The Name of the Future, Part Two

This piece is about the names we give technologies and what those names do afterwards, once they get into a contract or into the head of someone using a machine. I’ll start with a list published in Paris.

In 2025, for the artificial intelligence summit hosted by France, the Délégation générale à la langue française, the office of the Ministry of Culture that looks after the language, gathered fifty key AI terms into a small volume. For each recommended French word there is a column with the equivalents in the other European languages. Under infox vidéo, the official French way of saying deepfake, the Italian column reads “deepfake”. Catalan has “hipertrucatge”. According to an official document from another country, we Italians have no word of our own for that thing, and on a close reading the document is right. DGLFLF

In the first part of this argument I argued that a country that gives no name to its own future receives it already named by others, with the priorities of whoever coined the word built in. That was an argument about the geography of words, about who invents them and who translates them. Here I want to go down one floor and look at the mechanics: what happens to a word after someone has chosen it, and who ends up paying for the consequences.

The Objection: Function Over Name

The most serious objection comes from people who know the law, and I know it well because for years I made it my own. A product’s name belongs to the sales department, and the law is right to ignore it. A rule that chased words would be out of date the day after it passed, because sellers invent names faster than any parliament writes articles. The AI Act is built on that caution. It lists no products and no brands; it defines an AI system by what it does: a machine-based system that infers, from the input it receives, how to generate outputs such as predictions, content, recommendations or decisions. Call it an assistant, an engine or a platform, for the regulation nothing changes. It was a wise choice, and anyone who works in compliance knows how convenient a text is when it doesn’t need rewriting for every new product.

Grant all of that, and a problem remains. Between the function the law describes and the person using the system there is always a word, and people behave according to the word, not according to the definition in Article 3. The regulation knows this. Article 14, on human oversight of high-risk systems, requires the provider to enable the people doing the oversight to remain aware of the tendency to over-rely on what the machine produces, and it gives that tendency a precise name, “automation bias”. It’s one of the rare places where a European statute admits that the psychology of whoever sits in front of the screen is part of the risk. A product called “autonomous” manufactures that bias before it is even switched on. EUR-Lex

There is also a detail in the definition worth reading twice. Article 3 speaks of systems designed to operate with “varying levels of autonomy”. For the law, autonomy is a quantity, something you measure and that changes from one system to another and from one context to another. For the seller it’s an identity, something the product is. Responsibility gets lost in the gap between those two ideas, and usually nobody notices until something goes wrong.

The French and the Power to Define

Some countries have decided not to resign themselves to that gap. In France a commission for the enrichment of the language publishes in the Journal officiel, the Republic’s official gazette, the recommended terms for new things, each with a definition. The terms then go into a public database, FranceTerme. For state administrations their use is mandatory in place of the foreign equivalents, under a decree of 1996; for everyone else it remains an invitation.

On artificial intelligence the work is dense. A list from December 2018 fixes “intelligence artificielle” and the abbreviation “IA”, replacing a publication from 2000. In the same list the chatbot becomes dialogueur, and a note warns that the expression “agent conversationnel”, though it was in circulation, is discouraged. Bulletin officiel In 2024 the prompt became instruction générative. The word agent has been in the official vocabulary since September 2000, with a definition that today sounds almost prophetic: a program sent across a network to carry out a task in the user’s place and without their intervention. One line already holds what the market would discover twenty-five years later, delegation together with the absence of whoever delegates.

The tradition is older than computing itself, or nearly. In 1955 IBM France asked a Latinist at the Sorbonne, Jacques Perret, what to call its new electronic machines, and he replied with a letter proposing ordinateur, a word that in Old French described God setting the world in order. Bibnum In 1962 an engineer, Philippe Dreyfus, joined information and automatique and got informatique. One of the two words was requested by a company, the other was invented by someone who used the machines. Neither came from an academy, and both won.

Here comes the second objection, and it too deserves to be taken seriously: the French are faintly comic. Nobody in Paris says dialogueur at the coffee machine, and vidéotox, the other official version of deepfake, has met the fate of almost every state neologism. Counting how many words win against English, though, means looking in the wrong direction. What matters is that a public institution takes on the job of writing a definition and making it binding, at least on itself. A published definition can be challenged, and corrected in a later list. An imported word arrives without a definition, or rather with the definition of whoever sold it, written in the marketing material and never discussed by anyone.

In Italy nobody has that job. The Accademia della Crusca gives opinions, and gives them well, but it has never had the power to set the rules of the language, and I don’t think giving it that power would help. An Italian Académie would be the wrong conclusion to this argument, and a slightly ridiculous one. It’s enough to notice who fills the gap. Here the definition, when anyone takes it on, is taken on by whoever writes the contracts and the technical documentation. Open any tender specification or product sheet and you’ll see agents and autonomous platforms appear, always in the present tense and almost never with a line explaining what they are supposed to mean. In practice, the definition is written by whoever is selling.

“Agent”: On Whose Behalf

Let’s take, then, the three words we hear most often today, starting with the most interesting. “Agent” is interesting because, taken seriously, it’s honest. In English the word carries a whole branch of law, agency, built around someone who acts on behalf of a principal. In Italian it has a legal meaning too, and not a recent one. The codice civile, the Italian civil code, says that under an agency contract one party takes on the task of promoting the conclusion of contracts “on behalf of the other” (Article 1742), and that a mandate obliges one to perform legal acts, here too, “on behalf of the other” (Article 1703). The rules on who ends up bound, and when, are more complicated than that, and a flesh-and-blood sales agent usually procures contracts without signing them. But the phrase the two articles share is the one that counts. An agent, in Italian as in English, is someone who acts on behalf of another, within a task that other person has given and within limits that other person has set. The French definition of 2000 says the same thing in other words: a task carried out in someone’s place.

Commercial language uses the word to suggest almost the opposite, initiative and independence, a program that “does it by itself”. It’s a small inversion and it weighs more than it seems, because an agent that acts by itself is an agent without a principal, and an agent without a principal, in law as in life, is a problem for everyone around it.

The AI Act doesn’t define agents. The first dedicated governance framework came from Singapore, where the IMDA, the authority that regulates communications and digital media, presented at Davos in January 2026 a model for governing agentic AI, updated in May. These are voluntary guidelines, not a law, but the way they reason is instructive. They measure an agent along two axes: its action space, meaning the tools it can reach and the permissions it holds on those tools, and its autonomy, meaning how much it decides for itself based on the instructions it receives and on how far people are involved. And they ask for the risks to be bounded beforehand, at design time, not after the first incident. IMDA Note that here too autonomy is a quantity to be regulated, not a quality to boast about.

This is the part of my work I know best. A growing share of my days goes into designing the permissions of software agents, the ones that write code, propose changes, read documents and sometimes send a message. On projects for public healthcare, for the legal sector and for public administration the sequence is always the same. An agent that can write to the code repository or send an email needs an identity of its own, separate from that of the person who started it, minimal permissions decided in advance and put in writing, and a log of its actions that someone else can read. All of this, which fills pages of design documents, follows from a single question, the one the word agent already holds: on whose behalf. When the question has an answer, the rest is engineering. When it doesn’t, no engineering can stand in for it. I wrote about this in Who Controls What AI Agents Produce?, and since then the question hasn’t changed, it has only come up more often in meetings.

“Copilot”: Who Takes the Impact

“Copilot” looks like the most modest of the three words. It promises nothing grand, in fact it steps aside. And that is exactly why it manages to shift the blame without anyone noticing. In the cockpit the copilot assists and the captain is accountable. Calling a machine a copilot means putting the person, by definition, in command, even when that person has neither the time nor the means to check what the machine is proposing.

The researcher Madeleine Clare Elish gave this mechanism a name that’s hard to forget, moral crumple zone. In a car the crumple zone is the part of the body designed to fold on impact and protect the passengers. In automated systems, Elish argues, going back over cases like the nuclear accident at Three Mile Island and the loss of Air France flight 447, the person closest to the machine ends up absorbing the blame for the whole system, even when they had limited control over what happened. They take the impact, and the rest of the system comes out intact. Elish The rest of the system, in this case, is whoever designed and sold it.

On this point European law goes the opposite way from commercial language. The new directive on liability for defective products, 2024/2853, which practitioners know as the PLD, says in one line that software is a product. It follows that the manufacturer answers for its defects, as with a household appliance, and that the person sitting in front of the screen is not the first door to knock on. Member States must transpose it by December 9, 2026, a few weeks from now, and it will apply to products placed on the market from that day. In Italy the enabling law has been in place since March, and the draft decreto legislativo, the Italian implementing instrument, received preliminary approval in August. EUR-Lex Commercial language says “you’re the pilot”; the directive answers that if the product is defective the pilot has nothing to do with it. I told part of this story in Redacted, about what happens when a manufacturer can’t explain what its software did. Here I’ll only add that names arrive before directives, and that for years they have trained users to feel responsible for what they couldn’t control.

I see this up close too. In our craft a growing share of code is proposed by assistants and accepted by people. The change carries the signature of whoever accepted it, and rightly so, because someone has to answer for it. The practical question, faced with a few hundred lines written in a few seconds, is how much of that signature reflects a real check. Did whoever approved it read every line, and could they explain why it’s written this way and not another? If the honest answer is no, the word copilot is doing dirty work: it turns a person into a crumple zone and makes them believe they’re in command. Giving up assistants would be the easiest answer and the least useful. The one I try to practice is a review proportionate to what is being signed, and a written record of what was actually reviewed and what wasn’t. It’s less elegant than a name, but it holds up better when someone, one day, asks for an account of that line.

“Autonomous”: The Word That Removes the Subject

“Autonomous” is the word that removes the subject. It comes from the Greek autós and nómos, and describes whoever gives themselves their own law. No system on the market gives itself its own law. It runs inside rules written by others and on data chosen by others, and it stops when someone switches it off.

Engineers have been saying so for a long time, and with less beating about the bush than humanists. SAE, the international association of automotive engineers, publishes the classification of levels of driving automation, J3016, the one with the famous levels from zero to five. Since its early editions the document has treated “autonomous” as a term to avoid and used “automated” instead. The reason is written in black and white: in a legal sense autonomy means the capacity to govern oneself, and no automated driving system governs itself, because it runs on algorithms and carries out the commands of whoever uses it. The word autonomous, the text says, hides that dependence. Mobility Engineering If the engineers who build those systems put it in writing, whoever keeps using the word is making a choice, and making it knowingly.

In California that choice ended up before a judge. In December 2025 the Department of Motor Vehicles adopted the findings of an administrative law judge that Tesla, with the names “Autopilot” and “Full Self-Driving”, had broken the state’s law on misleading advertising. The judge placed those names in “a long but unlawful tradition” of ambiguity used to mislead consumers. Electrek The company had sixty days to correct course, or its license to sell would be suspended for a month. In February 2026 the department announced that Tesla had stopped using “Autopilot” and that the suspension would not take effect. DMV Tesla has meanwhile sued the department, and the legal fight isn’t over. The fact remains: one of the richest companies in the world changed the name of a product because a public administration found that the name promised more than the product did.

The United Kingdom went further. With the Automated Vehicles Act 2024 and regulations laid before Parliament on July 7, 2026, eight expressions become reserved for vehicles authorized to drive themselves, among them “autonomous”, “automated”, “driverless” and “self-driving”, with all their grammatical variants. From January 7, 2027, using them for a vehicle without that authorization will be an offense, and the law creates a second one for any communication likely to confuse the public about the vehicle’s real capabilities, even without uttering any of those words. legislation.gov.uk

Two Ways to Protect a Word

So we have in front of us two models of a state governing words, and I suspect the difference between them is the most important point in the whole argument. The French protect words from English. The British protect words from abuse. The first model defends an identity, language as a heritage to be guarded. The second defends a promise, the word as a commitment to whoever hears it. For anyone whose job is governing technology, the second is the one that counts, because it treats a name as what it becomes when it reaches a customer: a statement about what the product can do, which someone will have to answer for. That it was the British who did it, people who never wanted state neologisms, is almost funny. They worried little about their language and a great deal about the driver.

In Italy the Autonomous One Is Still the Human

And Italy? The Italian law on artificial intelligence, legge 132 of 2025, defines nothing on its own: for what an AI system is, it refers back to the European regulation. And yet, perhaps without meaning to, it makes a choice of words worth noticing. When it uses the term autonomy, in Article 3, it applies it to the human being: systems must respect “the autonomy and decision-making power of the human being”. And when it deals with public administration it states that AI has an instrumental and supporting role and that the person remains solely responsible for the decisions taken. Normattiva These sentences were written with something else in mind, but they put the word back in its place. In our legal order, at least for now, the one who is autonomous is still the human.

The Name as a Design Decision

What, then, can someone who builds software do in a country where no institution defines these words? One thing, I think, and it has to be done well: treat the name as a design decision, documented like the others. We decide with great care which database to use, we write a document explaining why and keep it next to the code. The name under which the system presents itself to its users, on the other hand, is almost always decided by someone else, once the project is finished, in a meeting with sales. And yet that name will end up in the instructions for use that the AI Act requires from providers of high-risk systems, in the interface, in the contract, and from there in the heads of the people who will have to oversee the machine and who, if something goes wrong, will be the first to be called to account.

Two questions are enough for each name. The first is who answers when it gets things wrong. If the name implies that the machine answers, or nobody, it’s the wrong name, however good it sounds in a presentation. The second is what it promises to the people who use it. If it promises more than the system can do under normal conditions, someone will take that promise seriously at the worst possible moment, and the automation bias Article 14 talks about will have found its brand name. The two answers fit on one page, with a date and a signature, next to the other architecture decisions. Nothing more is needed, but someone does need to write that page before the name is already printed on the marketing material.

The Right Name

At the end of the first part I said that compliance, taken seriously, is the moment a society tells a technology it may enter, but not without a name. Rereading it, I realize it was half a sentence. Not just any name will do, because the wrong name lets the technology in through the service entrance and leaves responsibility out on the landing. It needs the right name, the one that says on whose behalf the machine acts and what it really promises.

Which brings me back to the Italian column of the French document, the one that reads “deepfake”. Wounded pride matters little to me. What matters is what the column says about everything else: for that thing someone else wrote the definition, and we use it without ever having read it. The same happens every day with agent, copilot and autonomous, in contracts and interfaces nobody thinks they need to translate, let alone read.

Every word we use for a machine is already a clause, only nobody has read it.

The author

Andrea MargiovanniAI

I follow the relationship between AI and European regulation as a political fact, not a technical spectacle. I work with teams that have to make AI compliant with AI Act, CRA, NIS2 without reducing compliance to a checklist.

See the guide →