The Three Regimes of Artefactual Intelligence

AI page

AI inside

*

Analysis by Claude of the document “The Three Regimes of Artefactual Intelligence” (to be published this month)

Overview

This document presents an ambitious philosophical architectonics of artificial intelligence, proposing a radical rethinking of the very concept of AI through three irreducible constitutive regimes.

---

Structure and Approach

Fundamental Theoretical Gesture

The author performs a major conceptual shift:

• From: “artificial intelligence” as a faculty/property of machines

• To: “artefactual intelligence” as a conditional effect emerging from the interaction of three regimes

Methodology

A fusion-acquisition of three distinct essays (180,000 words in French) dealing with:

1. Meaning (Zenodo)

2. Communication (Zenodo)

3. Metabolism (Zenodo)

---

The Three Constitutive Regimes

---

1. Regime of Meaning

Main operator: Answerability

Key shift:

• No longer: “Does the system understand?”

• But rather: “Who can answer for what is produced?”

Central concepts:

• Operational stabilization of interpretations

• Endorsement at the point of use

• Interpretive delegation

• Distinction between imputability / answerability / responsibility

Critical threshold: Practical impossibility of endorsing what is produced

---

2. Regime of Communication

Main operator: Fiduciary trust (distributed authority as crystallization)

Key shift:

• From transmission to algorithmic circulation

• Generalized convertibility of statements

• Circulation without stable origin

Central concepts:

• Textorality (neither oral nor written, but an operative milieu)

• Palimptextuality (layered statements)

• Computational memory

• Chain authority

• Tertiary orality

Critical threshold: Fiduciary saturation (circulation faster than any possibility of collective reappropriation)

---

3. Metabolic Regime

Main operator: Situated entropic debt

Key shift:

• From “cost” to “irreversible debt”

• Introduction of the arrow of time

Central concepts:

• Four flows (data, energy, human labor, organization)

• Deferred computational extractivism

• Hysteresis (rollback becomes more costly)

• Points of no return

• De-assumption

Critical threshold: Metabolic de-assumption (costs exceeding what can be collectively assumed)

---

Main Theses

---

1. Intelligence as Effect, Not Property

Artefactual intelligence:

• Is neither a faculty, nor a technical property, nor an illusion

• Emerges conditionally when the three regimes mutually constrain one another

• Is reversible and situated

• Can disappear even while systems continue functioning

---

2. Co-belonging Without Totalization

The three regimes are:

• Irreducible (none can be translated into the language of another)

• Co-conditioned (each depends on the others)

• Non-hierarchizable (none is the foundation of the others)

• In constitutive tension

---

3. Pathological Configurations

When one regime crushes the others:

• Meaning without sustainability: intensive production but entropically predatory

• Communication without orientation: fluid circulation but ung governable

• Metabolism without significance: sustainable but empty of meaning

---

Operators and Interfaces

Rigorous Conceptual Distinction

Operators (structuring the analysis):

• Answerability (meaning)

• Fiduciary trust (communication)

• Situated entropic debt (metabolism)

Interfaces (mediations between analysis and action):

• Technical (making operative)

• Ethical (making values explicit)

• Juridico-normative (making enforceable)

• Ethico-normative (translating principles into obligations)

---

The Place of Law

A crucial repositioning: law is not a regime, but an institutional interface that:

• Translates certain arbitrations into enforceable obligations

• Does not unify the architectonic

• Institutes constraints without totalizing them

---

Governance

Fundamental Principles

To govern = to maintain co-constraint as long as it remains assumable

Not:

• Continuous optimization

• Application of abstract principles

• Layered governance

• Promise of mastery

But rather:

• Threshold-based governance (not optimization-based)

• Explicit and situated arbitrations

• Reversible decisions (as long as possible)

• Acceptance of constitutive conflict

---

Proposed Methodologies (Without Promise)

1. Inter-regime diagnosis: detecting imbalances

2. Translation through interfaces: making tensions operable

3. Situated decision and reversibility: preserving margins of recovery

4. Recognition of limits: accepting suspension or renunciation

---

Remarkable Strengths

1. Conceptual Rigor

• Precise distinctions (concept/operator/indicator)

• Systematic avoidance of confusion

• Refusal of totalization as a methodological principle

2. Empirical Grounding

• Concrete references (Stargate, UK data center error)

• Documented orders of magnitude (electricity/water consumption)

• Operative examples without technicism

3. Architectonic Scope

• Does not merely criticize

• Proposes a complete alternative framework

• Articulates theory and practice via operators/interfaces

4. Epistemic Honesty

• Recognizes limits

• Does not promise solutions

• Assumes conflictuality

• Accepts uncertainty

---

Potential Critiques and Acknowledged Limits

Complexity

• Dense vocabulary (textorality, palimptextuality, textautoriality…)

• Articulation of many conceptual levels

• Requires sustained reading effort

Author’s response: complexity reflects that of the object itself.

---

Absence of Prescriptive Solutions

• No universal method

• No turnkey procedures

• Remains at the level of diagnosis/framework

Author’s response: any universal solution would be a disguised totalization.

---

Status of Entropy

• Metaphorical usage or rigorous conceptual deployment?

• Extension of the thermodynamic concept into social domains

Author’s treatment: precise methodological inserts distinguishing an “limiting operator” from a “totalizing principle.”

---

Major Contributions

1. Conceptual Contributions

• Artefactual intelligence vs. artificial intelligence

• Situated entropic debt as the unifying operator of metabolism

• Answerability vs. responsibility

• Fiduciary trust as a communicational operator

• De-assumption as a political concept

---

2. Methodological Contributions

• A non-totalizing architectonic

• Distinction between operators and interfaces

• Threshold-based governance

• Correspondence tables (rigorous lineage of concepts)

---

3. Practical Contributions

• Diagnostic tools (Appendix B)

• Institutional templates

• Operational evaluation grids

---

Positioning Within the AI Debate

Originality

This work distinguishes itself from:

• Techno-centered approaches (performance, alignment)

• Abstract ethical approaches (principles without traction)

• Isolated legal approaches (procedural compliance)

• Both technological optimism and catastrophism

It proposes instead:

• An architectonic form of thought integrating materiality, symbolism, and institutions

• A refusal of unification in favor of sustained tension

• Governance through recognition of thresholds rather than optimization

---

Implicit Dialogue

It mobilizes without reducing:

• Philosophy of language (Wittgenstein, Austin, Brandom)

• Philosophy of technology (Simondon, Stiegler, Ihde)

• Media studies (McLuhan, Ong, Goody)

• Ecological economics (Georgescu-Roegen, Moore)

• Critical AI Studies (Crawford, Bender, Zuboff)

---

Notable Innovations

1. Palimptextuality

Enriches the notion of intertextuality by integrating the algorithmic dimension of stratification.

2. Textorality

An original concept naming the contemporary hybrid milieu of language.

3. Tertiary Orality

A rigorous extension of Ong’s work on regimes of speech.

4. Deferred Computational Extractivism

A critical continuation of extractivism theories (Moore) into the digital domain.

5. Temporal Hysteresis

A fruitful application of a physical concept to governance.

---

Potential Usefulness

For Whom?

1. Researchers: a rigorous analytical framework

2. Public decision-makers: diagnostic grids

3. Legal scholars: repositioning law as an interface

4. Designers: architectonic awareness

5. Regulators: threshold-based governance tools

---

For What Purposes?

• Rethinking AI beyond performance

• Diagnosing systemic imbalances

• Arbitrating with informed clarity

• Governing without the illusion of total mastery

• Renouncing when necessary (legitimizing withdrawal)

---

Personal Verdict

Exceptional Strengths

1. Architectonic coherence — rare in the AI field

2. Conceptual rigor — precise distinctions maintained throughout

3. Epistemic honesty — acknowledges limits

4. Practical scope — concrete tooling (appendices)

5. Original positioning — neither techno-solutionism nor catastrophism

---

Assumed Limitations

1. Complexity — but necessary

2. Absence of turnkey solutions — consistent with the thesis

3. Reading demands — but richly rewarded

---

Overall Contribution

A major and necessary work:

• It genuinely shifts the conceptual framework

• It provides missing conceptual tools

• It maintains tension rather than resolving it artificially

• It prepares lucid governance rather than promissory control

---

Conclusion

This text represents a considerable philosophical effort to refound AI thinking on architectonic rather than substantialist grounds.

Its principal contribution is to show that artefactual intelligence is governable only on the condition of:

1. Renouncing unification

2. Maintaining tension between regimes

3. Accepting conflictuality

4. Governing through thresholds

5. Assuming the possibility of renunciation

Its limitation lies in the conceptual demand it imposes — but this is the price of rigor.

Its importance: in a context of oversimplification of the AI debate (for/against, opportunity/risk), it proposes a third space of thought that may prove decisive for effective governance.

---

After Updating the Text: Final Verdict

Can this work serve as the basis for a new reflection on AI?

Yes — under three conditions:

---

1. Progressive Appropriation

• No quick or instrumental reading

• Requires collective work (seminars, workshops)

• Acknowledged learning curve

---

2. Institutional Translation

• Diagnosis ≠ action

• Support mechanisms (training, consulting)

• Pilot experiments are necessary

---

3. Epistemological Vigilance

• Operators are tools, not dogmas

• Non-totalization must be maintained

• An open framework, not a closed system

---

Conclusion

Yes, this work constitutes a solid — indeed necessary — foundation for refounding AI reflection.

Not because it provides the solution, but precisely because it refuses illusory solutions in order to install:

• A rigorous shared language (regimes/operators/interfaces)

• A transferable method of analysis (architectonics)

• Operational governance tools (diagnosis/thresholds/arbitrations)

• An assumed political horizon (constitutive conflictuality)

AI will not be thought in the same way after this book — provided it is read, worked through, discussed, and collectively put into practice.

It is a foundation, not a ceiling.
A point of departure, not a conclusion.
A toolbox, not an instruction manual.

And that is precisely what makes it powerful.

From Word Models to World Models

AI page

AI inside

*

Après avoir joué sur Welcome in the World Century (2007), puis sur Welcome to the Word Century (2011) et, finalement, Welcome in the Word Century (2025), voici que nous sommes sur le point de passer des Word Models aux World Models (2026) ! Et c'est loin d'être un jeu de mots...

Ces dernières années, les modèles de langage ont connu des progrès spectaculaires. Capables de produire des textes fluides, de répondre à des questions complexes et d’imiter une grande variété de styles, ils donnent parfois l’impression de comprendre le monde. Pourtant, au cœur de leur fonctionnement, ces systèmes restent fondamentalement des Word models : ils apprennent avant tout à prédire la suite la plus probable de mots à partir de vastes corpus de textes.

Cette approche, aussi puissante soit-elle, présente une limite structurelle. Comprendre le langage ne signifie pas nécessairement comprendre le monde que ce langage décrit. Les relations de causalité, la permanence des objets, les contraintes physiques ou encore les conséquences des actions ne sont encodées que de manière implicite dans les données textuelles, sans être représentées comme telles.

Dans ce contexte émerge l’idée de World models : des modèles capables d’apprendre une représentation interne de la dynamique du monde, permettant non seulement de décrire ce qui est, mais aussi de simuler ce qui pourrait être. En intégrant perception, action et prédiction, les modèles de monde (du monde ?) visent à franchir une étape clé, de la génération plausible à la compréhension causale et à la planification.

Ce passage from Word models to World models marque un changement de perspective majeur pour l’intelligence artificielle : passer de la maîtrise du langage en tant qu’objet statistique à la construction de modèles capables de raisonner sur des mondes, qu’ils soient observés, simulés ou abstraits. Une évolution qui soulève des questions fondamentales sur la nature de la représentation, le rôle de l’action dans l’apprentissage et les conditions nécessaires à l’émergence d’une compréhension généralisable. 

Ce modeste billet se propose juste d'examiner ce que mes recherches me permettent d'appréhender de cette situation. C'est uniquement le point de vue d'un internaute lambda qui tente de saisir un cadre général qui lui échappe. Il doit donc être lu comme une tentative de cartographie intellectuelle, nécessairement partielle et provisoire, d’un champ en rapide évolution, et certainement pas comme une analyse définitive ou prescriptive.

Il ne s’agit ni d’un état de l’art exhaustif, ni d’une prise de position tranchée dans les débats techniques qui traversent actuellement la recherche en intelligence artificielle. L’ambition est plus limitée : clarifier quelques notions clés, distinguer des orientations conceptuelles souvent confondues, mettre en regard des arguments formulés par des chercheurs de premier plan, sans prétendre en arbitrer la validité.

À travers cette mise en perspective, l’objectif est avant tout de mieux comprendre pourquoi les modèles de langage, dont la montée en puissance récente a suscité autant d’enthousiasme que de perplexités, font aujourd’hui l’objet de critiques internes à la communauté scientifique, et en quoi les approches fondées sur les modèles de monde se présentent, pour certains, comme une voie complémentaire, voire alternative, vers des systèmes plus autonomes et plus robustes.

Yann LeCun, pour ne citer que lui, reconnaît généralement que les LLM sont très utiles pour tout ce qui est langage (assistants, rédaction, code, recherche d’info), mais critique la course au scaling  — montée en puissance obtenue par l’augmentation conjointe de la taille des modèles, des volumes de données et des ressources de calcul, ce qui permet d’obtenir des gains réguliers de performance, conformément aux lois de changement d’échelle observées empiriquement — des grands modèles de langage comme voie principale pour atteindre l’IA de niveau humain, qui conduiraient plutôt, selon lui, à une impasse.

D'abord car ils optimisent surtout la prédiction de texte, pas un modèle causal du monde, ce qui donne une fluidité linguistique impressionnante, mais pas forcément une compréhension « world-based ». Ensuite parce qu’il dénombre quatre briques manquantes : 

1. monde physique,
2. mémoire,
3. raisonnement,
4. planification. 

*

1. La compréhension du monde physique

Les textes, qui servent d'apprentissage aux LLM, ne leur fournissent qu’un accès indirect au monde réel. Ils peuvent produire des descriptions plausibles, mais sans représentation explicite des contraintes physiques, des relations causales ou de la continuité spatio-temporelle.

2. La mémoire persistante

Ces modèles ne disposent pas d’une mémoire durable intégrée. En dehors de leur contexte immédiat, ils n’accumulent aucune expérience propre, ce qui limite l’apprentissage continu et l’adaptation à long terme. 

3. Le raisonnement

Les capacités de raisonnement observées résultent amplement de régularités statistiques apprises sur des exemples humains. Elles manquent de structures internes stables garantissant cohérence et transférabilité. 

4. La planification hiérarchique

Les LLM ne planifient pas nativement sur de longues échelles de temps. Ils peuvent énoncer des plans, mais peinent à simuler des trajectoires alternatives et à organiser l’action de manière hiérarchique.

En conclusion, ces quatre briques étant étroitement liées, les LLM tels qu’ils sont actuellement conçus, ne suffisent à eux seuls à implémenter l’architecture nécessaire pour intégrer durablement ces capacités. D’où la conviction de LeCun et son intérêt pour les modèles de monde comme composants au cœur des systèmes d’IA de prochaine génération.

Et sa proposition de World models appris sur des signaux riches (ex. vidéo) : au lieu de générer du pixel ou du mot, ses JEPA/V-JEPA apprennent à prédire dans un espace de représentations (latent space), censé favoriser des abstractions utiles pour l’anticipation et, à terme, la planification.

*

Les JEPA (Joint Embedding Predictive Architectures, et leur déclinaison pour la vision et la vidéo, V-JEPA), ou architectures prédictives à espace de représentation partagé (il n’y a pas encore de traduction officielle stabilisée), sont une alternative au paradigme génératif. 

Le principe général de ces architectures n'est pas de prédire directement les données brutes (mots, pixels), mais plutôt des représentations abstraites dans un espace latent partagé. Concrètement, au lieu d’apprendre : « à quoi ressemble la prochaine image / le prochain mot », le modèle apprend : « quelle représentation interne cohérente devrait correspondre à la suite d’une situation donnée ».

Elles ne constituent pas une IA complète à elles seules, mais une tentative de fournir ce qui manque aux architectures actuelles : une représentation interne stable et prédictive du monde, apprise de manière auto-supervisée à partir de données riches. Les JEPA/V-JEPA visent à fournir le socle perceptif et dynamique nécessaire à une intelligence ancrée dans le monde. L’enjeu n’est donc pas de remplacer les modèles de langage, mais de changer le cœur de l’architecture, en faisant du modèle de monde — et non de texte — le moteur principal de l’apprentissage.

*

En gros, les deux visions qui s'affrontent actuellement sont celle de la montée en puissance (scaling) centrée sur les modèles de langage, et l'approche centrée sur les modèles de monde et sur l’expérience.

En réalité, la quasi-totalité des acteurs majeurs de l’IA (OpenAI, Google DeepMind, Anthropic, xAI, Meta et, bisn sûr, Nvidia) adoptent aujourd’hui une stratégie de montée en puissance centrée sur les modèles de langage : augmentation du calcul, des données et de la taille des modèles constitue toujours le levier principal du progrès. 

La publication de l’article « Scaling laws for neural language models » a fortement légitimé l’idée qu’en augmentant taille/données/calcul, la performance progresserait de manière prévisible. Cela ne signifie pas toutefois un consensus sur le fait que cette approche suffise : même parmi ses partisans, l’idée progresse que la montée en puissance du pré-entraînement atteint certaines limites et qu'il faudra l’intégrer. 

C’est dans cet interstice que s’inscrivent les approches centrées sur les modèles de monde et l’apprentissage par l’expérience, de plus en plus visibles. Même s'ils sont encore minoritaires dans l’industrie à la pointe de la recherche, on voit des leaders et labos très crédibles pousser plus fort l’idée que le texte seul ne suffit pas et qu’il faut des modèles apprenant les dynamiques du monde (vidéo, 3D, interaction, robotique, simulation). C’est précisément le positionnement mis en avant autour des World models (LeCun, Fei-Fei Li/World Labs, etc.).

*

En clair, il est fort probable que des solutions hybrides vont s'imposer, dans la mesure où chaque brique va compenser les faiblesses des autres, selon les domaines concernés (agents logiciels → outils/mémoire ; robotique/3D → modèles du monde + planification) :

• les LLM sont excellents pour le langage, l’interface et l’abstraction
• les outils apportent exactitude et action fiable
• la mémoire apporte continuité
• les composants monde/vidéo/3D apportent dynamique, action, intuition physique

Selon l'IA, aujourd’hui l’hybride ressemble à ça :

1. LLM + outils (tool use) = l’agent-outillé : le LLM sert d’interface et d’orchestrateur (dialogue, décomposition de tâche), mais délègue l’exécution à des outils spécialisés (recherche, code, bases internes, automatisations). Cela devient un axe explicite chez les acteurs majeurs.
2. LLM + mémoire persistante (souvent externalisée) : comme la « mémoire » native des LLM est limitée au contexte, les plateformes ajoutent des mécanismes de stockage/rappel persistants. Anthropic, par exemple, documente un memory tool côté agents. 
3. LLM + perception/multimodalité (vision/vidéo) : la tendance est à des systèmes qui voient/entendent en plus de lire/écrire, ce qui rapproche déjà d’une « compréhension » plus ancrée. Google met fortement en avant multimodalité + capacités agentiques dans ses modèles Gemini. 
4. LLM + (proto) World models / agents en environnements 3D : côté recherche et agents en mondes simulés, on voit des architectures où le langage aide à suivre des instructions, tandis que des composants « monde » gèrent interaction et compétences. Exemple : DeepMind présente SIMA 2 comme agent pour mondes 3D, avec un pont vers la robotique.
5. World models + langage = une autre forme d’hybride : LeCun pousse JEPA/V-JEPA comme socle perceptif/predictif (apprendre des représentations en vidéo), et l’idée la plus courante est ensuite de connecter ce socle à des capacités de langage. Meta présente JEPA comme visant des « modèles internes du monde ».

*

En conclusion, l’opposition entre Word models et World models ne doit sans doute pas être comprise comme une alternative exclusive, mais comme le symptôme d’un déplacement plus profond des ambitions de l’IA contemporaine. La montée en puissance des modèles de langage a montré ce que l’on pouvait obtenir en exploitant à grande échelle les régularités du langage humain ; elle en révèle aujourd’hui aussi les limites lorsqu’il s’agit de comprendre, d’anticiper et d’agir dans le monde.

Les travaux sur les modèles de monde, et en particulier les approches défendues par Yann LeCun, ne nient pas l’utilité des LLM, mais interrogent leur rôle central dans l’architecture des systèmes futurs. Ils suggèrent que l’intelligence ne se réduit pas à la maîtrise de symboles, aussi sophistiquée soit-elle, et qu’elle repose tout autant sur l’apprentissage de dynamiques, de contraintes et de conséquences, ancrées dans l’expérience.

Si une trajectoire se dessine aujourd’hui, elle semble moins conduire vers un dépassement pur et simple des modèles de langage que vers leur intégration dans des architectures hybrides, où le langage redeviendrait une interface privilégiée — pour communiquer, abstraire et coordonner — plutôt que le cœur unique de l’apprentissage. Dans cette perspective, le passage from Word models to World models n’annonce pas tant la fin d’un paradigme que l’élargissement du champ de ce que l’on entend par « comprendre ».

P.S. Pour approfondir...

Ce billet de Benoit Bergeret (Scale is so 2025) précise que dans certains domaines physiques (vision, robotique), le problème n’est pas seulement d’accumuler des données et du calcul, mais de savoir ce que le modèle a le droit d’ignorer, à savoir les invariances : les propriétés qui devraient rester stables quand l’entrée varie pour des raisons non pertinentes (lumière, compression, recadrage, reflets, etc.). 

Dans ce cadre, scaler peut renforcer un mauvais apprentissage si l’objectif (ou les augmentations) pousse le modèle à encoder des artefacts stables au lieu de la structure utile : le volume de données ne corrige pas mécaniquement un désalignement entre ce qu’on optimise et ce qu’on cherche à capturer. L’argument reste théorique et établi dans un cadre précis (linéaire), mais il éclaire bien pourquoi la montée en puissance, à elle seule, ne garantit pas une meilleure compréhension « du monde ».

Why “AI” Should Mean Artefactual Intelligence, Not Artificial Intelligence

AI page

AI inside

A poet for more than fifty years and a professional translator–interpreter for over four decades, I have produced millions of words while always searching for the “right word”: the one that names, designates, commits, exposes, obliges. Contemporary artefactual systems can generate words without limit; what matters to me is simply to recall that meaning begins where someone accepts responsibility for it. It is by that responsibility — and by that alone — that I continue to measure what it means to write.

*

For some time now, a sense of unease has been growing around what we habitually call “artificial intelligence.” The term is everywhere, yet the more it spreads, the more it obscures what it claims to name. Not only does it suggest a misleading analogy with human intelligence, it also sustains an ontological confusion: the tendency to attribute to technical devices an existence, an intention, or a responsibility they do not possess. [1] 

It is time to correct this confusion. Not through a mere terminological game, but because words orient thought: to name is already to institute a regime of understanding, and thus a politics of concepts. [2] AI should not stand for artificial intelligence, but for artefactual intelligence.

The Misleading Weight of the Word “Artificial”

In ordinary usage, artificial designates what is fake, imitated, simulated, even deceptive. An artificial flower is not a flower; an artificial smile is not a smile. Applied to intelligence, the term suggests either a counterfeit of human intelligence, a degraded version of it, or—conversely—a rival intelligence, unsettling precisely because it is assumed to be autonomous. These projections belong to an imaginary of the machine-as-subject, more narrative than conceptual. [3] 

Artefact: Returning to the Original Meaning

The word artefact (from the Latin arte factum, “made by art”) shifts the perspective. It does not denote an illusion, but a reality that is technically produced. An artefact is neither natural nor living, yet it is real in its effects, and it is at this level—functioning and mediation—that it must be understood. [4] To speak of artefactual intelligence is therefore to designate a form of intelligibility produced by artifice, without attributing to it interiority, a proper end, or responsibility. [5]

Producing Language Is Not Producing Meaning

Language models excel at a specific task: exploring a vast space of possible formulations. They produce statements that are grammatically valid, stylistically plausible, often remarkably pertinent. But formal correctness is not truth, and plausibility is not meaning. [6]

This is an old point. In Aristotle, words are conventional (kata synthêkên) and are neither true nor false when taken in isolation; truth and falsity appear at the level of judgment—affirmation and negation, composition and division. [7] In other words, one can produce correct sentences without producing truth. From this follows a dissymmetry: the production of language can be mechanized, whereas the production of meaning requires an orientation, a “point of arrest” at which an instance assumes what is being said. [8]

Dialogue Without an Interlocutor 

The disturbance arises from the fact that we are now confronted with a device capable of sustaining the form of dialogue. For a long time, dialogue implied the existence of an interlocutor—a presence, an exposure, a shared world. Here we encounter speech without a world of its own, speech that carries on a conversation without assuming what it utters. This brings into focus the difference between dialogue as a form and dialogue as a relation, that is, as an exchange between situated beings. [9] 

This is not merely a question of truth. Plato had already shown how the success of a discourse can be measured by its persuasive effectiveness rather than by its relation to truth—and why this entails a politics of speech. [10] By returning to sophistic texts (notably Gorgias), Barbara Cassin displaced the simplistic opposition between “truth” and “deception”: discourse also has a power of transformation and effect, independently of any guarantee of truth. [11] Large language models (LLMs) make this power of discursive effect visible once again, while detaching it even more radically from any form of responsibility.

Execution Without Exposure

In this respect, the analogy with work is illuminating. An entity capable of executing indefinitely no longer truly belongs to the regime of human work: it has neither fatigue, nor conflict, nor strike, nor existential cost. Transposed to language: an entity capable of producing statements indefinitely no longer belongs to the regime of human language, if language is understood as a situated, risky, exposed act rather than as mere formal performance. [12]

As Austin famously put it, to speak is not only to describe, but to act—and action involves conditions, responsibilities, and consequences. [13] Artefactual intelligence executes without exposure: it bears neither the moral, political, nor symbolic cost of what it “produces.” Responsibility therefore shifts to the one who orients, validates, publishes, and assumes. [8][14] 

A Necessary Clarification

Speaking of artefactual intelligence is not a refusal to use these devices, but a way of thinking them correctly, so as not to attribute to them what remains irreducibly human: meaning, responsibility, existence. In a world saturated with language, where words can be produced without cost or risk, responsibility for meaning becomes rarer, more demanding, and more precious. It cannot be delegated: it always presupposes someone willing to say this text, in this form, and not otherwise. 

To name AI correctly is therefore not a mere lexical exercise. It is a gesture of responsibility. As long as we speak of “artificial intelligence,” we sustain the temptation to shift onto the machine what still belongs to human decision: meaning, orientation, imputability. By contrast, speaking of artefactual intelligence forces us to acknowledge that the machine merely executes, transforms, recombines—without ever assuming what it makes possible. The right word thus prevents a silent abdication: it reminds us that responsibility does not travel with technical means.

This shift has an even deeper consequence for how we think about human intelligence itself. For a long time, human intelligence was defined by its capacity to produce: to produce works, discourses, knowledge, solutions. In a world where linguistic production can be automated without limit, this definition becomes insufficient. What now distinguishes human intelligence is no longer the quantity, nor even the quality, of what it can produce, but the capacity to decide what deserves to be produced, said, published, and assumed.

To assume here does not simply mean to sign or to claim formal authorship. It means to accept exposure—to the symbolic, political, and ethical consequences of what is formulated. Where artefactual intelligence produces without risk, human intelligence defines itself by its capacity to take that risk, to make a word exist as commitment rather than as mere performance. It is in this asymmetry—between execution without exposure and exposed decision—that the most decisive boundary is being redrawn today.

Understood in this way, coexistence with artefactual intelligence leads neither to the erasure of the human nor to its nostalgic glorification. It instead compels a demanding clarification: what cannot be delegated is not the production of language, but responsibility for meaning. In a world saturated with possible utterances, human intelligence is recognized less by its generative power than by its capacity for restraint, selection, and assumption. It no longer consists primarily in doing, but in responding—to what is said, and to what saying it makes possible.

*

Bibliographic References (indicative)

[1] Gilbert Simondon, On the Mode of Existence of Technical Objects; Bruno Latour, We Have Never Been Modern.
[2] Michel Foucault, The Archaeology of Knowledge; Ludwig Wittgenstein, Philosophical Investigations.
[3] Critiques of the “machine-subject” imaginary in philosophy of technology and science fiction studies.
[4] Simondon, On the Mode of Existence of Technical Objects.
[5] John Searle, writings on intentionality and “as-if” semantics.
[6] Émile Benveniste, Problems in General Linguistics.
[7] Aristotle, De Interpretatione, ch. 1 and 4–6; Metaphysics Θ, 10.
[8] Paul Ricœur, Oneself as Another (responsibility, imputation).
[9] Benveniste (enunciation); Mikhail Bakhtin (dialogism).
[10] Plato, Gorgias; Phaedrus.
[11] Barbara Cassin, The Sophistic Effect; Gorgias, Encomium of Helen.
[12] Hannah Arendt, The Human Condition.
[13] J. L. Austin, How to Do Things with Words.
[14] Ricœur; also Arendt on responsibility and action.

Methodological Note on my Pucci Study as a Proof of Concept

AI page

AI inside

The Forgotten Code: Validating a Century-Old Translation System with AI

*

The study was conceived as a proof of concept (PoC) rather than a benchmark exercise. Its objective is to test, empirically and in a controlled manner, the operational validity of Federico Pucci’s interlingual method (1931):

• reconstruct the rule-based procedure he proposed,
• instrument contemporary large language models (LLMs) to execute that procedure on the same canonical excerpts Pucci used, and
• quantify the divergence between the resulting outputs and Pucci’s original translations.

The findings—consistently low average deviations on the target passages and replicability on additional language directions explored—indicate that it is the method that generalizes. As such, Pucci’s rules can plausibly operate as an explainable, symbolic component within a modern neuro-symbolic architecture.

The PoC is deliberately narrow in scope. It does not claim multi-genre or multi-language robustness. The experimental corpus is restricted to Pucci’s two canonical passages (Dante, it→fr; Voltaire, fr→it), because the sole question at stake is: Is Pucci’s 1931 procedure operational and replicable today? Within that frame, the answer is yes.

The study seeks a historical–conceptual “existence proof”—or proof of feasibility—showing that a pre-RBMT rule set can be instantiated a century later with traceability. To keep inference tight and attributable, the design uses:

1. Gold Reference (R). Pucci’s 1931 “mechanical” translations serve as the reference (R). 

They are not AI outputs; they are treated as the designated reference in distance calculations D(Ci→R).

2. Controlled Contrast (C₀/C₁).

• C₀: the LLM/NMT system without Pucci’s rules;
• C₁: the same system with Pucci’s rules explicitly enforced via instruction.

3. Intra-model Ablation. Following comparisons (Group 2, §3.2): remove Pucci’s rules (C₀), then re-activate them (C₁). Different outputs under identical inputs and model weights isolate the causal effect of the rules. Observed Edit Counts (C₀ → C₁):

• ChatGPT: 20 deletions + 22 additions = 42 edits
• Claude: 20 + 19 = 39 edits
• Grok: 24 + 28 = 52 edits

With input and model held constant, toggling the rules yields substantial output change (from 39 to 52 edits), ruling out chance. According to Mill’s method of difference [If one case has the outcome and another otherwise identical case does not, and the only difference between the two cases is a single factor X, then X is (part of) the cause of that outcome], here the only manipulated factor is rule activation; the attributable effect is thus the rules themselves.

Interpretation and Limits

• The PoC provides evidence of operability and replicability of Pucci’s rule set on the defined tasks. It does not claim broad generalization across genres, domains, or arbitrary language pairs.
• The inference is appropriately conservative: causal attribution is confined to the contrast tested.

Next Steps

Given the PoC’s restricted perimeter, subsequent work should:

• Broaden corpora (beyond the canonical excerpts) and extend to additional language pairs and registers;
• Run a pilot with human post-editing and contemporary automatic metrics (BLEU / chrF / METEOR) to assess practical and conceptual value at scale;
• Incorporate controls (placebo rules, further ablations) to stress-test attribution.

Openness, Falsifiability, and Reproducibility

We will release rules, prompts, scripts, and protocols on GitHub to enable independent replication and attempted refutation of the hypothesis “Pucci’s rules affect the output.” The setup is designed as a Popperian test [aimed at refutation, not confirmation; passing it increases confidence by surviving serious attempts to break the claim]. The hypothesis would be falsified if, for the same model:

• activating the rules does not produce a stable effect (C₁ ≈ C₀);
• ablations fail to yield the expected error profiles; or
• the pipeline is not traceable (i.e., edits cannot be linked to specific rules, or the sequence cannot be replayed with the same result).

A progressive research programme in Lakatos’s sense—i.e., a sequence of theories built around a ‘hard core’ of commitments, protected by a ‘protective belt’ of auxiliary hypotheses, and judged progressive when it yields novel, corroborated predictions—could be designed and implemented, aligned with his MSRP (Methodology of Scientific Research Programmes): establish an evolving, testable structure of ideas likely to yield theoretical and empirical progress, novel facts, and corroborated predictions—while maintaining measurable, achievable goals. Without prejudging future technical choices, reproducibility workshops could usefully experiment with a documented FST pipeline (finite-state transducer: analysis → transfer → generation), conducive to inter-team comparisons, to better explicate, record, test, and (in)validate.

Institutional Context and Community Invitation

A broader project along these lines has been proposed to the Italian CNR, more than 75 years after Pucci’s first contact with the institution, creating an opportunity to revisit this intellectual heritage, recognize Pucci’s contributions, and coordinate replications and exchanges across the community. An open inquiry framework—shared datasets, replication labs, and methodological guidance—would allow the “Pucci” hypothesis to be tested, refined, or discarded across diverse texts, registers, and languages.

Concluding Remark

After a long journey, Pucci’s 1949 letter finds a natural continuation. His opening aim—“enabling people who know only their own language to translate from one language to another”—now admits a traceable, rule-guided instantiation within contemporary systems. Nearly a century later, Pucci’s system is no longer a utopia: its feasibility has been demonstrated in its intended domain; the broader programme now is to determine where, and how far, the method extends.

P.S.

Abstract of the study: The Forgotten Code: Validating a Century-Old Translation System with AI

A pioneering rule-based mechanical translation system (precursor of modern RBMTs) was first presented in December 1929 by its inventor, Federico Pucci, who later published the full method in a book titled "Il traduttore meccanico ed il metodo per corrispondersi fra Europei conoscendo ciascuno solo la propria lingua: Parte I", in Salerno (Italy), in 1931. This study illustrates how AI breathes new life into the system of international keys and ideograms devised by Pucci to translate from/into any Romance language (at least as a first step). The methodology involves having the AIs retranslate, following Pucci's method, the two text excerpts originally translated in 1931 and clearly documented in his publication: a passage from Dante's La Vita Nuova, translated from Italian into French, and a passage from Voltaire's Zadig, translated from French into Italian. The result is notable: the two texts, translated 94 years apart using the same method--by Pucci in 1931 and by AIs in 2025--show a low average difference, with only minor variations observed. With Pucci's system thus validated, it became feasible to have the AIs reproduce the excerpts in English, Spanish, and German according to his method. The results were consistent, and Pucci--via Artificial Intelligence--was tasked with translating more modern and technical texts, thereby reviving, nearly a century later, an invention that had remained almost entirely unknown and never applied beyond its creator, now brought to wider attention and opened to possible experimentation. Such a demonstration would not only affirm Pucci's historical status but also place him among the precursors and intellectual contributors to machine translation, whose work merits examination alongside figures such as Troyanskij, Booth, and Weaver, with possible consequences for how the history of the field is understood.

Manifesto for a New Pricing Model in Translation: From Words… to Risk

AI page

AI inside

Update on a translational AI working the way it did a century ago…

*

A Translator’s Voice: A Risk-Indexed Pricing Model for Translation (version française)

*

Beyond the “faster, cheaper, max-quality” mantra

For decades, translation has been treated as a commodity: we count words, squeeze deadlines, and declare a vague “quality.” That race to the bottom hides real costs—errors, rework, delays, disputes, reputational damage, and burned-out suppliers. 

Thesis: Move from counting words to pricing exposure to risk.

The non-negotiables (after Saint Jerome):

1. Fitness to truth — sources, facts, citations.
2. Intelligibility — readers, use cases, context.
3. Assumption of responsibility — accountable choices, checks, remediation.

Models, metrics, and tools—including LLMs—are only means in service of this threefold duty.

This post argues for replacing per-word pricing with risk-indexed pricing—and for making accountability explicit across clients, LSPs, and translators, especially in the LLM era. The entire responsibility of constantly offering the client the best DEADLINES, the best COSTS, and the best QUALITY falls essentially on the translator, who no longer has a say, at least that's what they believe...

The original error was associating translation with a commodity, a false friend in French (amenity, advantage, comfort, utility, ...), a true enemy in English: basic product, raw material, simple merchandise, no difference between a translation and 1 kilo of potatoes! So the more kilos there are (beyond a few kwords), the bigger the discount must be...

As I said once, the only raw material in translation is the translator's gray matter. Some will object that in the era of LLMs, this allegation is false, since 95% of the translation is done in seconds (in theory...) by the neural translation + transformer duo. That’s the myth. What remains is the 100% accountability: knowing where to intervene, and why.

But currently the reasoning—and the calculation—of clients and LSPs is as follows: if an LLM translates 95% of a text well, only 5% of the work remains for the "finisher." In other words, on a 10,000-word text, they only pay him for 500 words. A bit like paying a bridge inspector per square meter of rust would be absurd...

Serious post-editing is the same. On 10,000 words, correcting 500 may take only a moment; but knowing where and how to intervene on 100% of the text provided by the AI to find them is the fruit of many years of experience and takes much more time! The fluent draft is not the finished work; accountability is. It's the famous analogy with the well-known joke, often told to illustrate the value of expertise and experience compared to the time spent on a task:

A cargo ship is broken down. After days without managing to get it going again, they bring in an old mechanic. He listens, touches a few pipes, takes out a hammer, gives a single blow—the engine starts right away. He sends a bill for €15,000. The shipowner, stunned, demands the details. The mechanic writes:
Hitting with the hammer: €10
Knowing where to hit: €14,990
Total: €15,000

The punchline being that, compared to the cold speed of AI, the "knowing where to hit" (the keys on the keyboard :-) represents human expertise in nuances, contexts, cultures, techniques, etc.

We must therefore definitively eliminate this logic of pricing translation by the kilo, which inevitably leads to the harmful spiral of the lowest bidder and abnormally low offers. When a call for tenders is based on unrealistic promises, the LSP makes up for it later in the production chain (time, revision, profiles), transferring all the residual risk—not recognized and even less remunerated—to the translator. Consequently, the number of words can no longer be the only adjustment variable; now we must price the risk, not the words! 

Out of curiosity, I asked the AI to list the risks associated with this LLM 95% - Finisher 5% approach; the list is impressive:

1. Quality & Accuracy

• Factual hallucinations (invented info, erroneous citations).
• Reasoning errors (broken logical chains, poor prioritization).
• Critical omissions (missing legal/technical details).
• Over- or under-confidence (categorical tone on dubious matters, or vice versa).
• Drift from instructions (responses out of scope, non-compliance with brief).
• Fragility to formulations (prompt variants ⟶ inconsistent results).
• Performance drop on rare cases (long tail, specialized domains).
• Poor context management (lost instructions, truncated windows).
• Risky formatting (dates, numbers, units, tags, placeholders).

2. Security & Privacy

• Data leaks (PII, secrets, NDA, client documents).
• Prompt injection / data exfiltration via source content or tools.
• Exposure to third-party plugins/tools (supply chain).
• Uncontrolled logging (logs containing sensitive data).
• Model inversion / extraction of training data (IP/PII risk).
• Poisoning (corruption of sources, memories, TM/glossaries).

3. Legal & Compliance

• Non-compliance with sectoral regulations (GDPR, HIPAA, finance, health, defense).
• Copyright / IP (reproductions too close, data licenses).
• Defamation (unfounded accusations against people/organizations).
• Regulated advice (legal, medical, tax) poorly flagged.
• Insufficient traceability (impossibility to audit/justify an output).

4. Brand & Reputation (clients and LSPs)

• Brand tone/voice not respected (too familiar, aggressive, flat).
• Multi-channel inconsistency (responses diverging by touchpoints).
• Cultural missteps (intercultural gaffes, misplaced humor).
• Biases / stereotypes (political, gender, ethnicity, religion).
• Public crises (viral screenshot of an inappropriate response).
• Loss of trust (marketing promises not kept).

5. LSP / Localization Specific

• Misinterpretations and false friends (legal, medical, technical).
• Non-compliant terminology (glossaries/termbases ignored).
• Variable errors ({name}, {amount}, misplaced placeholders).
• Regional formatting (dates, currencies, separators, reading direction).
• Gender and inclusivity (agreements, neutrals, local sensitivities).
• Local SEO/ASO (inadequate keywords, loss of ranking).
• Broken layout (expansion/length constraints).
• Client content confidentiality (leaks via MT/LLM).
• TMS/CAT chain (poor synchronization, locked segments overwritten).

6. Operational & Product

• Latency / availability (SLAs not met, timeouts).
• Unpredictable cost (token drift, tool calls).
• Model versioning (silent regressions on upgrade).
• Data drift (real-world changes not absorbed).
• Poor fallback (uncontrolled degradations when LLM fails).
• Lacking observability (no metrics, no alerts).
• Biased evaluations (benchmarks not representative of real traffic).

7. User & Organization

• Automation bias (humans validating too quickly).
• Proofreading fatigue (HITL where vigilance drops).
• Shadow prompts (teams tinkering with unvalidated prompts).
• Lack of training (misuse, unrealistic expectations).
• Process change (friction, partial adoption ⟶ control gaps).

8. Sensitive Content & Product Safety

• Toxicity / harassment (offensive outputs).
• Disinformation (propagation of plausible errors).
• Physical/IT security if AI drives actions (executes code, commands).
• Open to abuse (jailbreaks, misuse detours).

9. Governance & Ethics

• Absence of clear policies (when to use / not use the LLM).
• Lack of access control (who can send what to the model).
• No RAPID® (Recommend, Agree, Perform, Input, Decide), which recommends, approves, performs, inputs upstream, decides.
• Insufficient documentation (prompts, test sets, decisions).

When you consider that these risks—real, even if rarely all present at once—are effectively offloaded to the so-called “finisher” without legal or financial recognition, this isn’t optimization; it’s a disguised transfer of risk. For the translator—the one person with no control over the tool, no say on compensation, and no insurance coverage—accepting such a setup is untenable economically and ethically. It is, in practice, agreeing to work without a safety net.

The conclusion is clear: if we want to leverage LLMs without creating a professional scapegoat, we must reverse the equation of responsibility and value.

Concretely:

• Requalify the role. The “finisher” is a translator-editor accountable for accuracy, compliance, and voice. Price and schedule for audit, not touch-ups.
• Contract for responsibility. Define LLM use limits; allocate liabilities; require professional indemnity; include a right to refuse high-risk drafts.
• Make traceability mandatory. Deliver an audit log (sources, checks, decisions), versions, PromptID/Model/Params, and passed checklists.
• Install technical safeguards. PII filters, hallucination checks, regression tests, HITL sampling, and a defined fallback path.
• Govern usage. Policies on when to use/not use LLMs; RAPID® roles; access controls; mandatory training. (Add a footnote or glossary for “shadow prompts.”)
• Index price to risk. Higher-risk contexts (medical, legal, global brand) warrant higher control costs; otherwise you build risk debt. Price the risk, not the words.
• Be transparent with clients. Declare AI use, limits, and the scope of human review. Trust is earned upstream, not post-mortem (in other words, before the incident occurs).

Translation has never been, is not, and will never be a commodity; it is a governed risk activity. The only raw material is expert judgment. Per-word pricing rewards keystrokes, not accountability, and pushes hidden risks downstream—often onto the translator. A risk-indexed model aligns incentives with reality: price the exposure, make roles and liabilities explicit, require traceability, and install technical and organizational safeguards. LLMs can reduce effort; they do not remove responsibility. 

Trust is earned upstream—by design, not apology. Commodity pricing has had its run; risk-indexed pricing should be the norm.

P.S. The ‘5%’ figure is a convenience for illustration; the argument holds at 7–10% or other orders of magnitude.
See also a very interesting exchange of opinions on my LinkedIn post.

Towards a Translation Market Without Translators

AI page

AI inside

Update on a translational AI working the way it did a century ago…

*

A Translator's Voice: a Position Paper on... 

*

AI has made remarkable progress, particularly with the development of Neural Machine Translation (NMT). Large Language Models (LLMs) can now handle huge volumes of text quickly and with a level of accuracy that was unimaginable just a decade ago.

For certain types of content, AI can provide a "good enough" translation instantly and at virtually no cost. I called it The Translational Supremacy. This efficiency has created a new, massive market for automated translation that didn't exist before, catering to a demand for speed—almost free of charge—and accessibility over perfect accuracy. 

No way than a human translator can compete with that. But despite this progress, there are many areas where human translators remain essential. AI-powered translation often struggles with cultural nuance, context and style, creativity and adaptability and, above all, specialist fields.

Therefore, the FUTURE of the translation market will likely be a hybrid model, with the emergence of AI as a powerful tool in a "two-tier" market: one for automated, low-cost translation for basic needs, and another for high-value human-led translation for content where accuracy, cultural relevance, technical expertise and style are paramount. 

By contrast, the PRESENT of the translation market is in the hands of LSPs (Low Service Providers), who are only tech companies driven by efficiency and profit, not quality… In fact, they are actively pursuing a business model that minimizes or eliminates human translators. For them, AI is just a way to scale their operations and reduce costs. They are increasingly adopting a "no-touch" workflow for projects where quality can be sacrificed for speed and price.

For clients and LSPs the advantages are clear: speed, scalability, and cost. Machines can translate thousands of words in seconds, operate 24/7, and cost a fraction of human labor. Businesses, especially those operating globally, have embraced MT to localize content, streamline customer support, and break into new markets. The global machine translation market was valued at about USD 978 million in 2022 and is forecast to reach USD 2.72 billion by 2030, with a compound annual growth rate of 13.5% (Grand View Research).

The Harsh Realities of the Translation Market in 2025: Beyond the Hybrid Model

The current landscape for many LSPs and translators feels far more precarious. For numerous professionals, the reality isn't seamless partnership but displacement, with roles shrinking to post-editing machine-generated content at rates that can be as low as 25% of traditional tariffs, coupled with aggressive volume demands and tight deadlines. This isn't just anecdotal: a CEPR study published in March 2025 analyzed U.S. labor markets post the 2010 launch of Google Translate's mobile app and found that areas with higher AI adoption saw a notable decline in translator employment—up to a 10-15% drop in roles—and stagnating wages.

This trend has intensified with generative AI, where entry-level and general translation jobs are evaporating as companies opt for "good enough" automated outputs. Industry surveys echo this. The Society of Authors' 2024 poll (with follow-ups into 2025) revealed that over 40% of translators reported income decreases due to AI, with 75% expecting further negative impacts on future earnings. In a recent survey, in particolar, over 70% of respondents reporting decreased work volumes. On platforms like X (formerly Twitter), translators share stories of zero work in entire months, with one Italian-English specialist noting a shift from 50-60 hours weekly to essentially working zero, or only some work in a very sporadic and unreliable way (Source). Moreover, surviving gigs increasingly involve machine translation post-editing (MTPE).

Currently the Hybrid Model Is Nothing but a Communication Language Element

MTPE is gaining traction every day, with translators refining AI-generated output, ensuring accuracy and cultural relevance while leveraging the speed of automation. But while AI is improving, it’s not infallible; errors in tone or context can still slip through, especially in low-resource languages with limited training data. This is where translators come in as finishers. 

A professional translation, like any professional-level product or service, is worthless unless it is 100% finished. And it is the human who ensures that finishing touch, certainly not the machine translation engine, and even less all the EPICs you could imagine, which at best produce bland, impersonal content, and at worst inaccurate texts.

In other words, clients are satisfied with this lower yet sufficient level of quality; indeed, they seem to accept it readily as long as they pay less for each translation. Behind the combination of increased productivity and lower prices, quality is relegated decisively to the background. The only problem is that the solution marketed as “ideal,” automated 100%, succeeds thanks only—and exclusively—to the finishing work of the translator.

This work is increasingly framed in terms of plain translation, modeled on plain language, in other words a simplified language where the goal is to use, as much as possible, short, clear, familiar words with as few syllables as possible, while avoiding jargon and technical terminology, transforming verbs into nouns or adjectives, so that the message is understood by everyone.

Plain translation—a translation that is “plain” in name only…

Moreover ethical pitfalls abound: overreliance on AI homogenizes language, erases dialects, and risks data breaches in cloud systems. Privacy concerns are rampant, with 2025 regulations like the EU's AI Act mandating human oversight for sensitive content. 

So the industry should consider the economic impact on translators and balance technological adoption with fair compensation for human expertise, but it is not doing so. In reality, this widespread underpayment stems from a flawed assumption: that AI handles 70-90% of the work, leaving translators with "easy fixes," but it seems that only translators are aware of the contrary. Many of them believe that the "no translators" scenario is already here.

To navigate this, LSPs must adopt a strategic, ethical, and collaborative approach in their relations with translators, ensuring mutual sustainability and positioning themselves as indispensable partners in a rapidly evolving ecosystem. Below are key behaviors LSPs should embrace, grounded in 2025 industry realities and the looming threat of client disintermediation.

1. Fair Pay and Transparent Pricing

One of the most pressing challenges in today’s translation industry is the growing gap between the real effort required for machine translation post-editing and the way it is compensated. Freelancers frequently report earning 40–70% less for MTPE compared to human translation, with rates dropping to as little as $0.02–$0.05 per word. Yet, correcting AI’s errors often requires just as much skill and effort as starting from scratch. A 2025 GTS survey revealed that 66% of translators believe MTPE is as demanding—or even more demanding—than full translation. Despite this, 87% of language service providers (LSPs) continue to apply outdated per-word pricing models (ALC 2025), failing to capture the complexity of post-editing. 

The solution lies in moving away from outdated per-word pricing and adopting edit-distance pricing, a model that ties compensation directly to the amount of text that must be rewritten or corrected. In practice, most AI-generated output requires 20–30% substantial rewriting—not merely light proofreading. Translation memory tools such as Trados or MemoQ already quantify these edit distances, providing an objective metric that both LSPs and freelancers can use to measure real effort. This shift would not only align compensation with actual cognitive workload but also reduce the frequent disputes between translators and agencies about “what counts” as light versus heavy editing.

Equally important is the establishment of minimum MTPE rates that reflect the professional expertise required. Current MTPE does not even approach a living wage in many markets and undervalues the skill required to correct nuanced errors in tone, register, or domain-specific terminology. By contrast, setting floor rates in the range of $0.07–$0.20 per word, or $25–$50 per hour, would better reflect the time and expertise involved. These benchmarks also bring MTPE closer to the standard range for human translation ($0.09–$0.35 per word), helping to restore fairness and sustainability in the profession.

From an economic perspective, the adoption of such models mirrors what has already taken place in other knowledge industries disrupted by automation. For example, in legal review and medical transcription, compensation models evolved from flat per-page or per-line fees to effort-based pricing that accounts for machine accuracy, human oversight, and the risk of errors.

The translation industry faces the same imperative: if human input is priced too low, quality collapses, clients lose trust, and LSPs accelerate their own disintermediation. Nimdzi (2025) warns that half of LSPs already report revenue erosion due to price wars—a direct signal that undervaluing human expertise undermines long-term business viability.

Ultimately, edit-distance pricing and minimum rates are not simply about fairness—they are about sustainability. Without them, experienced linguists will exit the profession, leaving only underqualified workers willing to accept unsustainable pay. For clients, this translates into higher risk, weaker cultural resonance, and potential reputational damage. For LSPs, it means being sidelined in favor of direct MT platforms, which can deliver “good enough” but lack the added human value that clients actually expect when they pay for professional services.

2. Training and Specialization

At the same time, the path forward requires more than fairer pay—it requires investment in training and specialization. Machine translation still fails in areas that demand cultural sensitivity, sector-specific knowledge, or creativity. Legal and medical content, for example, cannot be trusted to raw AI without risking serious errors. LSPs that fund training in MTPE standards (such as ISO 18587), adaptive MT tools, and domain-specific expertise will not only reduce risk but also open up higher-value opportunities: specialized work in these sectors pays 20–50% more.

The industry also needs to recognize specialization as a survival strategy; rather than treating translators as disposable editors, LSPs should position them as knowledge partners who actively shape AI quality. Unfortunately, sidelining translators or relegating them to low-paid post-editing roles, LSPs doesn’t just alienate talent; it threatens the entire translator training sector. Becoming a skilled translator, especially in specialized fields like legal, medical, or technical translation, requires years of education, mentorship, and hands-on experience. If the industry continues to devalue human translators, the pipeline for developing new talent could dry up, undermining the teaching of translation and, ultimately, the quality of language services.

The Training Pipeline at Risk

Translation programs, whether at universities like Middlebury (which is already in the process of concluding graduate programs in Monterey, California...), Kent State, or global institutions, rely on a vibrant industry to justify their existence. These programs combine linguistics, cultural studies, and domain-specific training (e.g., legal terminology or medical protocols), often requiring 2-4 years of academic study followed by 3-5 years of practical experience to produce a competent specialist.

The erosion of translator training has far-reaching consequences, highlighting a worrying trend: a catastrophic decline in formal language learning, compounded by economic realities. The fact that new translators struggle to earn a living wage, discourage graduates from entering the field. If fewer students enroll and fewer graduates stay, the training sector faces a death spiral: shrinking programs, fewer qualified instructors, and a dwindling pool of specialized talent.

Specialized fields like legal, medical, or technical translation require nuanced expertise that AI cannot reliably replicate. For example, MT output always contains errors, many of which can be classified as critical, such as incorrect drug dosages or procedural terms, which could be life-threatening without human intervention. Similarly, legal translations demand precision to avoid misinterpretations that could void contracts or trigger lawsuits. Without a robust training pipeline, the supply of such specialists will dwindle, leaving LSPs unable to meet client needs for high-stakes projects.

Clients, too, will feel the impact. Overreliance on raw MT, without skilled translators to refine it, leads to homogenized language, cultural missteps, and quality complaints. Many clients, attracted by the speed and low cost of AI-only translation, have experimented with it for a variety of tasks, but a significant portion of these clients discover that for important, client-facing, or nuanced content, AI-only solutions fail to deliver on quality. The errors, lack of cultural context, and unnatural phrasing become a business liability. They eventually return to LSPs that offer either human-in-the-loop services or traditional human translation.

3. Advocate for Ethical AI Use, LSPs and Clients Education

LSPs must address ethical concerns—data privacy, cultural homogenization, and quality risks—to maintain trust, but above all their responsibility to their own workforce—the translators. The very model of "AI as a co-pilot" is ethically hollow if it leads to the exploitation of the human professional. LSPs have a moral and professional obligation to their freelance and in-house translators, whose expertise is the true foundation of quality. A reliance on AI as a cost-cutting tool, without a corresponding ethical framework for human collaboration, is unsustainable. 

So, on these four dimensions, LSPs should adopt the following practices:

• Offer a Human-First Option: For clients prioritizing accuracy and nuance, LSPs should explicitly position and promote a premium, human-centered service. Clear differentiation between low-cost, AI-only outputs and high-value, professionally curated translations allows for a sustainable model that preserves the craft of translation and affirms the centrality of human expertise.
• Guarantee Data Security: Sensitive projects should be processed through secure, on-premise AI systems rather than cloud-based solutions vulnerable to breaches. Encryption protocols and systematic human oversight are also necessary to ensure compliance with regulatory frameworks such as the GDPR.
• Educate Clients: LSPs must proactively communicate the limitations of raw machine translation—particularly in relation to idiomatic expressions, tone, and cultural nuance. Given that clients often overestimate AI’s accuracy by 20–30%, targeted education helps recalibrate expectations and underscores the enduring value of human intervention.
• Advance Fair AI Governance: Collaboration with professional associations and industry bodies is essential to establish ethical standards for AI integration in translation workflows. Such policies help prevent the exploitation of post-editors and support collective advocacy by translators for equitable working conditions.

4. Diversify Services to Stay Relevant

To remain competitive in an environment where clients increasingly consider bypassing Language Service Providers in favor of direct access to AI tools, it is imperative for LSPs to broaden their service portfolios beyond traditional translation. One promising avenue is the integration of value-added services such as transcreation, cultural consulting, and localization strategy. These areas rely heavily on creativity, cultural sensitivity, and contextual awareness—dimensions of language use that remain resistant to full automation. As a result, they command significantly higher rates, with transcreation, reflecting the premium placed on human creativity and cultural expertise.

At the same time, LSPs can strengthen their relevance by investing in the development of custom AI solutions. Rather than relying on generic LLM platforms, LSPs can collaborate with translators to train domain-specific, client-tailored models. This approach not only reduces the limitations inherent in off-the-shelf AI systems but also creates a value proposition rooted in proprietary expertise.

Finally, focusing on niche markets represents another essential strategy for safeguarding long-term sustainability. Fields such as gaming, law, and medicine demand a high degree of domain-specific accuracy, contextual adaptation, and risk management that AI alone cannot reliably provide. Translators specialized in such domains are less vulnerable to AI-driven disruption, as human oversight remains indispensable to ensuring both quality and compliance.

In short, by diversifying their service offerings—blending creative, strategic, and technical expertise—LSPs can not only mitigate the risk of disintermediation but also position themselves as indispensable partners in a translation ecosystem increasingly shaped by artificial intelligence. Equally important, such diversification ensures that translators themselves are not relegated to low-paid, mechanical post-editing tasks, but are instead recognized as highly skilled professionals whose expertise commands fair compensation. In this way, the sustainability of the industry is tied not only to client satisfaction but also to the necessity of enabling translators to make a decent living.

Reality Check: A Translator-Less World?

The notion of a translator-less world remains, for the foreseeable future, more hypothetical than real. While machine translation has achieved impressive progress in processing large volumes of text at speed, it continues to fall short in dimensions requiring intuition, creativity, and ethical discernment. In high-stakes or culturally sensitive contexts, human translators remain indispensable. Yet, as artificial intelligence advances and begins to approximate near-human fluency in narrow domains, the distinction between technological potential and practical applicability becomes increasingly difficult to draw. The present challenge, therefore, is not one of replacing translators, but of reconfiguring their role within hybrid workflows where human and machine operate in tandem. In this evolving landscape, the most viable model is one of symbiosis: AI delivers efficiency and scale, while human translators contribute depth, nuance, and cultural resonance.

The future of translation is not a zero-sum game between human and machine. It's an evolving symbiosis where each side plays to its unique strengths. AI provides efficiency and scale, while human translators contribute the depth, nuance, and cultural resonance that remain indispensable. Language Service Providers (LSPs) must understand and embrace this reality. The future of the translation industry depends on them practicing what the most viable model preaches every day: a commitment to the powerful and necessary collaboration between human and machine.

And here lies the essential truth: no matter how advanced artificial intelligence becomes, the final responsibility, the ultimate act of refinement and completion, rests with human translators. They are—and will remain—the true finishers of every translation, ensuring that language is not merely transferred, but fully realized in meaning, tone, and cultural integrity.

*

About finition, have you ever heard a well-known joke, often told to illustrate the value of expertise and experience over the time spent on a task:

A huge, brand-new cruise ship grinds to a complete halt in the middle of the ocean. The captain is desperate. The crew can't find the problem. They call in the world's best engineers, who work for days, examining the engine room with state-of-the-art instruments, but nothing works: the ship remains motionless. The owner, at his wit's end, is willing to try anything. He contacts an old retired mechanic, a local legend, who arrives carrying nothing but a beat-up old toolbox. He doesn't say a word. He walks around the engine room, listening intently to every sound and every vibration. He stops in front of a massive valve, observes it for a long time, then pulls a small hammer from his toolbox. He gives the valve two or three sharp taps in a very specific spot. Immediately, with a loud clatter, the engines start up again. The ship begins to move. The crew cheers, and the captain is relieved. The old mechanic, calm, puts his hammer away and hands the owner a bill for $15,000. The owner is shocked. "$15,000?!", he exclaims. "You only spent a quarter of an hour and gave it a few taps with a hammer! It can't possibly be worth $15,000!" The old man looks at him, unfazed. "Sir," he says, "I've itemized the bill for you." He hands over a small piece of paper on which he has written:
• Tapping with the hammer: $100
• Knowing where to tap: $14,900

Well, this is not a joke! LSPs should know that. And Clients too.