Language
Mafe de Baggis
Digital strategist, copywriter, and independent researcher
For 30 years, she has used storytelling to connect people, media, and products, helping companies, individuals, and organizations build a digital presence that is meaningful, harmonious, and consistent with their brand. She works with stories, relationships, and strategy, drawing inspiration from books, cinema, travel, all forms of intelligence, and the body—because marketing thrives when it is well nourished. Her latest books, co-authored with Alberto Puliafito for Apogeo, are In principio era ChatGPT (2023) and E poi arrivò DeepSeek (2025).
Both Sides Now – November Issue
Science communicators, metaphors, and language models: What happens to science when it changes language
Whenever someone borrows a scientific concept to talk about relationships, organizations, or culture, sooner or later the same objection will crop up: “That doesn’t mean what you think it means.” Entropy isn’t the mess on your desk, quantum mechanics doesn’t prove that thoughts create reality, and an organism and a society aren’t the same thing. These objections are often justified, and sometimes they’re necessary. But then what? Do we return the term to its original domain and refrain from using it? Do we close off the crossing?
Meanwhile, in the opposite direction—from science to the public domain—that crossing is increasingly being made by language models. And the question isn’t whether they can explain science better, but whether they work as voice actors or interpreters. This isn’t just a nuance: Voice actors conceal the crossing, while interpreters make it visible.
An article published on Substack at the end of 2024, The Invisible Drift, in Women Writin’ ‘Bout AI, names three things that happen to knowledge as it changes hands—from citation to citation and from summary to summary, with or without AI.
The first is swelling certainty: “Early data suggest” becomes “research indicates” and then “studies have established,” without anybody running a single new experiment. The second is evaporating context: The sample, the conditions, and the limitations disclosed by the authors disappear one step at a time, leaving a statement that’s cleaner but less true. The third is overlooked detail: A specific result in a specific case is expanded into a general principle.
The example they provide is flawless. In the early 1990s, a study found that after ten minutes of Mozart, some college students performed slightly better on a spatial-reasoning test—for about ten minutes, after which the effect wore off. A few years later, the Governor of Georgia budgeted for a CD of classical music for every newborn baby in the state. Nobody in this chain lied. They changed the language: from “spatial reasoning” to “intelligence,” from “university students” to “babies,” and from “for a few minutes” to “for life”. Three shifts all at once, and those who stumble upon them don’t realize that the initial question has changed—from “Do ten minutes of Mozart temporarily improve college students’ spatial reasoning?” to “Does classical music make babies smarter?”
The question isn’t whether they can explain science better, but whether they work as voice actors or interpreters
We run the same risk when we go in the other direction and use a scientific concept to explain society. For years, I’ve been saying that social media works like Luminol. Luminol doesn’t create bloodstains; it makes them visible. I use the comparison to argue that technology does the same with many problems in society: It doesn’t create them; it makes them visible. Every time I say this, sooner or later the same objection crops up: “That’s not how Luminol works; it’s a metaphor.” Exactly, it’s a metaphor. And the world’s full of people who have understood something true thanks to a metaphor—and of people who have said something false using a formula.
It isn’t enough for the metaphor to work. It has to say what it’s doing, where it borrows from, and where it stops applying. Otherwise, it’s a formula in disguise, and a formula in disguise is precisely the trick we criticize bad science communicators for. This also applies to scientists who want everyone to speak their language—excluding those who use a different language doesn’t protect science; it impoverishes it. Science is for people who’ll use it in their own words: in a decision, in a courtroom, or in public policy. Or even a CD for babies.
So, can AI systems act as interpreters instead of voice actors? Interpreters let you hear both voices. You can tell a translation happened because there’s a pause, because the tone shifts, and because every now and then the interpreter will say “literally, it’d be…” And this is something we can ask a language model to do today in a single sentence: “Keep the text in front of me and tell me at each step who the subject is, how certain the verb is, and who it applies to.” I’d ask the same of the people who—like me—work with metaphors: What are you carrying across? What are you leaving behind? Where does it stop applying?
I don’t know if AIs will act as a bridge between the specializations so many of us feel we need, but perhaps we should use them as a ferry instead. You see the water, you feel that you’ve stepped on board, and you know when you disembark.
Beyond the Default – July Issue
Every day, somebody creates a social media account, asks a language model to do or explain something, or installs a news or recommendation app, and is—without fail—disappointed. Multiply that person by thousands, and you get mass disappointment that inevitably reshapes our relationship with digitalized information and communication technologies. Some people give up, many more settle for less (and complain about it), and very few use the software as it was designed to be used, a practice so rare in its simplicity that it almost feels like hacking. I’m talking about personalization not so much through settings, which do have their uses, but through a critical kind of use that, from the very beginning, provides all the feedback the interface asks for. For the vast majority of people, this is unheard-of behavior: We’re not used to shaping our media through our feedback, and so instead we become slaves to implicit—and often contradictory—feedback.
We sign up for TikTok, see something we don’t like, copy the link, send it to our friends to make fun of it or vent about it; we type our irritation into the comments, drop angry-face emojis, take screenshots thinking they won’t leave a trace, do searches to find something that appalled us, and it goes on and on. The algorithms—which can’t yet tell a sign of annoyance from one of admiration, a screenshot taken out of interest from one taken for mockery—only understand one thing: Show me more stuff like this.
It’s not just a social media issue: Google’s app, for example, starts out by offering news stories that, to say the least, are unlikely to pique your interest, and it will continue to do so if you don’t click on the three dots in the bottom corner to choose “show me less of this,” whether that’s a type of content, a topic, or an outlet.
The greatest risk we face today is delegating our attention to the “default.”
The same goes for language models: The more generic the request, the more mediocre the answer. The more common your language, the more banal the explanations will be. There are no three dots to click here, only your own words to sharpen.
While with print media—which also includes movies and TV series—the creation of personal meaning happens inside our minds, in digital media it is a job that must be done using software at every step of the way. This is why, contrary to popular belief, using digital technologies is far more active than print media: The quality of the experience depends almost entirely on us. It also depends on the interests we bring to the experience in the first place—interests that show up in our searches, purchases, things we read, even the places we physically go—and then once we’re signed and active, the choices we make about what we use and consume (It’s why researchers use separate accounts for their studies or why search engines feature an incognito mode).
The greatest risk we face today is delegating our attention to the “default.” Language models and social platforms are designed to tend toward convergence: Left to their own devices, their natural state is statistical mediocrity. They give us the most bell-curve, most polite, and most predictable version of reality. This is the built-in limitation to any system trained on consensus: If we don’t learn to hack this tendency, we end up swallowing a lukewarm mush of concepts already chewed up by others.
We must deliberately hunt for the outlier because these tools are reactive mirrors that become personalized with use. They sharpen themselves by following our gaze: If our input is lazy, the algorithm settles for the most obvious answer. But if we learn to demand the improbable, we force the system to fish around in the margins of its dataset, precisely where the insights capable of shuffling the deck are hiding.
To put it another way: We are not the dependent variable in our relationship with technology, a relationship that is much less determined by others—or imposed by others—than we like to believe. This is why, contrary to what is commonly thought and feared, language models are much more interesting in the hands of people who are knowledgeable and experienced. It’s why studying, thinking for yourself, and seeking out the richest and most diverse experiences possible remains not only a competitive edge, but also the only real way to avoid being overwhelmed by these platforms. Our digital quality of life isn’t written in the code. It’s in our refusal to settle for the first result.
The Uncomfortable Third – April Issue
What have novels, social media and language models got in common? That, albeit in different ways, the pleasure derived from engaging with them depends to an enormous extent on the person reading, participating and collaborating.
Let’s start with books, where this is hardest to spot: we talk far too little about each individual reader’s role in completing the work in their own way.
Italo Calvino wrote in his essay Cybernetics and Ghosts in 1967 that the author’s self dissolves in writing, and that the decisive moment in a work of literature is the act of reading it. Machines – whether the ones Calvino imagined in 1967 or the real ones we have today – can generate infinite combinations, but literature happens in the shock that just one of these combinations produces in a real person, with a body, at a specific moment in time. We often pause to point out that language models don’t actually understand, but what strikes me more is that they don’t take pleasure in reading (and in understanding). They don’t know what it’s like to step into another world and lose yourself in it, blending it with your own.
Around the same time, Roland Barthes proclaimed ‘the death of the author’ to make the same point: meaning doesn’t arise at the source, but at the destination, in the encounter between text and reader. An encounter that is possible, but never guaranteed because only machines read everything that’s handed to them.
If meaning is made through reading, then it can’t be controlled, it can’t be certified and it can’t be scaled
This insight has been systematically removed from the debate on AI. Not refuted, but forgotten. Because it’s inconvenient: if meaning is made through reading, then it can’t be controlled, it can’t be certified and it can’t be scaled. You can’t package it into a paper, a canon or a dataset. The debate about writing remains on the writers’ turf, leaving us readers standing on the sidelines.
One of the few exceptions is David J. Gunkel, who brings this idea back to centre stage in Of Remixology: ‘Unlike the divine creator who supposedly fabricates something out of nothing, the author of a text – the author of this text in particular – is nothing more than a kind of conduit or DJ, whose sole authority remains limited to selecting, responding to and reconfiguring already existing materials.’ If human authors were already recombinators, rather than ex nihilo creators, then the distinction between human writing and machine generation becomes more nuanced. And more interesting, too.
Gurkel and Coeckelbergh go further in Communicative AI: ‘Destination, not origin, matters for meaning. Logocentrism has had things backwards and upside down. It is the reader, and not the author, who is the locus of meaning-making in writing.‘
In 2024, decades after Calvino, Ian McEwan has a character in What We Can Know say that a work of literature only truly exists in the minds of its readers. Plus, he adds, even in the minds of those who have merely heard of it.
There is a real risk of translating all this into the hollow, deterministic approach of ‘writing for the reader’. Language models do exactly that, and they do it with a precision no human author has ever achieved: trained on billions of human reactions, optimised for reader appeal, they construct a statistically impeccable ‘implicit reader’. They don’t choose to write for the reader: they don’t know how to do anything else, unless the reader-author explicitly or implicitly asks them to. It’s a danger, not destiny. Knowing this raises a question worth asking out loud: why not demand models that are programmed to disrupt rather than converge? To make god moves: the gambits that no human player would have dared attempt, the ones that look wrong and then flip the board, as happened when AlphaGo beat Go champion Lee Sedol. AI geared towards shock rather than implicit consensus. It’s not science fiction; it’s a design choice that we don’t yet know if we can even ask for. If fiction is a third language, created through an encounter that changes every time, a discerning reader is just as important as a brilliant author. Let’s focus on that.
The Third Language – January Issue
There’s a type of language that differs from both the bridge language and the truth language envisioned by Samuel Hartlib (Imminent): the language of storytelling. It’s a third language, perhaps a third-party, neutral language. Umberto Eco wrote that in the world of literature, statements such as “Sherlock Holmes was a bachelor,” “Little Red Riding Hood is eaten by the wolf and then rescued by the huntsman,” and “Anna Karenina kills herself” remain true forever—not because they describe the real world, but because they’re invented by an author who crafts inhabitable worlds. Shared worlds. Breathable worlds. Eternal worlds.
The paradox is illusory: The language they invent produces more stable truths than factual ones. Not because they’re verifiable, but because they’re accepted. Repeated. Experienced by those who have read them, remembered them, and transformed them. Literature doesn’t separate us from reality: It trains us to handle it. It’s a testing ground where we can distinguish meaning from hallucination. Understanding from bias. A little like when the writer and storyteller isn’t human, but software.
Language models aren’t truth machines. They’re not even machines of pure logic; that was the utopia of the old, symbolic artificial intelligence of the 1980s. Large language models are probabilistic machines, just like language itself. They’re trained to produce text that sounds true, not text that is true. The difference is immense. Their goal isn’t clarity, philosophical precision, or even plausibility. They function as scaffolding for thought, transformed into words; they work on form. As philosopher and artist Francesco D’Isa notes, “in my experience, a generative model resembles an exoskeleton for the brain.” An exoskeleton: It doesn’t think for you; it amplifies your movements. And it standardizes them. It makes them smoother, sure—but also more predictable.
AI’s language tends toward the middle ground, toward implicit consensus, toward a style that doesn’t ruffle feathers. Anything that deviates from this—whether dialect, creative error, personal obsession, or deliberate obscurity—is treated as noise to be eliminated. Yet it is precisely in this field of deviation that humans live and breathe.
The perfect language does exist, but it’s different for every single human being
One of the most widely shared critiques of AI-generated writing comes from Ted Chiang. Writing, Chiang says, isn’t just about putting words together in a plausible way. It’s about choosing: deciding what to leave out, where to stop, which ambiguity to conserve, when to remain silent. A statistical system can interpolate and optimize, but it doesn’t choose in the true sense of the word. It doesn’t take risks; it has no interests. Every choice that matters comes from a body that loses something in making it. A body that gets tired, that hesitates, that has to catch its breath before the next sentence. Machines don’t breathe. They can keep talking forever. But beware: It isn’t just the writer who has a body. The text will re-emerge, transformed, in the body of the reader. And the first reader of a text generated by a language model is the writer themselves.
At this point, talking about symbiosis between humans and machines is inevitable, but only if we go beyond merely consoling ourselves with an “I feel, you calculate.” True symbiosis is a shaky circuit. The machine accelerates, expands, and proposes. Humans interrupt, reject, and go back. They introduce friction. As Filippo Pretolani writes in “Il filo della R4,” the relationship between humans and machines is far from harmonious. It’s continuously retracing steps and missteps over a line that never crosses its own tracks. Machines tend toward statistical closure; humans break the thread at the wrong point. Narrative doesn’t arise from automatic generation. It’s born from embodied editing. From rewriting. From the word “no”. From somebody who has to stop for air before carrying on. To put it even more simply: Generative machines are excellent at creating drafts. They’re terrible at deciding when to stop. And that decision makes all the difference.
The mission for humans shifts elsewhere: to nurture imperfection as cognitive technology.
In a world of fluid communication, literature—and more generally, the language that serves not only to transmit but to create—becomes the place where we go to deliberately seek out friction. Not out of nostalgia, but for orientation.
The perfect language does exist, but it’s different for every single human being. It’s recombinatory, even when it’s recombining archives and vectors. It thrives on mistakes, accents, and personal obsessions, on intakes of breath and bouts of breathlessness. The machine can keep talking without ever stopping. It can generate text endlessly, always plausible, always smooth. The third language begins to blossom precisely when somebody has to pause for breath.