In this article
Today we talk about a small, everyday observation worth naming before it passes as if it hadn't happened. In cafés, in airports, in waiting rooms, in open-plan offices, the soft noise of typing has dropped noticeably between 2022 and 2026. The proportion of people articulating text versus those who read and tap short confirmations has changed. The thesis: the silence of the keyboards is a symptom of something wider than mere mobile digitalization. It's a symptom of the handover of verbal articulation to assistants that write for us and to workflows where the person approves instead of drafting. My take is biased by my own profession, which is writing. Form your own: go to a downtown café one day and count how many screens show long text and how many show buttons.
For two years I've had a little mental game when I walk into cafés or coworking spaces. I count the screens with long text —open documents, lengthy emails, half-composed drafts— and the screens with a tap-interface —Slack, Teams, WhatsApp, forms, dashboards. The proportion has shifted clearly. It isn't statistical proof, it's informal observation. But it matches what ergonomists, work sociologists and the few digital ethnographers who dare to study such subtle changes are reporting.
The observation, without alarmism
Before 2020, an average professional produced during the workday several thousand of their own words across different channels —emails, reports, messages, drafts. That production wasn't literature and wasn't evaluated as such, but it was sustained verbal articulation. It involved choosing words, building sentences, maintaining coherence between the start and the end of a paragraph, deciding tone according to the recipient.
Between 2020 and 2022, with the mass remote work brought on by the pandemic, there was a temporary increase in textual production because in-person meetings were partly replaced by written exchanges. That peak was observed in productivity reports and in corporate communication tools.
From late 2022 on, with the spread of ChatGPT and equivalent assistants, the reversal began. And from 2023 on, with the incorporation of assistants into the productivity platforms themselves —Copilot in Office, Gemini in Workspace, assistants built into Notion, Slack, Asana, Salesforce—, the change accelerated. The production of long text per average worker hasn't disappeared, but it has dropped substantially. What rises is the review, the approval, the marginal modification of text produced by machine.
The silence of the keyboards is, literally, that. Fewer keys pressed to produce new words. More clicks and short taps to approve words already produced.
What the typing did internally
Here comes the part that deserves detail, because the difference between articulating and reviewing isn't trivial.
When a person writes a long sentence, several cognitive things happen in parallel. They rehearse in their head —and in their fingers— different possible formulations before settling on one. They hold in working memory the start of the sentence while deciding the end, making sure the end answers the start without contradiction. They discard options by rhythmic intuition, the fruit of years of exposure to similar texts. When a sentence goes off course —clumsy rhythm, lexical repetition, referential ambiguity—, the brain detects it and corrects it on the fly, usually before even finishing typing it.
All of that happens at a very high frequency during a normal day of intellectual work. A hundred long sentences a day mean hundreds of small exercises of planning, working-memory maintenance, intuitive evaluation. The sustained accumulation over years produces a specific cognitive profile: that of the person used to articulating thought in their own language in real time.
When that same person switches to reviewing text produced by a machine, different things happen and, above all, fewer. Review activates shallower processing. You read, you assess whether it fits the general intention, you correct details. There's no initial planning, no sustained working-memory maintenance, no intuitive discarding of options, because the options have already been discarded by someone else —by the model. The processing is faster but also shallower. Less deep in its internal coherence.
The neurological difference between the two modes of work is starting to be documented. The Kosmyna and collaborators study from the MIT Media Lab that we saw in 0101 measures precisely this difference. The assistant users showed lower activation in regions associated with linguistic planning during production, which is consistent with the qualitative observation that the person delegates much of the planning to the model.
What confirmation doesn't train
For a person who has made that gradual change over three years, the measurable cognitive profile has shifted. Not dramatically —it isn't neurological deterioration— but it has shifted in concrete skills.
The capacity to hold a complex argument in working memory for minutes, while building it in real time, weakens. The skill still exists as potential, but the operational fluency is reduced.
The capacity to discard inadequate formulations by rhythmic intuition, before settling on them, too. The person who for years has read and accepted rather than articulated loses calibration of the inner ear.
The capacity to generate several expressive options simultaneously in order to choose the best one is reduced, too. The brain stops investing effort in producing alternatives when a reasonable alternative is available immediately on request.
And, most subtly, the capacity for transition between sentences —the small grammatical turns that connect the end of one idea with the start of the next— weakens. Those transitions are trained only by writing long prose. If you stop writing it, you stop training them.
None of these deficits is catastrophic. Any of them can be recovered with deliberate practice. But, without that explicit retraining, the cognitive profile of the average professional worker in 2026 is different from that of 2020. Faster at tapping approval. Less fluent at sustained articulation. The social accumulation of that change is what's worth observing.
The loss that isn't measured in exams
Standard cognitive assessments —IQ tests, academic exams, reading-comprehension evaluations, verbal-fluency tests under controlled conditions— don't capture this difference. The tests measure discrete capacities under bounded conditions. The skill of holding a sustained argument for minutes without assistance, while building it, doesn't appear in any standard test because the tests don't last long enough and are taken under conditions where there's no assistance available.
That means the aggregate cognitive change in society, if it's happening, will pass under the radar of the existing measurement systems. University exams will keep yielding results that probably look stable, while the real skill measured in a natural context changes. Workplace evaluations will show measurable productivity that probably looks the same or higher, while the quality of the underlying thinking changes.
This isn't speculative theory. It's what the studies on the effect of information-at-a-click have documented —Sparrow, Liu and Wegner published in Science in 2011 Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips— about the change in how the brain stores information when it knows the information is externally accessible. Risko and Gilbert extended the framework in Cognitive Offloading (Trends in Cognitive Sciences, 2016). The contemporary generalization to linguistic articulation by AI assistants is consistent with that earlier line of research.
The aggregate consequence that worries most
If the observation is correct —and the available data, still preliminary, suggest it is—, the aggregate consequence for public conversation deserves attention.
A society where most of the textual production in professional settings is done with assistants and approved by humans has a public conversation mediated by models in a deep and non-obvious way. Professional reports, corporate statements, press releases, technical proposals, legal drafts, pedagogical reports, academic writing: if a growing share of these texts passes through models as a first draft, the shared vocabulary, the frequent turns, the predominant argumentative structures begin to converge toward those the models produce.
That doesn't mean humans lose their voice. It means the human voice is increasingly channeled through frames the machine facilitates. The stylistic tics characteristic of the models —certain paragraph structures, certain turns like «not X, but Y», certain tendencies toward the tripartite enumeration— filter into average professional prose, and from there into public conversation.
This is observable today in specific sectors. Whoever has spent months reading corporate reports, press releases or institutional communications can recognize the models' style with relative ease. The proportion of texts with those traits rises each quarter. And, as those texts become part of the training corpus of the following models, the cycle closes. The model trains on its own output mediated by humans. What the sector calls, with justified concern, model collapse of human language.
The political question
This is personal opinion, but the accumulated observations back it. The silent concentration of professional linguistic articulation in the hands of AI assistants has cultural consequences that aren't being sized up.
Democratic public conversation depends on citizens' capacity to articulate their own arguments. When a significant share of the texts circulating in the public space —including texts signed by humans in professional and academic settings— passes through models as a first draft, the real diversity of argumentative frames is reduced. Not by censorship. By statistical convergence.
What I would ask for —again with limited realism— is that educational and professional institutions deliberately maintain spaces of articulation without assistance. University exams with writing on paper without an assistant, at least some. Professional work sessions with an explicit ban on assistants, at least some. Not out of nostalgia. For the preservation of the aggregate skill.
On the individual plane, the decision is each person's. But there's something worth knowing. The skill of articulating thought in your own language in real time is trained by continued practice. If you drop it, it weakens. If you want to keep it, you must practice it with some regularity, even if it isn't operationally necessary in your work. It's cognitive exercise, like walking is physical exercise. You don't need to compete; it's enough not to drop it entirely.
The hard fact to close on. According to the State of Workplace AI 2025 report published by the Microsoft Work Trend Index in May 2025, based on a survey of around 31,000 workers in 31 countries, 75% of the professionals surveyed used some generative AI assistant at least weekly in their work, up from 22% a year earlier. 39% declared that most of the text they handed in at work was produced with the assistant's initial help. That figure —39%— is the quantitative translation of the silence of the keyboards. It isn't the anecdotal perception of someone who writes; it's the majority reality of contemporary professional intellectual work. Whether that change is good or bad in aggregate depends on what's done with it. But ignoring it is failing to understand the cultural transformation we're going through.
Definitions
Working memory: a limited cognitive capacity to keep information temporarily active while processing it. It's critical for sustaining long arguments during writing.
Cognitive offloading: the practice of transferring mental tasks to external tools. Studied for decades for calculator, GPS, contacts; now applied to linguistic articulation via AI assistants.
Google effect: an effect documented by Sparrow, Liu and Wegner (Science, 2011) by which the brain remembers information worse when it knows the information is externally accessible. It's a conceptual basis for understanding the cognitive change associated with the use of assistants.
Sustained articulation: the capacity to hold a complex argument for minutes while producing it in language, without external assistance. It's the skill whose training is most reduced by the spread of assistants.
References
Betsy Sparrow, Jenny Liu, Daniel M. Wegner, Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips (Science 333:776-778, August 2011). The foundational paper on the change in memory due to the external availability of information.
Evan F. Risko & Sam J. Gilbert, Cognitive Offloading (Trends in Cognitive Sciences 20:676-688, September 2016). A theoretical review of how the brain distributes cognitive load to external tools.
Nataliya Kosmyna et al., Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task (arXiv:2506.08872, MIT Media Lab, June 2025). A study on the measurable brain effects of AI assistance in writing.
Maryanne Wolf, Reader, Come Home: The Reading Brain in a Digital World (Harper, 2018). A neurocognitive framework on the changes in the reading and writing brain.
Microsoft Work Trend Index, State of Workplace AI 2025 (Microsoft, May 2025). Figures on the use of AI assistants in global professional environments.
Lev Vygotsky, Thought and Language (MIT Press, 1962). A classic framework on the relation between articulated language and thought.
Nicholas Carr, The Glass Cage: Automation and Us (W. W. Norton, 2014). An analysis of how automation weakens previously exercised human skills.
To go deeper
John Sutton et al., The Cambridge Handbook of Cognitive Extension (in preparation, Cambridge University Press, 2026). An interdisciplinary framework on distributed cognition and its contemporary effects.
Susan Greenfield, Mind Change: How Digital Technologies Are Leaving Their Mark on Our Brains (Random House, 2014). A popular neuroscience framework on the brain effects of digital technologies.
Sherry Turkle, Reclaiming Conversation: The Power of Talk in a Digital Age (Penguin Press, 2015). A framework on the loss of conversational and articulative skills in the digital age, a direct antecedent of the current phenomenon.
Howard Rheingold, Net Smart: How to Thrive Online (MIT Press, 2012). A practical manual on how to keep cognitive skills against digital overload, transferable to the age of assistants.
You might also like
- The price of letting AI write for you
- Why starting a text keeps getting harder
- The domestication of AI
Comments0
No comments yet.
Leave a comment