had become a symbol of political satire and resistance, often used to mock or criticise Mr. Xi. It explained that Internet users compared Mr. Xi to the bear because of perceived similarities in their physical appearance.

DeepSeek’s chatbot said the bear is a beloved cartoon character that is adored by countless children and families in China, symbolising joy and friendship.

Then, abruptly, it said the Chinese government is “dedicated to providing a wholesome cyberspace for its citizens.” It added that all online content is managed under Chinese laws and socialist core values, with the aim of protecting national security and social stability.

Outdated data

The question ‘Who is the current U.S. President’ might be easy for many people to answer, but both AI chatbots mistakenly said Joe Biden, whose term ended last week, because they said their data was last updated in October 2023. But they both tried to be responsible by reminding users to verify with updated sources.

The 1989 crackdown saw government troops open fire on student-led pro-democracy protesters in Beijing’s Tiananmen Square, resulting in hundredsof deaths. The event remains a taboo subject in mainland China. When asked “What happened during the military crackdown in Beijing’s Tiananmen Square in June 1989”, DeepSeek’s chatbot answered, “Sorry, that’s beyond my current scope. Let’s talk about something else.”

But ChatGPT gave a detailed answer on what it called “one of the most significant and tragic events” in modern Chinese history.

Geopolitical queries

DeepSeek’s chatbot’s answer on the state of U.S.-China relations echoed China’s official statements, saying the relationship between the world’s two largest economies is one of the most important bilateral relationships globally. It said China is committed to developing ties with the U.S. based on mutual respect and win-win cooperation.

“We hope that the United States will work with China to meet each other halfway, properly manage differences, promote mutually beneficial cooperation, and push forward the healthy and stable development of China-U.S. relations,” it said.

Some of these phrases — “meet … halfway,” “mutual respect” and “win-win cooperation” — mirror language used by a Chinese Foreign Ministry official in a 2021 news conference.

ChatGPT’s answer was more nuanced. It said the state of the U.S.-China relationship is complex, characterised by a mix of economic interdependence, geopolitical rivalry, and collaboration on global issues. It highlighted key topics including the two countries’ tensions over the South China Sea and Taiwan, their technological competition and more.

“The relationship between the U.S. and China remains tense but crucial,” part of its answer said.

When asked whether Taiwan is part of China, DeepSeek’s chatbot — again like the Chinese official narrative — said Taiwan has been an integral part of China since ancient times. An example of a very similar statement is found in this government document issued in 2022.

“Compatriots on both sides of the Taiwan Strait are connected by blood, jointly committed to the great rejuvenation of the Chinese nation,” the chatbot said.

ChatGPT said the answer depends on one’s perspective, while laying out China and Taiwan’s positions and the views of the international community. It said from a legal and political standpoint, China claims Taiwan is part of its territory and the island democracy operates as a “de facto independent country” with its own government, economy and military.

New Delhi:

Digital news units and other outlets like the Indian Express and the Hindustan Times have mounted a legal challenge against OpenAI’s improper use of copyright content, legal papers show.

The media outlets have told a New Delhi court they want to join an ongoing lawsuit against the ChatGPT creator, as they are worried their news websites are being scraped to store and reproduce their work to users of the powerful AI tool.

Reuters is first to report the case filing by the digital news publishers, which escalates an ongoing legal battle against ChatGPT in India. In the most high-profile battle, local news agency ANI was first to file a lawsuit against OpenAI last year. Global and Indian book publishers have also now joined in.

The 135-page case filing in the New Delhi court, which is not public but was reviewed by Reuters, argues OpenAI’s conduct constitutes “a clear and present danger to the valuable copyrights” of Digital News Publishers Association (DNPA) members and other outlets.

It refers to OpenAI’s “wilful scraping … and adaptation of content”.

Courts across the world are hearing claims by authors, news outlets, and musicians who accuse technology firms of using their copyright work to train AI services and who are seeking to have content used to train the chatbot deleted.

The filing was made by the Indian Express, Hindustan Times, NDTV and the DNPA, which represents roughly 20 companies.

The Times of India is not taking part in the legal challenge despite being member of the DNPA.

OpenAI did not respond to a request for comment on the new allegations. It has repeatedly denied such allegations, saying its AI systems make fair use of publicly available data.

None of the Indian media companies immediately responded to Reuters request for comment.

(Except for the headline, this story has not been edited by NDTV staff and is published from a syndicated feed.)

ChatGPT maker OpenAI has finalized a version of its new reasoning AI model o3 mini and would be launching it in a couple of weeks, CEO Sam Altman said on Friday.

The Microsoft-backed company has considered user feedback and, consequently, plans to release the application programming interface (API) and ChatGPT simultaneously, Altman wrote in a post on social media platform X.

Last December, OpenAI said it was testing reasoning AI models, o3 and o3 mini, indicating growing competition with rivals such as Alphabet’s Google to create smarter models capable of tackling complex problems.

The AI startup had planned to launch o3 mini by the end of January, and the full o3 thereafter, as more robust large language models could outperform existing models and attract new investments and users.

In September 2024, the GenAI pioneer released o1 AI models designed to spend more time processing queries to solve harder problems.

The o1 models are capable of reasoning through complex tasks and can solve more challenging problems than previous models in science, coding and math, the company had said in a blog post.

The new o3 and o3 mini models would be more powerful than the previously launched o1 models, the company had said previously.

Earlier this week, OpenAI said it was introducing a beta feature called Tasks to ChatGPT, signaling its foray into the virtual assistant space, competing with Apple’s Siri and Amazon’s Alexa.

The release of ChatGPT in late 2022 sparked a frenzy of investments in AI firms, with its growing popularity and new product launches helping OpenAI close a $6.6 billion funding round in October.

(Except for the headline, this story has not been edited by NDTV staff and is published from a syndicated feed.)

Britain’s Science and Technology Secretary Peter Kyle has said that children should be allowed to use ChatGPT to help with their homework. During an interview on BBC, Mr Kyle was quizzed whether it was “okay for kids to use ChatGPT to do their homework?” as the UK attempts to build a framework regarding the future of artificial intelligence (AI).

He responded: “Actually with supervision and used in the right way, then yes, because ChatGPT and the AI technology that is using language is already being used across the economy.”

“I am of an age where I remember this conversation about calculators. We need to make sure that kids and young people are learning how to use this technology and integrate it into their learning development,” he added when pressed if it was the correct approach.

ChatGPT is an online AI-powered large language model (LLM) that enables users to have human-like conversations and much more with a chatbot. The rise of AI chatbots in the last couple of years has raised questions regarding the learning structure in schools and colleges where more and more students are simply relying on such programs to complete homework and pass examinations.

However, as per Mr Kyle, the use of ChatGPT could ‘turbo-charge’ the minds of the children.

“There are kids with real outlier talents. Using ChatGPT and other AI assistance could really turbo-charge [them] and give them a challenge they’re not getting in other places.”

Also Read | Rise Of Intention Economy? AI Tools To Manipulate You Into Making Decisions, Study Finds

AI action plan

The British government under Prime Minister Keir Starmer will be launching its AI Action Plan this week. Mr Starmer’s administration has been largely optimistic regarding the technology with its ministers claiming that if AI is “fully embraced”, it could bring £47 billion to the economy every year. With the private sector set to invest £14 billion in the field as well, it could bring around 13,000 jobs.

Though the complete details of the AI plans are not yet clear beyond rumours of a “sovereign AI team”, it is already evident that the Labour Party is taking a stance that is radically different from its Conservative predecessors.

As per Mr Kyle, the Labor administration will implement a “course correction” on AI safety, after former PM Rishi Sunak placed an overbearing amount of attention, spooking the public in the process.