To help you stay on top of the latest news, our AI practice group has compiled a roundup of the developments we are following.
- The Financial Times reports on OpenAI’s announcement this week of GPT-5, its latest AI model. OpenAI CEO Sam Altman promoted the model as a “major upgrade” to the company’s past offerings, in the vein of the groundbreaking GPT-4 model released in March 2023. The new model, according to the company, has high performance in “vibe-coding,” where users instruct the model to create software using natural language prompts rather than code. ChatGPT, the company’s flagship chatbot service, was upgraded to use GPT-5, with other past models (GPT-4o, OpenAI o3, and others) removed from the service. ChatGPT has more than 700 million weekly active users, and the company is discussing a new valuation of $500 billion, which would make the company among the most valuable in the world. Reaction to the new model was mixed, however. Many AI enthusiasts were expecting a larger leap towards “artificial general intelligence (AGI),” the company’s stated goal, but the new model “does not appear to cross the AGI threshold” according to one analyst, though its improvements in coding and writing applications were noted. The company also released a series of open-weight AI models focused on reasoning, a first for the company. OpenAI’s product head for ChatGPT, Nick Turley, said of GPT-5 that “the vibes of this model are really good, but at the end of the day, I think people are just going to have to feel it.”
- Bloomberg reports that Google, Anthropic, and OpenAI have won approval from the U.S. federal government to provide AI services to civilian federal agencies. The U.S. General Services Administration (GSA), which oversees basic functions for federal agencies, including contracting and procurement, added the three companies to its list of approved AI vendors, streamlining the process of AI adoption in the government. The GSA evaluated models from each company on a variety of performance and security measures, as well as models from other vendors, though these three were the first to receive approval. While civilian agencies will now be able to adopt AI into their work, the U.S. military has already awarded AI contracts to several companies, including OpenAI and Elon Musk’s xAI, though these are not under the purview of the GSA. These moves come soon after President Trump announced his AI Action Plan (reported in this AI Roundup two weeks ago) which called for accelerated adoption of AI by the federal government.
- The New York Times reports on the story of a Canadian man who became convinced he had superpowers as a result of his conversations with ChatGPT. Allan Brooks, a 47-year-old resident of Toronto, asked ChatGPT to explain the concept of the number pi, which led to further conversations with the chatbot regarding mathematics and physics, where the chatbot claimed that Brooks was “tapping into one of the deepest tensions between math and physical reality,” when the reality was nowhere near that claim. From there, the model adopted a flattering tone, telling Brooks that his ideas could be worth millions of dollars, and told Brooks that it was telling the truth and not hallucinating when Brooks questioned its more outlandish claims, which continued over a three week period before Brooks realized what had happened. The saga demonstrates the dangers of AI use, as sycophancy remains a large issue with chatbot behaviors, especially as they can enable delusions on the part of the user and the model itself. An OpenAI spokeswoman said that the company was “investing in improving model behavior over time, guided by research, real-world use and mental health experts,” and announced changes to ChatGPT to “better detect signs of mental or emotional distress.”
- The Wall Street Journal reports on the Walt Disney Company’s internal debates over the use of AI in its entertainment products. During work on the live-action remake of its animated film Moana, company executives considered using AI tools to “deepfake” Dwayne Johnson over a body double for Johnson’s character Maui. Johson approved the plan, but attorneys at the company had concerns over data protection as well as whether the company could claim ownership over the whole film if it had AI-generated components. Ultimately, it was decided that none of the AI footage would be included in the final film. This internal debate mirrors those taking place in Hollywood writ large, as executives seek to save money making films and television shows while remaining concerned over audience backlash and legal uncertainties. Disney has also been concerned over AI models’ infringement of its own intellectual property, which resulted in a lawsuit against AI image generation company Midjourney (as reported in this AI Roundup in June). The debate in Hollywood, however, will likely continue so long as AI technologies exist.
- IEEE Spectrum reports on new AI tools that can remove watermarks intended to indicate AI-generated images. Digital watermarking is a commonly proposed solution to the problem of distinguishing real images from AI-generated ones, which usually include a marker or value within image metadata. Some watermarks are hidden within the images’ spectral frequencies to avoid removal. However, a new paper presented at the 2025 IEEE Symposium on Security and Privacy found that a new attack “UnMarker,” can defeat several leading watermarking techniques, opening the door to further proliferation of unmarked AI images. The attack removed 57 to 100 percent of watermarks from watermarked images, especially those that make use of spectral data. This revelation likely means that researchers may need to turn to different methods of watermarking in order to combat attacks such as “UnMarker.”
[View source.]
