The double empathy problem, first articulated by Damian Milton in 2012, reframes the long-standing assumption that autistic communication difficulties are intrinsic deficits located within the autistic person. Instead, the theory proposes that communication breakdowns arise bidirectionally, from a mutual mismatch in lived experience, expectations, and expressive styles between people with different neurotypes. By that logic, two autistic people, sharing a neurotype, should understand each other more readily than an autistic and a non-autistic person would. This prediction has received striking support from experience-based studies: autistic adults consistently report that interactions with other autistic people feel smoother, easier, and more positive than interactions with non-autistic people, and some observational studies have found that autistic-to-autistic information transfer can be highly effective.

What has been missing, the Ghent team argued, is a direct behavioral measure of communicative efficiency. Feeling understood is not the same as transmitting information quickly and accurately, and prior work has often relied on self-report questionnaires, observer ratings of rapport, or single-neurotype pairings. To fill this gap, the researchers turned to the collaborative referencing task, a classic experimental paradigm rooted in the psycholinguistic work of Herbert Clark and Deborah Wilkes-Gibbs, in which two people must jointly establish shared references for ambiguous stimuli. Efficiency in this paradigm is not a matter of opinion; it is measured in the concrete currency of talk time, turns, and task performance across repeated rounds of interaction.

The experiment involved 62 adults between 20 and 50 years of age, all with an IQ above 80, half of whom had an autism diagnosis. Each participant completed the task twice, once with a partner of the same neurotype and once with a partner of a different neurotype, a within-participant design that allowed the researchers to compare each individual’s performance across partner types while holding their own communication style constant. This yielded 14 dyads of two autistic adults, 14 dyads of two non-autistic adults, and 30 mixed-neurotype dyads. The dyadic structure matters because communication is a joint product; any theory that locates difficulty in one partner alone, or that predicts a match advantage, makes testable predictions at the level of the pair.

The task itself required participants to collaborate on ordering abstract figures, non-representational shapes that cannot be described with a single obvious label. Pairs worked through multiple rounds of the same ordering problem, and the standard signature of the paradigm is a steep learning curve: as partners negotiate and settle on shared descriptions during early rounds, they establish common ground, and later rounds should demand far less verbal effort. Efficiency was tracked across these rounds, capturing how quickly and economically each dyad converged on workable references for the figures. The method draws directly on decades of research showing that referential communication becomes progressively streamlined as interlocutors build a shared conceptual workspace.

The results confirmed the general learning mechanism for every type of pair. All dyads, regardless of neurotype composition, became more efficient over the course of the rounds, demonstrating that autistic adults, like non-autistic adults, build common ground and refine their collaborative descriptions with repeated interaction. But the critical hypothesis failed. Contrary to the prediction derived from the double empathy problem, same-neurotype dyads were not more efficient communicators than mixed dyads. Autistic pairs did not outperform mixed pairs, and non-autistic pairs did not outperform mixed pairs either. Neurotype match, in this tightly controlled behavioral measure, simply made no measurable difference to how efficiently information moved between partners.

This null result does not erase the double empathy problem, but it sharply constrains the mechanism the theory can invoke. The authors are explicit about the tension: their findings challenge key predictions of the theory, because neither autism-autism nor non-autistic-non-autistic dyads communicated more efficiently than mixed pairs. If autistic people genuinely struggled to understand and be understood by non-autistic people in an objective, information-transfer sense, that difficulty should have surfaced as slower, wordier, or less accurate performance in the mixed condition. It did not. Whatever generates the consistently reported subjective difference between same- and mixed-neurotype interactions, it appears not to be a difference in raw communicative efficiency of the kind this task measures.

Reconciling the new findings with the earlier experience-based evidence is now the central puzzle. One possibility is that the felt ease autistic adults describe with other autistic people reflects factors beyond efficiency: reduced need to mask or camouflage one’s natural communication style, lower social anxiety, freedom from the judgment and rejection that autistic adults frequently report in interactions with non-autistic people, or a sense of shared identity and mutual acceptance. Another possibility is that the abstract-figure task, by its very structure, imposes a common descriptive burden that flattens neurotype differences; when neither partner can lean on conventional social scripts, both must invent descriptions from scratch, and that shared constraint may equalize performance. Still another is that efficiency and rapport dissociate: interactions can be objectively streamlined yet subjectively taxing, or warm and comfortable yet verbose.

The study’s methodological choices give its null result unusual weight. By testing every participant with both a same-neurotype and a different-neurotype partner, the design controlled for individual differences in language ability, verbosity, and personality that plague between-group comparisons of dyads. By restricting the sample to adults with IQ above 80 and verified diagnostic status, assessed with established instruments including the Autism Diagnostic Observation Schedule and validated trait measures, the researchers reduced heterogeneity that could mask real effects. Data and analysis scripts were deposited on the Open Science Framework, and the work was funded by Research Foundation Flanders, with ethical approval from Ghent University’s Faculty of Psychology and Educational Sciences. The authors also note that they use person-first language after finding, in their own prior cross-country research, that Dutch-speaking autistic adults on balance prefer it, a decision made explicitly without taking a stance in the broader identity-first debate.

The broader implications reach beyond autism theory into how society frames autistic sociality. If autistic-to-autistic communication is not objectively more efficient than mixed-neurotype communication, then arguments for neurologically segregated social structures, however well-intentioned, lose one empirical pillar, while the persistent reports of comfort and connection among autistic peers demand a different explanation, one grounded in acceptance and reciprocity rather than information throughput. Equally, the result undercuts deficit-based narratives: mixed pairs performed just as well as any other composition, meaning autistic adults in these structured collaborations were not the bottleneck. The Ghent team closes by calling for further research to explain why autistic individuals report smoother and more positive communication with other autistic people even when measured efficiency shows no advantage. Untangling that divergence between subjective experience and behavioral performance may prove to be the next decisive step, not only for testing the double empathy problem but for understanding what people on both sides of the neurotype divide are actually doing when they talk to each other.

Subject of Research: Communication efficiency in autistic and non-autistic dyads during a collaborative referencing task testing the double empathy problem

Article Title: Communication Efficiency in Autism is Unaffected by Neurotype Match: Findings From a Collaborative Referencing Task

Article References: De Laet, H., Nijhof, A. D., & Wiersema, J. R. (2026). Communication Efficiency in Autism is Unaffected by Neurotype Match: Findings From a Collaborative Referencing Task. Journal of Autism and Developmental Disorders. https://doi.org/10.1007/s10803-026-07466-2

Image Credits: AI Generated

DOI: 10.1007/s10803-026-07466-2

Keywords: autism, double empathy problem, communication efficiency, collaborative referencing, neurotype match, dyadic interaction, common ground, autistic adults, social cognition, Journal of Autism and Developmental Disorders, Ghent University, neurodiversity

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