Devant l’urgence, comprendre et faire comprendre les modèles - Franck Varenne

Summary

Why Models Matter Now

Franck Varenne links two urgent challenges: ecological and social-ecological change, and the rapid expansion of AI. Climate change, biodiversity loss, resource pressures, and social transitions involve many interacting processes and scales. Models are therefore important tools for anticipating developments and informing action.

Models have traditionally been valued for their modesty: they address a particular question or aspect of a system rather than claiming to explain everything. They can also be adapted and used alongside other models. But advances in computing have made it possible to build increasingly detailed, integrated models—not all of which are simple or readily understood.

A New Demand for Understanding

Varenne identifies two reasons why understanding models has become a pressing issue:

• Social and democratic: People may be asked to change their behavior, accept sacrifices, or support public policies. In a democracy, they need to understand the reasons for recommendations well enough to assess and act on them. This is especially important when humanity as a whole is implicated in the causes of a problem.
• Scientific and epistemological: Researchers and engineers increasingly build or use models whose internal operations they cannot fully interpret. Complex models may first be used to describe, predict, or calibrate, with understanding pursued later—or not achieved in a straightforward way.

These demands can conflict: the public wants intelligible guidance, while the problems may call for models too complex to be immediately intelligible to either the public or their designers.

What a Model Is—and What It Does

Varenne argues that models should not be defined only as simplified representations. A model can instead be understood as a substitute object that makes it easier to ask questions about a target system. It may simplify the questioning process without simplifying the system’s representation. For example, a laboratory organism may be almost as biologically complex as the organism it stands in for, yet make experimentation more feasible.

To describe and compare models more carefully, Varenne proposes a four-part “characteristic vector”:

1. Epistemic function: What kind of knowledge is the model intended to provide—such as description, prediction, explanation, understanding, experimentation, or decision support? Varenne says he has identified 21 possible functions, while allowing that the classification may change.
2. Nature: What is the model made of? It might be a physical object, a living organism, a mental construction, or a symbolic or computational model.
3. Principle: What makes the model useful for its purpose? Possibilities include analogy, representation or denotation, shared characteristics with the target, or compression of information.
4. Use: How is the model deployed in practice or argument? Its use may be informative, prescriptive, heuristic, or intended to orient research or decisions.

This framework helps ensure that models are assessed against what they actually claim to do, rather than against functions they were never designed to perform.

Distinguishing Kinds of Knowledge

A central lesson is not to treat several different aims as interchangeable:

• Description characterizes a system or its state.
• Prediction concerns what may happen and generally requires at least some representation of the system’s dynamics.
• Explanation identifies interactions or mechanisms that account for a phenomenon.
• Understanding, in Varenne’s account, is a person’s ability to grasp a phenomenon as a whole through a manageable mental representation. It is subjective and culturally shaped, and is not identical to explanation.

A model may perform one of these functions without performing the others. For instance, a model may predict effectively without providing a humanly intuitive explanation. Conversely, a model that represents mechanisms does not automatically provide a useful prediction.

Critiquing Complex Models and AI

Varenne uses numerical AI to illustrate why careful distinctions matter. Such systems can compress data, classify patterns, make predictions, support decisions, or generate outputs. Their success does not by itself show that they have identified meaningful concepts or mechanisms in the world. A system may exploit a pattern that works in its training data but fails in a different setting.

Nor, he argues, are models entirely free of assumptions. Even a model presented as “agnostic” or purely predictive depends on mathematical choices about how data are represented and measured. These choices carry at least minimal assumptions about the structure of the data—and therefore about the world being modeled. The appropriate response is not to reject models wholesale, but to make their assumptions and limitations open to scrutiny.

Modeling Complex Systems and Involving Stakeholders

The lecture discusses different strategies for tackling complex systems, including system-dynamics models and multi-agent simulations. Each has strengths and limits: some approaches seek stable patterns or attractors, while others simulate interactions among individual agents and may require empirical analysis to interpret their results.

For social and ecological systems, Varenne emphasizes participatory modeling. Stakeholders can help define the problem, make their practices and values visible, explore scenarios, and test possible responses. Participation is not merely a matter of courtesy: recommendations are less likely to work if the people expected to act on them have not understood or helped shape them.

Varenne nevertheless raises a difficulty: deliberation and shared understanding take time, while ecological crises may require action quickly. Stakeholder involvement is valuable, but it does not remove conflicts between values or guarantee that society will find solutions soon enough.

Proposed Standards: Feasible, Criticizable, Acceptable

Varenne’s guiding claim is that models should be feasible, open to criticism, and acceptable.

• Feasible: They should be technically workable, usable in their social context, and—as far as possible—constructible or explorable by people beyond specialist communities.
• Criticizable: Their functions, assumptions, methods, values, code, and data should be open to examination. Criticism should be appropriate to the model’s purpose and structure.
• Acceptable: People should be able to assess the evidence and recommendations, including what is known and unknown. Acceptability should not mean resignation or manipulation designed to preserve existing power arrangements.

Practical Guidance from the Lecture

When developing, evaluating, or presenting a model:

• State clearly what the model is meant to do: describe, predict, explain, support understanding, or guide a decision.
• Distinguish its function, nature, principle, and use rather than treating “model” as a single, self-explanatory category.
• Do not assume that a model must be simple, or that a complicated model is automatically explanatory.
• Match evaluation and validation criteria to the model’s stated function.
• Make assumptions, values, uncertainties, and potential biases as explicit as possible.
• Provide access to the code and data where possible, and allow users to explore the model when appropriate.
• Involve stakeholders in co-construction when their knowledge or actions are necessary for the model’s practical success.
• Explain findings without pretending that uncertainty has disappeared or using manipulative persuasion to secure acceptance.

Speakers and Sources Featured

Speakers

• Franck Varenne — the lecturer.
• Michel — thanked for welcoming and introducing the speaker; no further contribution is transcribed.

People and Intellectual Sources Cited

• Christophe Bonneuil
• Jean-Baptiste Fressoz — the surname is rendered unclearly in the subtitles.
• Jean-Marie Legay
• Jean-Louis Le Moigne
• Marvin Minsky
• Herbert Simon
• Jeff Forrester
• René Thom
• Max Weber
• Spinoza
• Freud
• Samuel Alizon
• Juliette Rouchier
• Pascal Picq
• Christian Anderson — likely a subtitle rendering of Chris Anderson.
• “Claudine Arrière” and “Robert Durne” — names as rendered in the subtitles; the transcription is uncertain.
• Bohler — referred to by surname only; the first name is not clear in the subtitles.

Organizations, Projects, and Other Sources Mentioned

• CIRAD
• EPFL
• Blue Brain Project
• Human Brain Project
• Physiome Project
• Digital Patient project
• IPSL
• Nature et Sociétés / “NSS” — referenced in connection with an article on comparing models.
• Science and Technology Studies (STS)
• The Ministry of Ecological Transition
• The book Theory and Models in the Human Sciences
• The article “Comparing Models: The Characteristic Vector”