OpenAI’s “PhD-level” AI claim is false, says DeepMind CEO

DeepMind CEO pushes back on PhD‑level AI claims

DeepMind CEO Demis Hassabis isn’t buying the hype around “PhD-level” AI. Speaking at the All-In Summit, he called claims that current AI models match human expert intelligence “nonsense.”

Hassabis argues that while AI can perform some high-level tasks, it still falls short of general reasoning and creative problem-solving like real human scientists.

Even the most advanced chatbots still make basic mistakes. Hassabis warns that we’re vastly overestimating AI’s current abilities, especially claims coming from competitors like OpenAI.

GPT-5 touted as a PhD expert

OpenAI has been marketing GPT-5 as a breakthrough, claiming it can act like a PhD-level expert in any area. CEO Sam Altman said the model now handles complex queries better than its predecessors and is “legitimate PhD-level” in knowledge.

Hassabis pushes back, pointing out that strong benchmark scores don’t equal true expertise. Tests may make GPT-5 look brilliant, but outside controlled settings, it still struggles to reason or connect ideas the way real scientists do.

AI still struggles with basics

Despite advanced training, today’s AI chatbots regularly hallucinate or make simple errors. Hassabis notes even high school math can trip them up, showing a clear gap between human reasoning and AI output.

According to experts, next-word prediction alone cannot replace reasoning. Hassabis emphasizes that a true artificial general intelligence (AGI) should not fail at these basic tasks, and current models still have a long way to go before approaching expert-level understanding.

What’s really holding AGI back

Hassabis explained that the missing ingredient isn’t more data or faster chips, but true reasoning skills. Unlike humans, who can link ideas from different fields, today’s AI still struggles to carry insights across domains. That gap keeps machines from moving beyond narrow expertise.

He suggested real breakthroughs will be needed before progress speeds up. In his view, it could take another five to ten years before researchers get close to building the kind of general intelligence some companies claim is already here.

The limits of current models

Even the most advanced systems, such as GPT‑5 or other frontier models, still hallucinate. These “hallucinations” can slip into answers as confidently as real facts, reminding users that reliability isn’t guaranteed. For anyone using AI in serious settings, that’s a major concern.

Hassabis explained that the flaw is built into how large language models work. They’re trained to predict words, not to reason, which is why they can sound convincing while producing nonsense. Until that changes, true problem-solving will remain out of reach.

Hype vs reality in AI claims

Hassabis argued that marketing often cherry-picks the best AI results and presents them as proof of human-level ability. He warned that this creates a false picture, because success in one narrow task doesn’t mean broad intelligence. The gap between flashy demos and everyday reliability remains wide.

For users and investors, that hype can be misleading. Believing the buzz too quickly risks disappointment when the tech doesn’t deliver. Hassabis called for more honesty about what AI can actually do right now, not what it might achieve years from now.

Hallucinations are more than mistakes

Hallucinations aren’t just small errors; they can change the meaning of AI’s answers entirely. Even confident-sounding responses from GPT-5 or Gemini may be completely wrong, which makes relying on AI for critical decisions risky.

Hassabis highlighted that fixing these issues is a priority before anyone can claim AI truly thinks like a human. Until that happens, labels like “PhD-level” are premature and misleading.

Reasoning vs next-word prediction

AI models are trained to predict text sequences, not solve problems creatively. Some AI researchers argue that next‑word prediction is not the same as reasoning, limiting real-world applicability.

This training method means AI can appear knowledgeable without actually understanding concepts. True AGI would need reasoning skills that allow transferring knowledge between subjects and solving novel problems, something current models cannot do.

A reality check for investors

Hassabis’s comments serve as a caution to investors and the public. Overhyping AI capabilities risks misplaced trust and unrealistic expectations about the pace of scientific breakthroughs.

True AGI is likely years away, requiring major advances not yet on the horizon. He stressed that true AGI is still years away and will require major breakthroughs.

The takeaway: investors and users should stay realistic and watch actual progress, not marketing spin.

Competition fuels marketing hype

OpenAI’s bold PhD claim contrasts with DeepMind’s cautious approach. Hassabis noted that rivalry between AI companies can push them to overstate what their models can do.

Bold claims grab attention, but they can also distract from real research and slow the adoption of genuinely useful tools. Companies that exaggerate risk lose trust with users, investors, and regulators alike.

Why general reasoning matters

AGI should not only store knowledge but also reason across disciplines. Hassabis points out that spotting patterns from one field and applying them elsewhere is key to human-level intelligence.

Current AI still struggles with this transfer of knowledge. Without the ability to connect ideas creatively, AI remains a powerful tool for automation but far from a replacement for human researchers.

Why transparency matters in AI

Even top AI labs can’t fully explain why models generate certain answers. Lack of transparency makes it hard for users to trust AI in critical tasks. When a system hallucinates or produces biased outputs, there’s often no clear way to trace why.

This gap isn’t just technical; it affects adoption and credibility. If AI is going to be relied on for research, healthcare, or law, users need ways to verify and audit its decisions, not just take flashy claims at face value.

Experts weigh in on limits

Other AI scholars agree that large language models excel at predicting text but fall short at reasoning. Hassabis’ critique reinforces that claims of human-expert equivalence are premature.

The gap between next-word prediction and true cognitive reasoning remains wide. Researchers caution against assuming AI can perform at human expert levels, even if it passes select benchmarks.

Why the hype is risky

Overstating AI abilities can mislead users and policymakers. Hassabis highlights that inflated claims may shape decisions before the technology is ready, potentially causing real-world consequences.

Responsible communication of AI capabilities is essential. Accurate messaging ensures users, regulators, and investors make informed choices instead of reacting to exaggerated claims.

Godfather of AI thinks the same

AI pioneer Geoffrey Hinton said that claims of “superintelligence arriving any day” are wildly exaggerated. He said GPT-5 might even be a small step backward and laughed at the hype. According to him, predicting when true AGI will arrive is impossible.

His best guess? Somewhere between five and 20 years, showing just how uncertain the path to advanced AI really is.

Hinton’s caution highlights that even with fast-moving breakthroughs, the timeline for human-level intelligence remains unclear, reminding investors, researchers, and enthusiasts not to buy into exaggerated claims.

Path to true AGI

Hassabis predicts five to ten years, and one or two key breakthroughs may be needed before AGI emerges. Current AI systems are steps along the path but far from replacing human researchers.

Realistic timelines matter. Researchers must manage expectations and clearly communicate the remaining challenges to avoid hype-driven misunderstandings about AI’s capabilities.

Can DeepMind really make AGI safe? See why experts are skeptical about DeepMind’s AGI safety paper and what it means for the future of AI.

Lessons for AI enthusiasts

Impressive benchmarks do not mean AI equals human intellect. Even state-of-the-art models hallucinate or fail simple tasks. Enthusiasts should be cautious about extreme marketing claims.

Patience and careful evaluation of actual capabilities will pay off as AI evolves. Observing real progress rather than being swayed by flashy claims is the best way to follow the field responsibly.

Want to know what DeepMind’s chief really thinks about Meta’s struggles? Check out what DeepMind CEO Demis Hassabis says Meta is behind in AI and scrambling to catch up.

Do you think AI hype is getting out of hand, or is this part of normal tech competition? Share your thoughts in the comments, and hit like if you follow the AI debates closely.

Read More From This Brand:

• DeepMind AI Now Solves Complex Math Science Problems
• Google Search Gets an AI Makeover
• Google AI Now Understands Images Like Never Before

Don’t forget to follow us for more exclusive content right here on MSN.

If you like this story, you’ll LOVE our Free email newsletter. Join today and be the first to receive stories like these.

This slideshow was made with AI assistance and human editing.