Most AI conversations in manufacturing stall on the same question: can we trust it? Mistral CEO Arthur Mensch has a blunt answer. AI is software, and software can be controlled. He is not making a philosophical point about machine consciousness. He is describing an engineering reality: a model is code and weights running on hardware you can audit, constrain, version, and switch off. If that sounds obvious, ask your vendor to show you the controls. Many cannot.

That reframing gives you a practical test for every AI system you buy or build, from vision inspection on the line to a document assistant in your quality department. Below, we break Mensch’s claim into the five control questions worth asking before you sign anything, and what good answers actually look like.

Your Board Thinks AI Is a Black Box. Your Engineers Know Better.

In the boardroom, AI gets discussed like weather. Unpredictable, powerful, something you prepare for rather than direct. Meanwhile, the people actually running models on your shop floor treat them like every other production system: versioned, logged, gated behind approval steps, rolled back when a release misbehaves. One group is having a philosophy debate. The other is doing change management.

Mensch’s comments to Le Monde are the sharpest counter to the black-box story yet, and they come from someone shipping models rather than commenting on them. Mistral AI builds systems customers can run on their own infrastructure, which forces the question of who holds the controls into the open.

So the useful question is not whether AI control is possible. It is which controls you actually need, and how you verify a vendor has them.

What Mensch Actually Said in Le Monde, and What He Didn’t

The technical argument: weights, logs, and versioning

Strip out the headlines and the argument is narrow. A trained model is a file of numerical parameters. Given the same input, the same weights and a fixed sampling configuration, you get the same output. That is not magic, it is arithmetic at scale, and it sits inside the same release discipline as any PLC firmware update or MES patch.

Mensch also pushed back hard on existential-risk framing, arguing it distracts from the engineering questions that actually determine whether a system behaves. He is right about that. Risk in production comes from bad data pipelines, silent model swaps by a vendor, and nobody logging what the system decided. Those are all fixable with configuration management, not moral philosophy.

The commercial subtext: why a European vendor benefits from this framing

Now separate the technical claim from the sales pitch. Mistral sells open-weight models you can run on your own hardware, which makes “AI is controllable software” a description of its product rather than a neutral observation about the field. European AI sovereignty and on-premise deployment are also its clearest differentiators against US hyperscalers whose models you rent through an API and never see.

That does not make the argument wrong. It makes it incomplete. A model served from someone else’s cloud, retrained on their schedule, with a terms-of-service page instead of a change log, is software in theory and a moving target in practice. The controllability Mensch describes is real, but it comes from the deployment architecture, not from the fact that AI happens to be code.

Where ‘Controllable Software’ Holds Up on a Production Line

A vision inspection model that classifies weld seams is a piece of software in your quality system. Treat it that way. It gets a version number, a validation record, a defined input specification, and a documented owner. When it changes, the change goes through the same gate as an MES configuration change, because the failure mode is identical: a silent shift in what gets passed and what gets scrapped.

The same logic applies to a deviation-triage assistant reading nonconformance reports. It suggests a category and a likely root cause. A human approves or overrides, and that decision is logged with the model version that produced the suggestion. Nothing about that workflow is novel. It is the approval routing you already run, with a new input source.

The control stack a quality manager can actually audit

Six controls decide whether your AI is governable. Pinned model versions, so nobody silently upgrades you mid-quarter. Deterministic inference settings, meaning fixed temperature and sampling config for anything touching a quality decision. Input validation that rejects out-of-spec images or malformed records instead of guessing. Human-in-the-loop gates on any disposition affecting product release. Audit trails capturing input, output, model version, timestamp, and the human who signed off. Documented rollback to the previous validated version.

Ask your vendor for evidence of each. If the model runs behind an API with no version pinning and no inference log you can export, you do not have AI control, you have a dependency. That gap is exactly what an IATF 16949 or ISO 13485 auditor will find, and it is why European AI sovereignty arguments about self-hosted, inspectable models land harder with quality teams than with boards.

Where the Software Analogy Breaks Down

Conventional software fails loudly. A null pointer throws an exception, a PLC faults, the line stops and somebody gets paged. Models fail quietly. They keep producing confident outputs while the accuracy underneath them erodes, and nothing in your alarm stack notices.

Drift is the failure mode nobody budgeted for

Change the supplier of a coating and the surface reflectance changes. Swap a camera after a crash and the white balance shifts two percent. Neither event triggers a code change, so neither passes through your change control. The model keeps running against inputs it was never trained on.

What follows is a silent quality escape. Defects get classified as pass, the scrap rate looks better than last quarter, and the problem surfaces as a customer complaint months later. Line-level traceability does not help here, because no line of code is wrong. The data moved and the software did not.

Non-determinism versus your validation protocol

Fixed weights and fixed sampling give reproducible outputs. That holds for a classifier with a locked configuration. It stops holding the moment you use a generative model where temperature is above zero, or where a vendor updates the hosted endpoint beneath you. Your validation record then describes a system that no longer exists.

Validating non-deterministic output is genuinely expensive. You cannot test one input and sign off. You need a held-out set, an acceptance threshold expressed as a rate rather than a pass or fail, and re-validation on a schedule. Real AI control requires drift monitoring, input distribution logging, and a defined rollback trigger. Most plants have built none of that, and buying the model is the cheap part.

Five Questions to Put to Every AI Vendor Before You Sign

Put these in the RFP, not the pilot review. Five questions, asked before money moves:

1. Can we pin and reproduce a specific model version? Same input, same version, same output, on demand, twelve months from now.
2. Where do inference and data physically run? Name the region, the provider, and what leaves your network.
3. What does the audit log capture? Input reference, model version, output, confidence, timestamp, and who acted on it.
4. How is drift detected, and who gets alerted? Thresholds, monitoring interval, and a named recipient inside your plant.
5. What is the rollback path? Time to revert to the previous version, and whether you can do it without a support ticket.

What a credible vendor answer looks like

A good answer is specific and slightly boring. “Model versions are immutable, tagged, and retained for five years. Inference runs in our Frankfurt region. Rollback is a config change you make yourself, effective in under a minute.” That is a vendor who has already been through a regulated customer’s audit.

A bad answer reaches for adjectives. “Enterprise-grade,” “continuously improving,” “our models learn from your data.” Continuous improvement without version pinning means your validated system changes silently under you. Ask what happens to your validation record when they push an update. If the answer is a shrug, walk.

The ROI of asking now is measured in revalidation cycles you never run. Discovering post-deployment that you cannot reproduce a classification decision means re-qualifying the process, re-running samples, and explaining the gap to an auditor. Weigh that against a thirty-minute procurement conversation. Vendors like Mistral compete on exactly this ground, so the controls exist. You just have to demand them in writing.

The Real Divide in 2026: Teams That Govern AI vs. Teams That Debate It

The controllability argument is effectively over. Whether you find Mensch persuasive or not, the practical question has moved from “can AI be governed” to “have you actually governed it.” That is a much less interesting question and a much more expensive one to get wrong.

Two kinds of manufacturers are visible now. The first has a named owner for every model in production, a change log, and a monitoring threshold someone checks. They deploy in weeks because the approval path already exists, and they survive customer audits because the evidence was collected as they went, not reconstructed afterwards. The second is still running steering committees on AI ethics while three departments quietly pay contractors to key in data by hand.

The cost of the second position is not theoretical. Every quarter spent waiting for certainty is a quarter of manual triage, manual inspection review, manual document handling, paid for in headcount you could have pointed at process improvement. Meanwhile your competitor’s quality team has already absorbed the tooling into their standard work and stopped talking about it.

European sovereignty rules will harden this split further. Buyers who can name where inference runs and produce a version history will clear procurement. Buyers who cannot will find their AI projects blocked by their own customers’ compliance teams, which is a slower and more embarrassing failure than a bad pilot.

One control this quarter: build a single register of every AI system touching your operation, with the owner, the version in production, and the date it was last validated. Most plants discover they have more models running than they thought, and no one accountable for half of them.

Source: lemonde.fr