AI2Day Daily Brief — 4 Aug 2026

Good morning, it is Tuesday the fourth of August, and overnight the conversation about what AI can and cannot do moved from the research lab to the opera house, the battlefield, a Sussex bank account, a Hollywood boardroom, and an Apple research paper. Let us get through all five.

Bayreuth Festival in Germany has been staging Richard Wagner's operas since eighteen seventy-six. Its hundred and fiftieth anniversary production of the Ring of the Nibelungen, sixteen hours of music across four operas, was handed to an AI-assisted creative team led by stage director Marcus Lobbes. The official word was that the production was curated rather than directed. Lobbes fed AI tools prompts drawn from the Ring's themes: power, capitalism, mythology, gender roles. The staging opened this week. Early reviews are not generous. The Guardian called the result banal and dramatically empty. The takeaway for anyone building AI creative tools is blunt: generating plausible visual ideas from a reading list of themes is not the same thing as building dramatic tension across an evening, let alone four.

That is a hard lesson to learn at that scale. Moving from the stage to an active war zone.

Ukraine has been using cheap first-person-view drones to destroy Russian armoured vehicles and helicopters. Those drones are now getting an autonomy upgrade. From mid-July, the Ukrainian military started receiving Shrike attack drones fitted with what the American company Auterion calls the Skynode S strike kit. A human pilot still flies the drone into the general area and designates a target up to roughly half a mile away. After that, the AI system tracks the target and guides the drone without further human input. Ukrainian manufacturer SkyFall and Auterion say they plan to deliver fifty thousand equipped drones over the coming months. Each Shrike costs around four hundred dollars. The combination of low unit cost and autonomous terminal guidance is a significant shift in what small drones can do on a battlefield.

Fifty thousand is a production number that changes the calculus. Next, a story that sits at the intersection of AI fraud and banking response time.

Zoli Rutter, a businessman in Sussex, lost fourteen thousand, two hundred and forty-four pounds from his Metro Bank account to fraudsters who spent every penny buying credits for Claude, the AI chatbot made by the San Francisco company Anthropic. Credits are prepaid tokens that let users send messages to the chatbot. Rutter says he noticed the unauthorised transactions and contacted Metro Bank. The withdrawals continued anyway. The Guardian reported the story and Rutter is now fighting to recover the money. The case does not say anything specific about a flaw in Claude itself. What it does highlight is a question about bank fraud detection: small, repeated purchases of digital credits are not a new fraud pattern, and the question being asked is whether UK banks are catching AI-linked versions of it quickly enough.

And that question about institutional response is unlikely to go away. Now to Hollywood.

Wired published a wide-ranging conversation between its editor Katie Drummond and Matt Belloni, a founding partner at the media newsletter Puck and a long-standing Hollywood reporter. The picture they describe is of an industry already deep into AI use that has not settled on whether to say so publicly. Netflix reported in April this year that three hundred of its shows already use generative AI, the kind that produces new images, video, or text from scratch. Netflix separately paid five hundred and eighty-seven million dollars to acquire Ben Affleck's AI startup for post-production work. The Writers Guild of America permits members to use tools like ChatGPT, as long as a credited human writer remains on every script. Animation is flagged as the most immediately disrupted sector, because AI can complete scenes and generate storyboards without human artists. The central tension Belloni identifies is not whether AI is in the building. It is already in the building. The tension is over who in Hollywood is willing to say so.

Transparency as the actual fault line, not the technology itself. And finally, a piece of foundational research.

Are you at risk? Attackers do not break in any more, they log in, and your people are the way in. Train2Secure teaches your employees to spot the email before they click it, and proves it works with real phishing simulations and compliance-ready reporting. From $1.59 per user, per month. Start free today at Train2Secure dot com. That's Train, the number two, Secure, dot com.

Apple Machine Learning Research published a study this week on why AI models that process both images and text still hallucinate, meaning they describe pictures with confident-sounding details that are simply not there. Researchers call this class of system a multimodal large language model. A text-only model hallucinating means it states a wrong fact. An image model hallucinating can mean describing a red car as blue, inventing a sign on a shop wall, or misreading a person's expression. The Apple study looks at a training technique called preference alignment, which teaches a model to favour accurate answers over plausible-sounding ones. This is well established for text-only AI but much less studied for image-understanding models. The research matters beyond the lab because these models power image captioning, visual search, and accessibility tools that are already in everyday use.

If you want the full context on any of these stories, the longer Monday show AI Today Weekly is where we go deep. That is your AI briefing for today. Every story is at A-I-2-Day dot live. That is A, I, the number two, D-A-Y, dot live. We are back tomorrow morning. If this was useful, hit the thumbs up and subscribe, so the next one finds you.