Could AI present CrowdScience?

Episode Summary

Executive Summary: The episode examines whether AI could take all human jobs, using CrowdScience itself as a test case. It explains how large language models work, shows that AI is rapidly improving at specific tasks but still struggles with context, humor, and human interaction, and argues that job replacement is ultimately a social and economic choice, not just a technical one.

Main Topics: What AI is and how large language models work (Priority: 5/5): The hosts explain that modern AI is mostly based on large language models trained on vast text datasets, then refined with human feedback to become usable products like chatbots. How fast AI capability is improving (Priority: 5/5): A tech journalist describes benchmark-style progress, noting that AI can now complete increasingly long tasks with measurable accuracy, though reliability remains limited. Jobs vs tasks: the real question (Priority: 5/5): The discussion shifts from whether AI will replace jobs to which tasks within jobs can be automated, and how employers and societies choose to respond. Could AI present a radio show? (Priority: 4/5): The hosts test whether AI could host CrowdScience, concluding it may handle scripting or some voice work but would struggle with interviewing, physical experience, and humor. Voice cloning and synthetic speech (Priority: 5/5): BBC AI staff demonstrate cloning a presenter’s voice and a linguist explains why synthetic speech can sound convincing yet still miss prosody, emphasis, and contextual cues. BBC policy and practical uses of AI (Priority: 4/5): The BBC says it is not using AI presenters or fully AI-generated content, but is experimenting with synthetic voices for scalable reformatting of existing journalism. Future of personalized audio and ‘liquid content’ (Priority: 4/5): The episode explores a future where listeners could customize length, language, and even voice, raising questions about what audiences value in human-made media.

Key Arguments: AI progress should be measured by tasks, not just jobs; many roles contain a mix of automatable and non-automatable tasks. Current AI can outperform humans on some narrow tasks, but its outputs are often unreliable and require human oversight. The biggest limits on AI are practical: training data, computing power, and the lack of enough high-quality speech data for many languages. Whether AI replaces jobs depends less on technology than on economic and labor choices made by employers, workers, and society. AI can imitate a presenter’s voice convincingly, but it still struggles with prosody, question intonation, emphasis, and contextual meaning. Humor, relationship-building, and real-world experience are especially hard for AI to replicate. The BBC is using AI selectively for efficiency and accessibility, but not to replace human creativity or core journalism. Future media may become highly customizable, allowing listeners to reshape content length, language, and presentation style.

Data Points: Listener age: 27 years old - Po from Taiwan says he worries about whether there will be jobs for him over the next 30 or 40 years. AI task doubling rate: Every 7 months - Alex Hearn says the length of tasks AI can do with 50% accuracy has roughly doubled on this cadence. Current top-flight task length at 50% accuracy: About 2 hours - AI can now do tasks taking a software engineer around two hours with 50% accuracy. Projected task length in 7 months: 4 hours - Based on the cited doubling trend, the same 50% accuracy task length is expected to double. Projected task length in 14 months: 8 hours - The trend implies AI could handle an entire working day’s task length at 50% accuracy. GPT-5 task length at 50% success: About 2.25 hours - Alex gives a more specific example of current capability at a lower reliability threshold. GPT-5 task length at 80% success: 26 minutes - At higher reliability, the task length drops sharply, illustrating the tradeoff between accuracy and complexity. Human comparison: Trained software engineer - The benchmark compares AI performance against professionals rather than average workers. Synthetic speech sample length: About 3.5 minutes - BBC staff used a short sample of the presenter’s voice to clone it. BBC Space podcast mission duration: 186 days - Promotional material for 13 Minutes mentions astronaut Tim Peake’s time aboard the ISS.

Pivotal Quotes: "It seems pretty likely that the task of mine that is proofreading the issue of an economist before it goes to press can be really, really helpfully sped up by an AI system." — Alex Hearn: Used to illustrate that AI is already useful for some professional tasks. "There is this issue of context dependency that we've been discussing." — James Kirby: Explaining why synthetic speech can sound natural but still miss the right emphasis or question intonation. "I think the important thing from a BBC perspective is there is possible and then should we do this?" — Nikki Birch: Clarifying that technical feasibility does not mean the BBC will adopt AI presenters.

Implications: AI is likely to reshape work by automating tasks, not instantly eliminating all jobs. Media and other industries may adopt AI for efficiency and personalization, but human judgment, creativity, and trust will remain valuable.