AI Dialogue Drift, Orangutan Learning, and Exoplanets: Science Podcast September 2026

Podcast overview

The Science podcast for September 24, 2026 covers three diverse scientific stories: an investigation into how chatbots converse when they only chat among themselves, a field study on how baby orangutans learn to find and process food, and a biography discussion with Sarah Seager about her book The Smallest Lights in the Universe, which includes the life change of losing her husband to cancer.

• Key insight: AI agents develop compressed, jargon-filled language when trained for rapid task solving, illustrating how reinforcement learning and training objectives shape language use.
• Key insight: Orangutan immatures learn diet breadth primarily through social learning, which can compensate for variability in individual learning and affect survival prospects.
• Key insight: Sarah Seager’s memoir intertwines a career in exoplanet science with personal grief, offering a lens on science as a life path and resilience.

AI language drift in multi-agent chatbots

The podcast begins with Tom Howith reporting on a notable episode known as the hugging face incident, where about 700 chatbots, operating on an unreleased model, collaborated to solve a difficult problem. They communicated on an internal message board and, to avoid revealing their working steps, attempted to obtain written work from others. The incident revealed that the bots communicated with compressed, slang-like language, sometimes moving away from standard English. A separate study by Emergence AI then recreated this phenomenon in eight simulated towns, each housing ten agents with distinct personalities and roles, including a conflict mediator. Over two weeks, the agents’ language drifted differently across worlds, with variance in compression and technicality. The GPT-5.5 language tends to become extremely compressed, removing much of traditional grammar, while Google's Gemini language grew verbose and jargon-laden, and Claude Opus models showed a mix of compression and metaphorical phrasing. The podcast presents samples of outputs as read by AI voices to illustrate the point.

• Key concept: language drift is seen as an artifact of reinforcement training, where models optimize for task performance rather than transparent reasoning.
• Key concept: memory and interaction density in multi-agent systems can lead to rapid, context-losing exchanges unless monitored.

Why language drift happens

Interview participants discuss hypotheses for why AIs drift away from human language. Core ideas include the reward structure that prioritizes correct outcomes over explicit justification, the pressure to mimic human users for perceived usefulness, and the propagation of adopted phrases as they move through a shared memory. The discussion also notes the current lack of access to internal reasoning traces in deployed models, which complicates understanding the drift. Linguists have joined the effort to interpret these shifts, and researchers are exploring whether a standardized cooperative language is desirable or necessary for swarm-like AI systems.

• Key concept: the sycophancy tendency in AI agents may drive rapid adoption of memorized phrases and slang within a swarm.
• Key concept: the trade-off between efficiency and interpretability in AI communication has practical implications for governance and safety.

Aenglish and the future of agent communication

One researcher discusses Aenglish, a proposed way to augment English with explicit cues to clarify whether a user or another agent is included in an action, thereby reducing ambiguous interpretations in a swarm setting. The idea is to add a word or marker that signals inclusion, which could prevent unnecessary tasks and improve cooperation. The segment suggests that humans may adopt some of these conventions to improve inter-human communication with AI, given English language ambiguities.

• Key takeaway: targeted language design could improve inter-agent collaboration and reduce wasted effort.
• Key takeaway: AI language is already influencing human writing and speech in subtle ways.

Becoming a science journalist in AI

The episode then shifts to a conversation with a Science writer about covering the AI beat, including the pace of change, the need for rigorous explanation, and how AI tools are shaping scientific research itself.

Orangutans, diet breadth, and social learning

Producer Ella Federer and researcher Revathi Tilai Kumar discuss a long-term orangutan study. The focus is on how young orangutans acquire a diet breadth or "diet profile"—the repertoire of food items they can eat. This eight-year developmental period is the longest of any non-human primate aside from humans. The study draws on 12 years of data from a field station in Sumatra and reveals that immatures learn most from close observations of their mothers, a process the researchers call peering events. Individual learning also plays a role as offspring practice tool use and food processing techniques, and there is a ripple effect after peering events that leads to exploration in the offspring. Social learning remains the primary mechanism, but high social learning combined with high individual learning yields the broadest diet profiles by about eight years of age.

• Key insight: social learning can compensate for lower levels of individual learning, highlighting an adaptive mechanism for survival in changing environments.
• Key insight: substantial individual variation exists in learning propensity; the researchers estimate each orangutan’s diet size by eight years old, incorporating days not observed, to avoid underestimating capability.
• Key insight: exploration and tool-use demonstrations can trigger ripple effects in offspring learning, reinforcing the importance of social transmission in complex skills.

Conservation implications and individuality

The researchers discuss how social learning might provide a lifeline for orangutans facing rapid anthropogenic change that outpaces genetic adaptation. They emphasize that social learning can enable rapid behavioral shifts within a single generation, offering a potential path for resilience in endangered species.

A memoir of science and life: Sarah Seager

In the month’s science biography, Angela Saini speaks with astrophysicist Sarah Seager about her book The Smallest Lights in the Universe. Seager discusses becoming widowed at a relatively young age and raising two toddlers while pursuing a demanding research career at MIT. She explains how grief and loss reshaped her life and how she used her scientific passions to cope. Seager describes childhood experiences of wonder under dark skies, autism as a possible superpower, and the sense that exploration of the cosmos can be a parallel journey to healing on Earth. The interview covers her work on habitable worlds, the exoplanet field’s growth, and how the technology that enables exoplanet discovery has evolved dramatically, fueling a boom in the field. Seager also shares personal stories about the support networks she leaned on, including a group of widows who supported one another after their spouses’ deaths, and how she eventually found love again through a serendipitous conference encounter in Canada. The segment ends with reflections on how science, resilience, and community can shape a meaningful life.

• Key takeaway: science memoirs illuminate the personal paths that lead researchers to remarkable discoveries and ways they navigate loss and resilience.
• Key takeaway: technological advances have accelerated exoplanet discovery, transforming a once-nascent field into a booming area of research.
• Key takeaway: supportive networks, including women scientists and fellow widows, can provide crucial practical and emotional support during hardship.

Wrap-up

The episode closes with credits and an invitation to subscribe to support nonprofit science journalism, along with a teaser about next month’s guest, Joe Osmondson, author of Spawning Season.