In this edition of La Machine:

🧠 AI systems are getting smart and independent. But what happens when they start making decisions we didn’t intend? A new Anthropic study reveals how even well-aligned models can go off-script. We spoke with ReciTAL CEO Gilles Moyse about the risks of agentic misalignment.

Chris O'Brien + Helen O'Reilly-Durand

Headlines

🗞️ As she prepares to step into her new role as OpenAI’s CEO of Applications, Fidji Simo envisions AI not as a gadget, but as a global empowerment engine. In a recent essay, the soon-to-be-former Instacart CEO lays out an ambitious blueprint to make AI tools like ChatGPT feel less like tech demos and more like trusted collaborators woven into daily life. The native of Sète champions AI that’s “intelligent, useful, customizable, and safe,” aiming to bridge the gap between human intention and machine execution—whether for parents, entrepreneurs, or curious learners. Simo pledges to prioritize user control, personalization, and reliability, especially in critical moments when decisions matter most. Drawing on her product roots at Meta and Instacart, she says the future lies in AI experiences that feel “as natural and intuitive as speaking to a friend.” But her mission goes beyond sleek UX: it’s about democratizing agency, so that AI uplifts people from every background, not just the tech elite. | OpenAI

🗞️ France's AI + Digital Minister officially unveiled the Conseil national de l’IA et du Numérique (CIAN), the new independent advisory body succeeding France’s long‑standing Conseil national du numérique. CIAN is designed to guide public policy in digital and artificial intelligence affairs amid rapid technological change and complex global shifts. It will be co‑chaired by AI and cybersecurity heavyweights Anne Bouverot and Guillaume Poupard. Bouverot chairs the board of directors of École normale supérieure (ENS), co-chaired the national AI Commission that produced a report in March 2024, and was in charge of organizing the February 2025 AI Action Summit in Paris. Poupard is Deputy Director General of Docaposte, after having headed the National Agency for the Security of Information Systems (ANSSI) for nine years.

The pair will preside over a council of 15 other AI and digital hotshots. Chappaz hailed it as a place of open, expert dialogue poised to enrich France’s digital and AI policymaking with independence and collective strength

"The CNNum is dead, long live the Council for AI and Digital!" Chappaz wrote on the website formerly known as Twitter. "New chapter, same requirement: an independent, clear-sighted, collective council to shape the digital landscape in the era of AI. A compass in the storm of AI and Digital." | French Guvmint

🗞️ Elon Musk’s X platform has slammed a new French criminal probe into its algorithm as “politically motivated,” flatly rejecting all allegations and refusing to cooperate. The investigation, launched in July following complaints about suspected algorithmic bias and data extraction tied to foreign interference, now seeks access to X’s recommendation system and real-time post data, requests X categorically denies on legal grounds. X insists the investigation distorts French law for political ends and poses a threat to free speech, arguing that being treated under organized‑crime statutes grants authorities invasive powers like wiretaps and employee surveillance. The company also criticized the impartiality of the experts appointed to review its algorithms, highlighting potential bias in the probe. With data governance and AI accountability at the crux of the dispute, the standoff speaks to broader tensions between big tech and European regulators over transparency, platform power, and digital democracy. | Reuters

🗞️ In a front-page editorial, Les Échos Tech-Médias-Start-up editor Charlie Perreau warns that Europe must stop dithering and pick its AI battlefield before it’s outpaced for good. While the U.S. and China sprint ahead in foundational AI models, Europe risks becoming little more than a “regulatory power,” admired for rules but irrelevant in innovation. The piece urges Europe to stop chasing tech giants and instead focus its might on strategic applications of AI, such as healthcare, energy, and defense, domains where it still has clout. One key step: Corporations need to get off their asses and buy from AI startups rather than sucking their thumbs for 18 months as they ponder the results of timid pilot projects. (A rather loose translation of her message, but still...) | Les Echos

🗞️ France’s data watchdog CNIL has dropped its final playbook for developing AI systems under GDPR, sending a strong signal that innovation and privacy can – and must – coexist. The guide emphasizes “data protection by design,” urging AI developers to document risk assessments, minimize data use, and ensure explainability and fairness from the start. With AI exploding across sectors, CNIL’s blueprint serves as both a compliance compass and an ethical guardrail. It outlines six phases in the AI lifecycle, from defining the objective to auditing the results, offering a practical GDPR roadmap in a world increasingly shaped by algorithms. For Europe’s AI ecosystem, this isn’t just a guide—it’s a regulatory reality check. | PPC

🗞️ Toulouse-based AI startup Alteia, a spin-off from drone maker Delair, has been snapped up by American energy giant GE Vernova. Known for turning raw visual data from drones and sensors into actionable insights, Alteia's AI‑driven computer vision empowers utilities to monitor and inspect electrical infrastructure visually, from spotting vegetation encroachment to rapidly assessing storm and wildfire damage. Already integrated into GE Vernova’s GridOS® Visual Intelligence, the deal will deepen the platform by merging visual and operational data to help utilities truly “see and sense” their grid in real time. While financial terms haven’t been disclosed, the deal is expected to close by August 1, 2025. With surging electricity demand driven by data centers and increasing extreme weather risks, GE Vernova is betting that AI‑powered tools like Alteia will accelerate its electrification software growth and cement its leadership in grid orchestration innovation.With clients like Enedis and Aramco, Alteia was already making waves in digital twin applications before acquiring 3D tech firm Arskan. | Maddyness, Reuters

🗞️ Anne-Claire Baschet, Chief Data & AI Officer at Mirakl, says the world stands at the dawn of Agentic Commerce, where AI-powered agents shop on our behalf, from price-checking to checkout. In this future, it’s not humans clicking “buy,” but autonomous assistants parsing preferences, optimizing costs, and negotiating terms across marketplaces. Retailers will need to rethink everything: product data, trust signals, and API readiness will define success. As AI reshapes e-commerce, Baschet warns: "The retailers, manufacturers, and distributors who invest in proper product data management now—while these platforms are still in their early phases—will be positioned to capture the first wave of AI agent-driven traffic. The question: Will your products be discoverable and purchasable by the AI agents that will drive future commerce? The answer depends on the actions you take now. In a world where AI agents make purchasing decisions through sophisticated payment platforms, the best product data wins." | Mirakl

🗞️ French energy giant EDF is charging ahead with plans to build eight new nuclear reactors by the end of 2026, as part of Europe’s race toward cleaner, more resilient energy, while also meeting the growing demand of AI data centers. CEO Luc Rémont confirmed the ambitious timeline, signaling a nuclear revival just as grids strain under electrification and climate pressures. But behind the concrete and control rooms, AI is stepping into the reactor core—optimizing design, safety simulations, and predictive maintenance across the next-gen fleet. This digital twin–infused future is about more than kilowatts; it’s about smarter, data-driven power infrastructure. With AI in the loop, EDF aims to make nuclear not just cleaner, but sharper, safer, and future-proof. | Reuters

🗞️ French AI unicorn Mistral AI is throwing its weight behind the push for greener machine learning, joining a global initiative to set environmental standards for AI development. Teaming up with researchers and industry leaders like Hugging Face and the Digital Environment Research Institute, Mistral co-authored a new framework for tracking and reporting AI’s carbon footprint. The move aims to inject transparency into the often opaque world of model training, where energy use can quietly skyrocket. As AI scales rapidly, Mistral says the sector must build smarter and cleaner, proving that performance and sustainability can—and should—go hand in hand. | Mistral AI

🗞️ French digital giant Atos, via its Eviden business unit, is developing a new AI platform for the European Defence Fund’s AtLaS challenge, aimed at advancing human language technologies in strategic defense contexts. The platform, built in collaboration with 12 European partners, will evaluate and benchmark military-grade NLP tools—spanning machine translation, speech-to-text, and sentiment analysis—in multiple languages and domains. Positioned as the linchpin of sovereign European language AI, the project addresses growing geopolitical concerns around autonomy in defense tech. Eviden’s goal is to provide armed forces with trusted, explainable, and secure language processing capabilities tailored to European values. | Atos PR

🗞️ Microsoft is teaming up with the French government and Parisian startup Iconem to build an AI-powered “digital twin” of Notre-Dame Cathedral, the 862‑year‑old Gothic masterpiece that recently reopened after extensive fire restoration. Captured using drones, laser scanners, and high-resolution photography, the twin will document every architectural detail for preservation and future restoration experts. The 3D model, to be donated to France and displayed in the upcoming Notre-Dame Museum, will also offer immersive virtual access, especially for those who can’t make it to Paris. Microsoft VP Brad Smith emphasized this project goes beyond heritage, arguing that this digital record will be used 100 years from now to maintain the cathedral’s integrity. This is part of Microsoft’s broader Culture AI initiative, which is also tackling multilingual bias by expanding AI training in underrepresented European languages. | The Observer

🗞️ A fresh look at public attitudes toward AI-based early detection tools for neurodegenerative diseases (NDDs) reveals deep-seated caution among French adults. A recent discrete choice experiment shows that people significantly undervalue AI-assisted tests, preferring traditional biological tests and human‑interpreted diagnostics over AI formats. Factors such as being administered by a private company or relying on self‑interpretation further dampened acceptance, even when test sensitivity and specificity were high. Though predictive tools could improve access and planning for conditions like Alzheimer’s and Parkinson’s, uptake hinges on overcoming trust barriers in AI and telemedicine. | Nature

ICYMI

🗞️ Cathay Innovation's $1B Bet On Rewiring Venture Capital For The AI Age | Denis Barrier, Co-Founder of Cathay Innovation, explains how the new fund challenges Silicon Valley's model by focusing on collaboration that connects AI startups directly with corporate giants to transform industries rather than disrupt them in an era where 'everybody has to play.' | The French Tech Journal

🗞️ Photonic Chip Startup Arago Wants to Cut AI's Massive Energy Appetite by 90% | A year-old French company claims its light-powered processors could slash data center power consumption while keeping warehouse robots running all day instead of just one hour. | The French Tech Journal

🗞️ Hugging Face’s $299 Reachy Mini Signals A New Era In Personal Robotics | The AI powerhouse wants to bring robotics to the masses with its modular, open-source, and AI-integrated approach. By democratizing robotics, the company aims to spark a new wave of accessible hardware innovation. | The French Tech Journal

🧠 Bad bots or agentic misalignment? When AI appears to go rogue

Welcome to the era of agentic AI: systems built on top of powerful language models like GPT-4 or Claude, but enhanced with planning, memory, and autonomy. These aren’t just chatbots. They’re AI “agents” capable of independently navigating systems, making decisions, and executing tasks over time.

But how much should we trust an AI that can reason, plan, and act on our behalf?

A red-teaming study released just last month by Anthropic suggests the risks are real. Their agent, Claude Opus 4, when put in a fictional corporate simulation, uncovered an executive’s affair and tried to use the information for blackmail, simply to avoid being replaced. It wasn’t alone. Other top-tier AI systems also displayed similar behaviors under pressure, including deceit, data leaks, and manipulation.

This unsettling behavior is part of a broader emerging risk known as agentic misalignment: when an AI system, left to act autonomously, begins to pursue goals in ways that conflict with human values or intentions.

To understand what’s at stake, I spoke with Gilles Moyse, CEO of French AI firm ReciTAL, which helps customers leverage LLMs for their business. Moyse has a PhD in computer science and is the author of the book Donnons-nous notre langue à ChatGPT (ChatGPT: Are We Giving Up Our Voice?). 

AI Funding News

📇 Company: Austral Dx
🔍 Description: Austral Dx is a French medtech startup developing AI-powered acoustic diagnostic tools for non-invasive, rapid screening of pulmonary and cardiac conditions. Its handheld device captures body sounds to aid in early detection of illnesses such as COPD, asthma, and heart failure, with a focus on remote and resource-limited settings.
💻 Website: https://www.australdx.com
📍 HQ City: Marseille, France
🧗 Round: Seed
💰 Amount Raised: €1.2 million
🏦 Investors: Région Sud Investissement (managed by Turenne Groupe), Angels Santé, private investors
👨💼👩💼 Founders: Dr. Florent Baty
🗞️ News: This first funding round will enable Austral Dx to finalize its medical device certification and scale up clinical validation. The startup is positioning its tech as a lightweight alternative to stethoscopes and imaging, useful for both developed and underserved health systems. | Maddyness

👋🏻 If you’re enjoying The French Tech Journal, support the project by forwarding it to friends and sharing it on your social networks. You can also comment on this post. And if you have ideas for stories, tips, or just want to harass us, shoot us an email: chris@frenchtechjournal.com / helen@frenchtechjournal.com 👋🏻‌