The study, published in the journal AI & Society by Yasmin A. Carruthers and Johannes M. Heim, compared podcasts generated by NotebookLM with naturally produced human podcasts on the same topics, examining two hallmarks of human conversation: the timing of turn-taking and the way speakers adopt contextually appropriate conversational roles. The researchers argue that naturalness in dialogue requires far more than getting the timings right; it demands a contextually appropriate, addressee-oriented approach that responds to how a conversation actually unfolds.

Human turn-taking is a marvel of predictive coordination. Decades of research, beginning with the landmark work of Sacks, Schegloff and Jefferson in 1974, have established that conversation is governed by an organized system in which only one person typically speaks at a time, transitions are smooth, and gaps or overlaps are minimized. Listeners exploit morphosyntactic, semantic and prosodic cues to anticipate when a speaker’s turn will end, preparing their responses while the previous speaker is still talking. Average turn transitions fall in the range of roughly 200 to 250 milliseconds, far shorter than the approximately 600 milliseconds humans need to plan and produce a spoken response. This means that human interlocutors must be predicting, not merely reacting, and traces of this ability appear remarkably early in development, with infants showing sensitivity to conversational turn structure from birth.

Yet the “no-gap, no-overlap” principle is not absolute. Subsequent research has shown that overlaps can account for a substantial share of speaker transitions, and that brief overlaps often signal engagement rather than disruption. Backchannels, the brief vocal acknowledgments such as “mm-hm” or “right” that listeners produce without claiming the floor, add a dynamic and expressive quality to human dialogue. In podcast settings, hosts and guests also follow genre-specific social scripts: hosts introduce topics and then step back into a novice role to give guests a platform, while guests predominantly adopt expert stances. These role expectations, rooted in the broadcasting heritage of podcasting, make conversational behavior highly predictable despite frequent shifts between expert and novice positions.

Against this backdrop, the researchers annotated approximately ten-minute excerpts from two human podcasts and two NotebookLM-generated podcasts built from their transcripts, using the acoustic analysis software Praat. They coded speaker contributions, role shifts, turn durations, overlaps and transitions, achieving very high inter-annotator agreement. The human podcasts came from a series on productivity, featuring a recurring host and invited guests, while the AI versions were generated by NotebookLM using Gemini-based script writing and Google’s own text-to-speech voices, with no additional training data supplied.

On the surface, the AI podcasts behaved similarly to their human counterparts: questions structured the conversation, topics were expanded through follow-ups, and both hosts and guests provided feedback through backchannels. A statistical model confirmed that guests, whether human or artificial, produced significantly more expert contributions than hosts, in line with the conventions of the podcast genre. But the deeper measurements told a different story. Human turns were dramatically longer, averaging 22.3 seconds for hosts and 49.4 seconds for guests, compared with just 11.3 and 14.8 seconds for the AI agents. Human distributions of turn length were skewed toward long, developed contributions, while AI turns were short and symmetrically distributed.

The most striking divergence appeared in overlapping speech. Human guests overlapped for an average of more than four seconds and human hosts for over two seconds, whereas AI overlaps lasted mere milliseconds. AI backchannels rarely overlapped at all; instead, the current speaker paused, waited for the backchannel to be uttered, and then resumed, producing a mechanical politeness that no human interlocutor would display. The researchers interpret human overlaps as signals of engagement and emotional alignment, made possible by the fact that human listeners continuously predict where turns will end. NotebookLM, by contrast, sidesteps the problem of endpoint prediction entirely by pre-planning the entire dialogue script before converting it to speech, strictly enforcing the avoidance of gaps and overlaps in a way that humans never do.

The study also tested a newer NotebookLM feature that allows a human to join the AI conversation live, and the results were revealing. Every human intervention disrupted the pre-planned script: transitions slowed dramatically, overlaps vanished, audio glitches caused sudden voice switches, and the agents could only acknowledge the human’s contribution after long, silence-based delays before returning to their prepared narrative. Because spontaneous intervention requires resource-intensive, real-time context integration that pre-planning was designed to avoid, the hybrid dialogues lost the fluidity that made the fully scripted versions sound convincing. The authors conclude that pre-determining role behavior and turn-taking only goes so far, and that NotebookLM clearly hits its limits when responding to unpredictable input.

Taken together, the findings suggest that AI-generated dialogue can mimic individual features of natural conversation in isolation but fails to synthesize them into the dynamic negotiation of information and social relation that defines human talk. The AI agents spoke shorter, overlapped far less, and behaved more like two equally informed participants than like a host giving a guest a platform, breaking with genre conventions that listeners intuitively expect. Human conversation, the authors argue, is not optimized for seamless information exchange but for building common ground, an emergent and collaborative process shaped by mutual attention, verbal feedback and shared, growing context. Variable turn-taking timings and adaptive conversational role behavior remain unmatched human skills, and achieving genuine naturalness will require AI systems that can interpret social contexts and respond dynamically to the fluid negotiation of interlocutor relations, not merely imitate their surface elements.

Subject of Research: Turn-taking timings and conversational role behavior in AI-generated podcast dialogues compared with natural human conversation

Article Title: Not your average podcaster: turn-taking timings and conversational role behavior in AI-generated dialogues

Article References: Carruthers, Y. A., & Heim, J. M. (2026). Not your average podcaster: turn-taking timings and conversational role behavior in AI-generated dialogues. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03365-3

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03365-3

Keywords: AI-generated podcasts, NotebookLM, turn-taking, conversational roles, dialogue generation, backchannels, overlaps, knowledge asymmetry, common ground, pragmatics, human-machine interaction, speech technology

Single-modal sensors, which detect one quantity at a time, are now mature enough to be printed, woven and laminated onto skin-like substrates. But the human body and its environment rarely deliver stimuli one at a time. A wearable patch pressed against sweating skin experiences mechanical deformation, a rise in humidity and a shift in temperature simultaneously, and its output is a convolution of all three. Multimodal intelligent sensors are designed to capture this multidimensional information in a single device, yet the review emphasizes that their performance remains constrained by overlapping material response windows, signal crosstalk caused by structural coupling, and the sheer difficulty of decoding the resulting complex datasets. Two channels on the same chip can end up answering each other’s questions.

To organize the field, the authors adopt a structured narrative-review framework built on four pillars: advanced material design, perception-decoupling strategies, artificial intelligence-driven data analysis and system deployment. The first pillar surveys the major transduction pathways used to convert physical and chemical stimuli into electrical signals, covering pressure, strain, temperature, humidity, gas and biochemical sensing. Each pathway carries its own trade-offs between sensitivity, response time, dynamic range and mechanical compliance, and the choice of transduction mechanism largely determines how well a device can later separate one stimulus from another.

The materials themselves form the second pillar, and the review catalogs a strikingly diverse toolbox. Hydrogels, with their tissue-like softness, ionic conductivity and controllable swelling, offer a natural route to humidity and strain sensitivity while remaining comfortable against skin. Carbon-based composites and two-dimensional materials bring exceptional electrical conductivity, large surface areas and tunable band structures, enabling highly sensitive resistive and capacitive readouts at low power. Janus heterogeneous structures, which combine two chemically distinct faces in a single particle or film, exploit asymmetric responses so that one side reacts to one stimulus class while the other responds preferentially to a different one. Biomimetic textiles weave these functions into fabrics, embedding sensing into clothing that people actually want to wear. Across all of these platforms, the unifying design goals are flexibility, conductivity, interfacial regulation and multifunctional integration, ensuring that adding modalities does not destroy the mechanical comfort or durability that makes wearables viable in the first place.

The technical heart of the review is its deep treatment of decoupling strategies, the engineering answers to crosstalk. Structural spatial decoupling physically separates sensing elements so that each stimulus interacts primarily with its designated channel, an approach that is intuitive but costly in device area and packaging complexity. Orthogonal responses of functional materials take a more elegant route: materials are selected or engineered so that their response matrices to different stimuli are as linearly independent as possible, meaning temperature changes one output in a pattern that pressure cannot mimic. Microstructure and interface engineering tunes porosity, surface chemistry and layered architectures to sharpen selectivity at the material level. The construction of independent signal channels gives each modality its own electrical pathway, reducing parasitic coupling, while feature-representation-based assisted unmixing pushes part of the separation task into the software, using learned feature spaces to statistically disentangle mixed signals that no passive design can fully separate.

Read together, these strategies describe a layered defense against ambiguity. The review makes clear that no single technique achieves low-crosstalk, high-fidelity perception under genuinely concurrent multi-stimulus conditions; instead, state-of-the-art devices combine spatial layout, orthogonal material chemistry and algorithmic unmixing, assigning each layer the portion of the separation problem it handles best. This cross-layer view is one of the paper’s most useful contributions, offering practical design guidelines for material selection, structural engineering and algorithm configuration rather than treating each as an isolated discipline.

Artificial intelligence completes the loop. The review examines how data-driven models now underpin multimodal signal recognition, fusion-based decision-making and scenario understanding. Machine learning models trained on labeled multimodal datasets can learn the characteristic signatures of, for example, a pulse waveform riding on a temperature drift, and can classify complex gestures or physiological states that no single channel could distinguish. Fusion-based approaches combine evidence across modalities to make decisions that are more robust than any individual sensor’s verdict, while scenario-understanding models move the system from raw perception toward contextual interpretation, such as recognizing that a subject is exercising rather than feverish. This is where the title’s promise of intelligence becomes literal: the sensing hardware provides rich but ambiguous data, and the learning algorithms provide the decoding machinery.

The applications surveyed span wearable health monitoring, electronic skins and human–machine interaction. In health monitoring, multimodal patches can track pulse, respiration, skin temperature and humidity together, offering clinicians a multidimensional physiological picture rather than isolated vital signs. Electronic skins for prosthetics and robots must distinguish a hot object from a heavy one at a glance, a task that demands exactly the decoupling and fusion capabilities the review describes. In human–machine interfaces, gesture recognition and tactile feedback both depend on reliably separating intentional mechanical signals from environmental noise. The authors frame the field’s trajectory as an evolution from device-level sensing toward system-level cognition, in which the sensor, the material, the algorithm and the application form a single designed pipeline.

The review does not shy away from the field’s bottlenecks. It identifies material stability, decoupling capability, sensing accuracy, long-term reliability, system integration and scalable manufacturing as the major obstacles between current laboratory prototypes and deployed products. Hydrogels dry out; two-dimensional films delaminate; machine learning models degrade when training data fail to match real-world distributions; and fabrication processes that work at wafer scale rarely translate to roll-to-roll textile production. Looking forward, the authors call for next-generation multimodal sensing platforms featuring high selectivity, self-powering capability, manufacturability, interpretability and edge-intelligence-enabled collaboration, so that sensing and computation can be distributed across networks of low-power devices rather than concentrated in the cloud.

What emerges from this synthesis is a coherent blueprint for a technology that could quietly become as ubiquitous as the smartphone camera. A flexible sticker that simultaneously reads touch, temperature, sweat chemistry and ambient gas composition, decodes the mixture on a nearby edge processor and reports a meaningful health state is no longer a speculative concept but an engineering program with named materials, named algorithms and named hurdles. By binding materials science, device physics and machine learning into one analytical framework, the review offers researchers a shared vocabulary for the interdisciplinary work ahead, and offers the rest of us a preview of electronics that will not just touch the world but genuinely understand it.

Subject of Research: Development of AI-empowered multimodal intelligent sensors using advanced functional materials and signal decoupling strategies for wearable and robotic applications

Article Title: AI-empowered multimodal intelligent sensors integrated with decoupling and advanced functional materials

Article References: AI-empowered multimodal intelligent sensors integrated with decoupling and advanced functional materials. (n.d.). https://doi.org/10.1007/s42114-026-02067-0

Image Credits: AI Generated

DOI: 10.1007/s42114-026-02067-0

Keywords: multimodal sensors, flexible electronics, wearable technology, electronic skin, hydrogels, signal decoupling, machine learning, human-machine interaction, health monitoring, two-dimensional materials, biomimetic textiles, edge intelligence