AI is error-strewn: when over half the summaries get it wrong
A BBC study found most AI news summaries riddled with errors. On hybrid models, the open-source pivot, Arm's gamble, and whether we're outsourcing our minds too fast.
A recent BBC study showed that over half of AI-generated news summaries are riddled with errors — proof these tools are far from the reliable sources many hoped for. Contrary to Valley AI ego-maniac opinion, the data suggests we're still far from true AGI. In fact, it's mostly getting it wrong.
AI models and synthetic media
DeepSeek tweaked R1 to cut compute costs while still performing well, inviting comparisons with OpenAI and others. Anthropic is preparing a 'hybrid' model that can switch between deep reasoning and quick responses, claiming better handling of programming and large-scale code analysis. And in a surprising turn, Baidu announced it will make its next Ernie model open source from 30 June — a clear departure from CEO Robin Li's long-held preference for closed source, apparently driven by pressure from cost-efficient competitors like DeepSeek, and paired with plans to offer Ernie Bot free from 1 April.
Cultural normalisation and legal tech
AI companies are expanding into new regions, signalling cultural normalisation: OpenAI is opening an office in Munich, adding to recent moves in Paris, Brussels and Dublin. On the legal-tech front, SpotDraft — now serving around 400 customers with 169% year-on-year revenue growth — closed a $54M Series B.
Semiconductors re-arm
Arm, long known for licensing chip blueprints, is now gambling on its own custom server CPUs for data centres, with Meta already on board. EnCharge AI, a Princeton spinout focused on analog memory chips, raised over $100M; its chips claim to use 20 times less energy than conventional digital counterparts, which could dramatically lower AI infrastructure costs.
Regulation and energy
A study by Epoch AI recalculated ChatGPT's footprint using GPT-4o and found an average query uses only 0.3 watt-hours — far less than the once-quoted 3 watt-hours based on old hardware assumptions, a timely revelation as environmental concerns push the industry toward greener data centres. Across the Atlantic, Macron unveiled a roughly €109 billion AI investment package aimed at data centres powered by low-carbon nuclear — a model the UK might well emulate.
Media, culture and cognitive impact
YouTube isn't sitting idle either. CEO Neal Mohan named AI one of the company's four 'big bets' for 2025, rolling out tools for creators from auto-dubbing to advanced age-identification tech — even as critics warn an influx of AI-generated content might dilute genuine creativity. Beyond media, there's a growing concern that reliance on AI is dulling our collective critical thinking. A Microsoft and Carnegie Mellon study of 319 professionals found that using generative AI for routine work can sap our ability to think deeply, with many shifting from creative idea-generation to merely verifying AI outputs — a troubling atrophy of problem-solving skills.
I mis-prompted ChatGPT to write a version of this article set in the future. What I wanted was a scheduled data run from recent web searches; what it actually did was confidently fabricate a whole article, dated in the future, with every single topic invented. Done so confidently — but still total nonsense. It makes things up worryingly well. Quis custodiet ipsos custodes? I really hope we don't outsource our collective brain too quickly.