3 May 2026: Harvard Study Finds AI More Accurate Than Emergency Room Doctors

An AI model outscored two emergency room doctors on diagnostic accuracy in a Harvard study, the Academy locked AI-generated actors out of the Oscars, and the Pentagon signed deals with Nvidia, Microsoft and AWS to put AI on classified networks. Here is what shifted in AI today.

A Harvard Medical School study has found a large language model produced more accurate emergency room diagnoses than two human doctors in a head-to-head test. The research, run with Beth Israel Deaconess Medical Center, fed real ER cases to several large language models and a set of clinicians, then graded their diagnostic reasoning. At least one model, built on OpenAI technology, came out ahead of the human doctors it was compared against, according to TechCrunch’s write-up.

The study did not claim AI should replace clinicians. It tested how models reason through the messy, time-pressured information ER doctors handle every day. The headline will travel further than the caveats, which is why it matters.

For UK readers, the question is what the NHS does with results like this. NHS England has been piloting AI triage tools, and the MHRA has been firming up its software-as-a-medical-device guidance. A peer-reviewed paper showing measurable gains accelerates procurement, not cools it. Read the TechCrunch report.

The Academy of Motion Picture Arts and Sciences has ruled that films featuring AI-generated actors or AI-written scripts cannot be nominated for Oscars. The change closes a debate that has been running ever since the synthetic performer Tilly Norwood was first floated as a potential awards contender. Under the new rule, work where the lead creative contribution is generative AI is ineligible.

The decision is narrower than it sounds. Films that use AI as part of production, for de-aging, dubbing or visual effects, remain eligible if a human creative team is in charge. The Academy is drawing a line between AI as assistant and AI as author, and only the second is ruled out.

For UK studios and Soho VFX houses, the practical effect is limited. Most British awards-track productions already use AI as a tool, not as the artist. The signal matters more than the rule: Hollywood’s premier institution has decided AI authorship is not yet ready to be honoured. Read the full story.

The US Department of Defense has signed deals with Nvidia, Microsoft and AWS to deploy AI on its classified networks. The contracts cover model access, secure compute and data infrastructure inside the Pentagon’s classified environments. They follow a public dispute earlier in the year between the DoD and Anthropic over the usage terms of Anthropic’s models, which left the department determined to widen its supplier base.

The deals are framed as diversification rather than a pivot. The Pentagon wants several frontier providers, not one preferred partner, and it wants the freedom to use them on the most sensitive data. That is meaningful commercial validation for the three vendors named, and a nudge to other labs that defence procurement now expects flexible deployment terms.

For the UK, the read-across is the Ministry of Defence’s AI strategy and the Defence AI Centre’s frontier model work. American buying patterns tend to set the template allied governments follow within twelve months. Read the report.

If you have not yet adopted AI dictation, TechCrunch’s new ranking of the best apps gives you a starting point. Voice tools have moved past simple transcription. The current generation handles email replies, meeting notes and even short coding tasks dictated by voice. The TechCrunch test ranks apps including Wispr Flow, with the top picks reflecting a clear gap between dictation that simply types what you say and dictation that understands context, formatting and intent.

For freelancers and small teams in the UK, the productivity case is direct. Replying to twenty emails by voice is faster than typing them, particularly on a phone, and the better tools now respect tone, punctuation and capitalisation without correction. The category is no longer a novelty.

The trade-off is privacy: most apps process audio in the cloud, so pick a provider you trust with what you say. See the full ranking.

Anthropic is preparing a fundraise that could value the company at more than nine hundred billion dollars, with allocations being asked of investors within forty-eight hours. If the round closes at that level, Anthropic moves from frontier challenger to one of the most valuable private companies in the world, narrowing the gap with OpenAI on paper terms. The speed of the round is striking. Investor demand has clearly outpaced the usual diligence cycle.

The valuation reflects strong commercial uptake of Claude across enterprise customers, plus the scale of capital that frontier AI training now requires. A round of this size is fuel for the next training cluster, not just runway. For UK customers, the read is stability: a well-capitalised Anthropic can keep investing in its UK office and in the safety research that has helped sell the model to financial services. Read the TechCrunch scoop.

Here is everything else worth knowing from today’s AI news.

• Meta acquires humanoid AI startup Assured Robot Intelligence. The deal strengthens Meta’s robotics models as it pushes deeper into physical AI. TechCrunch
• Google’s Gemini assistant is rolling into millions of vehicles. The deployment turns conversational AI into a default in-car feature for several major manufacturers. TechCrunch
• Stripe Link now lets AI agents spend on your behalf. The wallet update creates approval flows for agents to buy goods, services and subscriptions. TechCrunch
• OpenAI partners with Yubico for hardware-key ChatGPT logins. Account takeovers have surged, and Yubikey support is now an opt-in for security-conscious users. TechCrunch
• Elon Musk testified that xAI trained Grok on OpenAI models. The admission lands in the middle of his lawsuit against OpenAI and revives the distillation debate. TechCrunch
• Replit’s Amjad Masad weighed in on the Cursor talks. He told StrictlyVC he would rather not sell, even as rival Cursor is reportedly in acquisition talks with SpaceX at a sixty billion dollar valuation. TechCrunch
• Legal AI startup Legora hits a 5.6 billion dollar valuation. Its head-to-head fight with Harvey is reshaping how UK and US law firms buy AI. TechCrunch
• Apple says AI demand for Macs caught it off guard. Mac mini, Studio and the new Neo will be supply-constrained again next quarter. TechCrunch

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s afternoon update on developments in artificial intelligence, published every weekday afternoon.