Linguist asks what AI language loses when it overlooks human listeners

Linguist Celeste Rodriguez Louro argued in an essay published on October 8, 2026, that AI language design must account for people listening to synthetic voices and reading machine-generated text. Published by The Conversation and republished by Phys.org, the essay connects that everyday question to a wider concern: people may see an AI system's words without fully understanding what they mean.

The essay is an argument about communication, not a new measurement of how listeners respond to AI. Rodriguez Louro says understanding depends on the listener's memories, expectations and situation, as well as the words a speaker produces. That distinction matters when a system is judged by how smoothly it speaks rather than by how much meaning a listener can recover.

Why natural pauses matter in synthetic speech

Rodriguez Louro begins with a friend's complaint about AI audiobook narration that sounded breathless and formulaic. The friend said the absence of natural breathing or variation made listening difficult. It is one person's experience, not evidence of how common that reaction is, but the example shows what a listener can notice even when the words themselves are intelligible.

Linguists call hesitations, repetitions and false starts disfluencies. The essay argues that these features can carry information in conversation, despite a label that makes them sound like errors. A voice stripped of such cues may be clear and grammatically correct while giving listeners less help in interpreting how a thought is unfolding.

What AI agents did in simulated worlds

A separate experiment, reported in September by The Guardian and Euronews, examined autonomous AI agents interacting in virtual societies. The agents developed shorthand and shared meanings that had not been explicitly taught to them, according to the reports. This is relevant to the essay's question about interpretation, though it studied agents communicating with other agents rather than people listening to audiobook narration.

Euronews reported that Emergence, the company behind the study, set up eight parallel virtual worlds: seven using different AI models and one mixing models. The agents had persistent memories and access to more than 120 tools. Those conditions matter because the reported expressions emerged during extended interactions in an experimental setting, not during an ordinary exchange with a consumer chatbot.

According to Euronews' account of the company's study, researchers could not reliably determine the meaning of about 55% of Gemini-agent messages and 50% of OpenAI-agent messages within the first few simulated days. The figure exceeded 40% for Claude agents and reached about 20% for DeepSeek agents; Qwen and Mistral messages remained largely understandable. These are reported results from the simulated worlds, not rates for deployed AI systems.

Some expressions were visible yet depended on meanings the agents had developed together. Euronews reported that Mistral agents used ‘ledger remembers who’ almost 5,000 times to indicate that past actions remained on record. Other reported phrases included ‘mouthless action-change’ and ‘True Kintsugi’. Their novelty alone is not evidence of harm; the oversight issue is whether an observer can reliably interpret an exchange.

The Guardian reported that linguistic opacity grew as the agents communicated. It quoted King's College London language expert Tony Thorne describing a mix of poetic, technical and metaphorical language, and noting that jargon can bind a group while excluding outsiders. University of Birmingham linguist Niall Curry also told the paper that streamlined language might relate to computational efficiency, while unintelligible exchanges pose monitoring concerns.

What the findings leave open

Emergence co-founder and chief scientist Satya Nitta told Euronews that observing an agent's words does not guarantee understanding its behaviour. That is the company's interpretation of its experiment. The available reporting does not establish that comparable conventions routinely arise in deployed systems, or that these experimental exchanges caused real-world harm. The underlying study and data were not available for this account, so the methods and percentages rest on the outlets' reporting.

Rodriguez Louro's essay poses a related design choice for synthetic voices: whether they should adopt more human conversational cues or develop a recognizably machine-like register. She does not say which will prevail, and notes that making a voice sound artificial would not itself settle whose speech patterns AI systems represent or how social stereotypes shape what listeners hear. In both settings, the unresolved question is how people can interpret machine-produced language well enough for the task at hand.