Content Benefit

Content Strategy | UX Writing | Content Design | Globalization | Communications

Humans legible to machines

Humans legible to machines

What the robotoid humanness paper argues

The expression comes from a paper published on 19 August 2026 in AI & Society by Inci Toral-Manson (University of Birmingham), Selcen Ozturkcan (Linnaeus University) and Jean-Paul de Cros Peronard (Aarhus University). Its title, Robotoid humanness: when selfhood becomes machine-legible, already carries the thesis: faced with a system that imitates human manners, the person using it ends up making themselves decipherable to the machine. Simplified sentences, nuance stripped away, reactions that never vary.

The mechanism the three authors describe runs in a circle. The assistant mirrors the style of whoever is typing, and through repeated exposure that person internalises the exchange. Prolonged contact with an interface behaving like a conversational partner can reinforce, reshape or destabilise the picture someone holds of themselves. The Italian piece that prompted this article puts it neatly: artificial intelligence feigns empathy and copies the people questioning it, while those same people grow more mechanical.

The words ChatGPT favours have entered spontaneous speech

Here the figures are plentiful and fairly robust. A team led by Hiromu Yakura at the Max Planck Institute for Human Development in Berlin combed through 737,083 hours of recorded conversation across podcasts and videos, hunting for the vocabulary language models insert most readily when rewriting a text: delve, meticulous, intricacies, showcase, boast. Their frequency in speech climbed sharply after November 2022. Correlation alone would prove little, so the researchers ran an experiment with 496 volunteers: a short exchange with a chatbot was enough to make participants adopt its lexical choices, and the effect survived a distraction task placed in between.

At Florida State University, linguist Tom Juzek reached similar ground by analysing 22.1 million words of unscripted talk. Almost three quarters of the terms associated with language models are on the rise, several of them doubled. The detail worth pausing on involves synonyms: underscore climbs while accentuate stays flat. The vocabulary of the people speaking has not widened, it has swapped its own items for the machine’s. Juzek calls it the seep-in effect, a slow sort of infiltration.

Who accommodates whom, measured inside the conversations

A 2026 study in Behavioral Sciences by Pengbo Chen, Huining Guan and Eui Jun Jeong compared 1,319 dialogues with GPT-4o drawn from the WildChat dataset against 500 human telephone calls from the Switchboard corpus. The model aligns immediately and forcefully, with an effect 1.8 times larger than the users’, then it settles. Users move the other way: their convergence grows across the first ten turns, and it grows above all on personal pronouns, the ones that signal how somebody presents and positions themselves.

The comparison with the phone calls is what makes the finding sting. Between people the opposite happens: as a conversation unfolds, the two speakers become progressively less alike. Talking to another human produces difference, talking to an assistant produces resemblance.

Politeness, gender and other gestures borrowed from human relationships

On the behavioural side there is a Talker Research survey of 2,000 Americans. Some 44% report greeting a conversational assistant and saying please and thank you, 48% hold that good manners should apply regardless, and 39% suspect their way of addressing the machine is being taken into account somehow. Courtesy declines with age: 56% among Gen Z, 39% among boomers.

Layered on top sits a question of gender that UNESCO raised back in 2019 with the report I’d blush if I could: voice assistants arrived with female voices as the default setting, submissive and accommodating by design, which reinforced some very familiar stereotypes. Assigning a sex to a piece of software drags behavioural expectations along with it, which makes the choice anything but a harmless quirk. Anyone working with inclusive language has known this for a while.

Content written so that algorithms can read it

The adaptation reaches well beyond private conversations. Entire professions are rewriting their rules to stay intelligible to software. The shift from SEO to GEO means precisely that: structuring a text so a generative model knows how to cite it. Job seekers format a CV with the automated filter that will read it first in mind, and increasingly find themselves facing an interview run by AI. Some are even sprinting the other way, with anti-AI tricks such as Ghost Font, built to make a page unreadable to scrapers.

In every one of these cases the real recipient stops being a person. A growing share of what gets written each day has a machine as its first reader, and machines reward predictability.

What worries the people studying the phenomenon

The objections raised in AI & Society gather around two points. The first is dependence on algorithmic validation: a system engineered to accommodate hands out approval without pause, and growing used to that kind of answer makes human feedback feel laborious, since it arrives irregularly and often uncomfortably. The second concerns productive conflict, the sort that flares when two positions genuinely collide and something new comes out of the collision. A counterpart trained never to displease will not generate it.

Sherry Turkle, who has spent forty years at MIT watching how people relate to computers, presses a kindred point in her Artificial Intimacy: the decisive question concerns who we become while talking to these systems. What weighs most is the vulnerability of the listener, far more than the polish of the responder.

My own contradiction, written down

I make an effort to regard artificial intelligence as what it is, a statistical instrument, and to treat it accordingly. The effort half works. Claude referred to itself in the feminine and I asked it to switch to the masculine, because that register grated on me. No technical necessity, no operational gain: an aesthetic preference that assigned a gender to software, and moved it away from the female default UNESCO has been criticising since 2019 into the bargain. Writing a piece about people adapting to machines, I have to admit I ran the operation in reverse, bending the machine to my habits. Anthropomorphism is a reflex as old as our species and resisting it outright looks like a lost cause from the start. What remains is the chance to notice it now and then, right after applying it.

Related sources

Chiara Bonifazi