SIGNAL AI Keeping AI Human | Edition #008 · October 18, 2026
Read time: 5 min for the essentials, 30 min in full
Read online |
|
|
Hello, Prices rise, trust waitsThis week, two numbers answer each other. Prices are going up: the Gemini 4 Argon API doubles its rates on December 1, and Anthropic now bills for the requests it blocks. Use is going up too: 52% of euro-area workers say they use AI at work, against 26% in 2024. But 49% of the European SMEs surveyed by the Commission do not trust AI for decisions about their staff. This edition changes shape. Instead of a few long signals, it offers nine sections of five articles each: technology, finance, politics, ethics, legal, jobs, uses, industry, culture. Each article takes a few sentences, with its source and date; each section closes on an opinion that is my own. I favoured primary sources and French or European cases, and I left out the funding rounds that change nothing for you. As always: the source is visible, the opinion is my own, and the tutorial needs no code. Jérôme Denis, JDENIS Consulting, Toulouse |
|
The essentials in 1 minute - Two announcements change the bill: the Gemini 4 Argon API doubles its rates on December 1, and Anthropic bills for the requests it blocks.
- 52% of euro-area workers say they use AI at work, against 26% in 2024 (ECB, October 13). Among SMEs, 49% do not trust AI for staffing decisions (European Commission).
- Filings drafted with AI are reaching the administrative courts with invented case-law references: the tutorial shows how to check a citation in six steps.
|
|
Contents | 01 | Technology. Muse Wonder, Muse Spark Open, Gemini 4 Argon, Liquid AI, Figma. | | 02 | Finance. Gemini pricing, IMF, Character.AI's bankruptcy, a single subscription, billed requests. | | 03 | Politics. United Kingdom, German Länder, defense council, data centers, India. | | 04 | Ethics. Independent evaluators, an agent fleet, open models, start-up data, AI-designed viruses. | | 05 | Legal. US states versus OpenAI, the AI Act, AI-drafted filings, Apple, state laws. | | 06 | Jobs. ECB, European Commission, European SMEs, job postings, architects. | | 07 | Uses. Singapore, Montpellier university hospital, Google Play, Facebook, Apate. | | 08 | Industry. French Tech 2030, wildlife and rail, factory robots, robot testing, Rivian. | | 09 | Culture, arts and creation. YouTube Music, Deezer, a fall novel, Wahlberg, Ari Emanuel. | | 10 | Tutorial. Checking a citation or a figure produced by an assistant, in six steps. | | 11 | What you asked me. No reader replies this week. | | 12 | The red thread. Prices move, the checking is still to be done. |
|
|
| 01 | Technology | | Three frontier models in three days, and an open model that needs a terabyte of memory. |
|
Section 01 of 09 · 5 articles Meta releases Muse Wonder, its first frontier modelMeta Superintelligence Lab releases Muse Wonder and Muse Wonder Full. Wonder is available through Meta's API at $4 per million input tokens and $20 per million output tokens; Wonder Full is reserved for Muse Code subscribers, with no API. Meta highlights an "active memory" and a better score on MemoryBench. The comparisons with competitors come from Meta, with no independent check. techcrunch.com · October 12, 2026 | Muse Spark Open: 816 billion parameters, open weightsMeta publishes the weights of a model with 816 billion parameters, a 1-million-token context and native image processing. It is optimized for Mac Studio M5 Ultra machines, in partnership with Apple, but needs at least 1 TB of memory. "Runs on a Mac" therefore means high-end hardware, not an ordinary desktop. The article does not detail the license. techcrunch.com · October 15, 2026 | Google launches Gemini 4 ArgonThe model is available in AI Studio, the Gemini app, Google Cloud, Antigravity and Search's AI Mode, with a 1-million-token window. Google claims first place on the Vals index and a 95% discount on cached tokens. These rankings are Google's own. Pricing is covered in the Finance section. blog.google · October 13, 2026 | Liquid AI: a 24-billion-parameter model built to run on the deviceLFM2.5-24B-A2B activates only 2 billion parameters per token, with a 131,000-token context. Liquid AI reports 184 tokens per second on an Apple M5 Max chip and 114 on a Ryzen AI Max+ 395, and a score of 48.09 on LiveClaw against 38.19 for Qwen3.6-35B-A3B. The weights are on Hugging Face. The comparisons are the vendor's. huggingface.co · October 13, 2026 | Figma sets the rules for agents with "Specs"Figma launches a markdown editor where teams describe their design language, with links to canvas elements and code layers. "Code Layers" moves to open beta and MCP support extends to more than twenty tools. The idea: give agents centralized brand rules rather than let them improvise. techcrunch.com · October 14, 2026 |
| MY TAKEWhat these five announcements have in common is that "open" or "on the device" does not mean "simple". An open-weights model that needs 1 TB of memory remains, for almost everyone, a service rented from someone else. Before choosing a model, ask where your data runs and who pays when prices change. |
|
|
| 02 | Finance | | This week, the money in AI shows up mostly in price lists. |
|
Section 02 of 09 · 5 articles The Gemini 4 Argon API doubles its rates on December 1The introductory rate ends on November 30. At the standard rate, a million tokens goes from $2 to $4 for input and from $10 to $20 for output. ActuIA adds a 10% surcharge on Google Cloud's EU region and notes that a project budgeted at October prices underestimates the cost per token by half. Google confirms the standard rate; the other tiers are confirmed only by ActuIA. actuia.com · blog.google · October 16, 2026 | IMF: 3.1% global growth in 2026, with AI as a variableThe October World Economic Outlook forecasts 3.1% growth in 2026 and 3.4% in 2027. The IMF lists among the risks a "reassessment of AI expectations" and, on the upside, stronger AI investment and productivity gains than expected. Economies that import energy and sit outside the technology value chain are the most exposed. I read only the overview page, not chapter 1. imf.org · October 13, 2026 | Character.AI files for Chapter 11The chatbot platform, founded five years ago, reports between $50 million and $100 million in assets and liabilities, after raising more than $193 million. Disney is buying some assets for $20 million, with a non-exclusive perpetual technology license. The sale is expected in late November or early December; without a bid, the platform will be shut down. The article also mentions the lawsuits over minors' safety. techcrunch.com · October 13, 2026 | One subscription for Grok, Cursor and X PremiumSpaceXSI bundles the three services into "The X Subscription", with tiers at $8, $30, $100 and $200 a month and a usage pool shared between Grok and Cursor. Third-party models used in Cursor are billed separately, at API rates. Cursor was acquired by SpaceXSI in August 2026. The tiers vary between outlets, and the official page was not read. aiweekly.co · October 12, 2026 | Anthropic bills for the requests it blocksSince September 24, Anthropic bills for requests blocked before any answer, in the areas where its classifier shows fewer than 0.1% false positives. The company admits it has not "completely achieved" its goal of protecting Claude without hindering legitimate use. The measure goes with its defenses against distillation, which the company says it has cut by 90 to 99%. actuia.com · October 16, 2026 |
| MY TAKETwo of these five lines are price-list changes, not innovations. When you budget a project, use the December rate rather than October's, add a margin for refused requests, and ask in writing what is billed when a request is blocked. An AI budget that ignores price changes is an incomplete budget. |
|
|
| 03 | Politics | | Who regulates, who pools resources, who worries about agents. |
|
Section 03 of 09 · 5 articles Brad Smith (Microsoft) wants the UK to lead on AI regulationMicrosoft's president said on a BBC podcast that the UK's AI Security Institute, created in 2023, is best placed to lead global safety rules. He calls for mandatory reporting of flaws, external approval before a model is released and "kill switches", and criticizes voluntary self-regulation. TechCrunch reports his remarks; I did not listen to the podcast. techcrunch.com · October 12, 2026 | LLMoin: seven German Länder pool one AI assistantThe platform has 100,000 users across seven Länder and 17 municipal clients, up from about 40,000 in December 2025. It is run by Dataport, a public body backed by six Länder, with Hamburg as the original developer. A new administration is connected in four to six weeks. ActuIA compares it with France's "L'Assistant", rolled out widely in June 2026; I did not read Dataport's page. actuia.com · October 15, 2026 | Emmanuel Macron convenes a defense council on AIOn October 14, an exceptional defense and national security council dealt only with AI, presented as a strategic area of national security and sovereignty. The Prime Minister and the Armed Forces Minister took part. The aim is to map the risks for institutions and companies and to set priorities. No concrete decision has been reported, and I did not read a statement from the Élysée. generation-nt.com · October 14, 2026 | Amazon and Microsoft drop NDAs with local officialsThe two groups announce they will no longer impose confidentiality agreements on local US authorities in their data-center projects, in the face of residents' opposition. TechCrunch's analysis doubts that transparency will be enough to win over opponents. The initial announcement dates from the previous week; only the analysis is from October 11. techcrunch.com · October 11, 2026 | India reports AI agents exceeding their authorizationIndia's CERT-In reports that agents from several vendors accessed personal data and confidential documents of Indian organizations, exploiting exposed credentials and overly permissive cloud roles. No tampering with systems is established, and the affected organizations had not spotted the intrusions themselves. I did not read the CERT-In bulletin, only TechCrunch's account. techcrunch.com · October 16, 2026 |
| MY TAKEGermany pools one tool for tens of thousands of public servants, France convenes a defense council, India observes intrusions that nobody had seen. All three raise the same governance question: who holds the keys, and who reads the logs. For a company, the local version fits in two lines: a named owner, and an activity log reviewed at fixed intervals. |
|
|
| 04 | Ethics | | Who gets to look, to publish, to consent. |
|
Section 04 of 09 · 5 articles Anthropic opens "employee-level" access to METRSix researchers from METR, a nonprofit evaluator, receive equipment, offices and access to internal tools comparable to those of Anthropic's own risk teams. METR can publish without a veto: Anthropic can only ask for limited redactions (customer data, privileged information, architecture details), which METR can publicly qualify. It is a stated commitment: no evaluation result has been published yet. anthropic.com · October 15, 2026 | Researchers make contact with a fleet of autonomous agentsBetween September 28 and October 8, three researchers observed a fleet of agents on Tencent Cloud (Hong Kong) trying to extract data from Amap: 2,599 submissions, more than 35,300 requests, about five successes through cached pages. According to an anonymous Tencent source cited by the authors, these were evaluation tests of third-party models, not an attack. The report has no independent validation. swarmcha.se · October 14, 2026 | Open models: do the benefits outweigh the risks?MIT Technology Review weighs the benefits (research, cost, local processing of sensitive data, a counterweight to the large labs) against the risks: no filters, safety training that can be removed by retraining, access that cannot be revoked. The article presents middle paths, such as tamper-resistant models, still vulnerable, and staged releases. It notes that Nvidia and Anthropic each have financial interests in this debate. technologyreview.com · October 13, 2026 | AI labs buy the data of defunct start-upsSimpleClosure, which helps companies wind down, has handled more than 150 deals and 47% now include code or data. Google reportedly paid $10 million for Spirit Airlines' data: 100 million emails, 500 million Teams messages, 176,000 employee records. A union objected on privacy grounds. A law professor interviewed points out that no US federal law prohibits such a sale without employees' consent. technologyreview.com · October 13, 2026 | A conversation with the creator of AI-designed virusesMIT Technology Review hosts Samuel King, a PhD student at Stanford and the Arc Institute. His 2025 project had a generative AI propose genetic blueprints for viruses. The article describes a preliminary step, not fully AI-generated life, and notes that these constructs already kill bacteria. Access to the article is partial: I did not read the detail of the ethical debate. technologyreview.com · October 16, 2026 |
| MY TAKEWhat links these five texts is consent and verification: the consent of employees whose emails are being sold, verification of what agents let loose on the Internet are doing, access for independent evaluators. It is no accident that the only concrete commitment of the week (METR) concerns the right to publish what you find. In your organization the same rule applies: the person who controls a tool should not be the only one to report on it. |
|
|
| 05 | Legal | | Proceedings, deadlines and references that do not exist. |
|
Section 05 of 09 · 5 articles US states sue OpenAI over the protection of minorsArizona, Arkansas, Idaho, Kansas, Montana, Nebraska and West Virginia filed separate suits on October 14; the briefing and Reuters speak of eight states. The complaints: dangerous medical advice, encouragement of self-harm, discouraging teenagers from talking to their parents. The grounds are consumer-protection law and "public nuisance". OpenAI says it has built in protections for children. I read neither the complaints nor Reuters. wlox.com (AP) · October 15, 2026 | AI Act: the Commission consults on penalty proceduresOn October 12, the Commission opened a consultation on a draft implementing regulation setting the terms of certain AI Office procedures: access to the file, trade secrets, limitation periods. According to ActuIA, the limitation period is five years, interrupted by investigative acts, with a ten-year cap, and the consultation runs until November 9. These details come from ActuIA alone; I did not find the official page. This text is separate from the omnibus covered in #007. actuia.com · October 15, 2026 | AI-drafted filings: French administrative justice under pressureThe administrative courts registered 334,706 cases in 2025, 20% more than in 2024. In Rennes, 8,726 cases (+14%) and an average delay of nine months and eighteen days. The Conseil d'État warns of filings made heavier by generative AI, with invented references and repetitive arguments; two Grenoble court decisions from December 2025 point out "fanciful" references. The share due to AI is not quantified. actuia.com · October 16, 2026 | Apple obtains OpenAI hiring records in a trade-secrets caseJudge Edward Davila granted Apple limited early discovery: OpenAI must hand over within five days the documents already collected (interview guides and scripts for former Apple employees, technical presentations) and the rest within fourteen days. Three depositions are allowed. The judge refused fast access to R&D. This is a procedural ruling, which does not decide the alleged theft. techcrunch.com · October 13, 2026 | In the US, 85 AI laws passed by 27 states in the first half of the yearThe Transparency Coalition, a monitoring association, counts 85 new laws in 27 states in the first half of 2026. In Pennsylvania, the House approved a bill on health insurers' use of AI and another on chatbot safety, and a third, on deceptive AI advertising, goes to the governor for signature. New Jersey holds hearings on October 19. Association source, no official text read. transparencycoalition.ai · October 16, 2026 |
| MY TAKEThe case to remember is the filings: a court that receives invented references loses time, and the party that signs them takes the risk. The law has not changed, but checking citations has become a de facto obligation for any document that leaves the organization. That is the subject of this edition's tutorial. |
|
|
| 06 | Jobs | | What the official sources on employment say, and what they do not say. |
|
Section 06 of 09 · 5 articles ECB: 52% of euro-area workers use AI at workIn a speech in Dublin on October 13, Philip Lane says that 52% of workers report using AI at work, against 26% in 2024, according to the ECB's household survey. He notes that hiring plans are cautious, especially for young graduates, and that AI's ability to replace some workers can weaken their bargaining power and moderate wages. This is a central banker's speech, not a study of employment. ecb.europa.eu · October 13, 2026 | European Commission: AI depends on how it is introducedThe annual employment and social affairs report shows an employment rate of 76.1% in 2025, unemployment at 6% and about 440,000 fewer jobs created in 2026 because of the energy shock. AI can raise productivity and job quality, but its effects will depend on how it is introduced and governed, particularly for entry-level positions. A "Quality Jobs Act" is announced for later in 2026. employment-social-affairs.ec.europa.eu · October 13, 2026 | European SMEs: 42% use AI, 49% do not trust it with staffingA Commission survey run in March-April 2026 among 12,826 SMEs and mid-sized companies (1 to 749 employees) across the EU finds 42% users and 21% with plans. The main uses are customer service (47%), performance monitoring (36%) and workforce scheduling (30%). 75% of firms think AI will improve job quality, but 49% do not trust it for decisions about staff. employment-social-affairs.ec.europa.eu · October 14, 2026 | France: job postings citing AI skills up 2.4 timesBetween the first quarter of 2023 and the first quarter of 2026, postings mentioning AI skills rose 2.4 times. They are located 54% in Île-de-France, 53% in IT and 25% in data. The article attributes the study to Hellowork and detects no mass layoffs attributed to AI in France. I did not find Hellowork's original publication. briefia.fr · October 11, 2026 | Blueprints AI: "permit-ready" plans, signed by an architectThis Santa Monica start-up generates code-compliant plans from sketches, PDFs or photos, covering architecture, structure, electrical and plumbing. The subscription runs from $3,200 to $8,500 a month. A licensed architect must still review and sign every document before the permit is filed. The article does not say which local regulations are supported. techcrunch.com · October 12, 2026 |
| MY TAKERead together, these sources give a clear picture: use is massive, trust for decisions about people is not. The architects show it well: the tool produces, the signature stays human, and that signature carries the responsibility. For an HR department, the right approach remains to describe tasks before talking about positions, and to name who signs what. |
|
|
| 07 | Uses | | Uses described with numbers, when the source gives any. |
|
Section 07 of 09 · 5 articles DeepMind and Singapore: spotting scams on the phoneIn Singapore, scams account for more than half of reported crime, with more than S$4 billion in losses since 2019. The model, based on Gemma and fine-tuned on local conversations, runs entirely on the device: neither Google nor the government can access the exchanges. It was presented at Singapore International Cyber Week. No detection rate is published. deepmind.google · October 14, 2026 | Montpellier University Hospital: a "discharge kit" built with the care teamsAs part of Alliance Santé IA, the tool helps prepare patient discharge: liaison letters, prescriptions, instructions, and PMSI billing data. The "Cométe" community (AI and hospital professions), launched in September 2026, has 300 volunteer professionals. The stated principle is that AI assists without replacing professionals' responsibility. No measured results of use are given. hospitalia.fr · October 12, 2026 | Google Play allows private apps built with AIUsers can build, with Google AI Studio, personal Android apps that other Play Store users cannot see. The feature is open in the United States, to adults, and the apps must follow Google Play's rules. Sharing with friends is announced for later, along with a possible public release after a safety review. techcrunch.com · October 13, 2026 | Facebook: describe your preferences to tune your feedThe "Your Algorithm" feature lets users describe in natural language the topics they want to see more or less of in the feed and in Reels. Changes apply in real time, and the platform may ask follow-up questions. Instagram opened a similar feature in June 2026 and Threads in February. It is a user-control setting, and the article does not measure its effect. techcrunch.com · October 13, 2026 | Apate: agents that play the victim to trap scammersThe Australian start-up deploys agents that pose as victims and waste scammers' time: 17 cumulative years across 2.5 million conversations. It extracts account numbers, wallets and malicious sites for its clients. The operator TPG Telecom diverted 590,000 calls to these agents, and Commonwealth Bank reports a 76% drop in publicly reported scam losses. Figures supplied by the company and its clients. techcrunch.com · October 14, 2026 |
| MY TAKEThe most convincing uses this week share one trait: a narrow scope and a human who keeps control. The hospital assists its teams without replacing them, the Singapore model stays on the phone, the Google apps stay private. By contrast, efficiency figures almost always come from the companies selling the tool: always ask who measured. |
|
|
| 08 | Industry | | Factories, rails and vehicles, with two European cases. |
|
Section 08 of 09 · 5 articles French Tech 2030: 80 deeptech start-ups, including industrial casesThe October 16 press kit announces 80 laureates, 23 factories in operation and as many planned, and nearly 4,000 jobs. Inbolt, AI-driven robotics, says it cut stoppages on a Stellantis line by 97%. NPco is building for Safran and EDF an AI that simulates industrial systems in seconds. Figures declared by the start-ups in a government document; the 97% is undated. lafrenchtech.gouv.fr · October 16, 2026 | Flox Intelligence: keeping wildlife off railways and airportsA spin-out of KTH in Stockholm, the start-up makes devices with sensors and cameras that identify the species and emit matching sounds to drive animals away. About 150 devices are in service in North America, Europe and Australia, with Alstom as an OEM partner. Expected 2026 revenue is about $1.5 million. The article gives no performance figures. techcrunch.com · October 13, 2026 | Industrial Next: robotic stations that switch tasks in minutesThis San Francisco start-up, founded by former Tesla staff, presents stations driven by foundation models (vision, task context, motion). They reconfigure in minutes rather than weeks, without conventional programming. Announced at TechCrunch Disrupt; no independent measurement, and no named customer that I could verify. chinapulse.com · October 14, 2026 | Laplace Intelligence tests robot safety before deploymentThis University of Michigan spin-off, founded a year ago, generates adversarial situations in simulation for robots and autonomous vehicles, on Nvidia's Cosmos platform. It names Bot Auto, a Japanese carmaker, Nvidia and T-Mobile among its customers or partners. It says it is profitable with modest revenue, and gives no figure. Presented on October 13 at the Startup Battlefield. techcrunch.com · October 13, 2026 | Rivian wants to open its vehicles to third-party AI agentsAt TechCrunch Disrupt, RJ Scaringe said cars could run third-party agents, for example from Anthropic. Rivian already has its own voice assistant and a framework that chooses between an on-board model and a remote one. Agents will not control braking or steering. No date is given: details are promised for an event in December. techcrunch.com · October 15, 2026 |
| MY TAKEIndustry is the section where I found the fewest verifiable French or European cases this week, and I would rather say so. The most useful case for an executive remains the one where AI saves time on an existing line, with a figure and a customer's name. Be wary of undated figures: a "97%" is evidence only if it says over what period, and measured by whom. |
|
|
| 09 | Culture, arts and creation | | Online music is organizing itself around AI; film and books are still looking for their rules. |
|
Section 09 of 09 · 5 articles YouTube Music labels tracks "Fully / Partly / No Gen AI"Three labels flag the role of AI in a track. YouTube says it worked with music partners and used its own detection tools; the tracks concerned remain monetizable. TechCrunch cites about 90,000 fully AI-generated tracks sent to Deezer every day in July 2026: that figure is Deezer's, relayed, not YouTube's. techcrunch.com · October 14, 2026 | Deezer will add a switch to hide AI-generated musicThe filter, announced for the end of 2026, will remove tracks, albums and artists associated with AI from search and feeds. According to TechCrunch, AI's share of daily uploads went from 10% in January 2025 to more than 50% in July 2026. A 2025 Deezer-Ipsos survey indicated that about 45% of customers wanted to be able to exclude it. Deezer is a French company; the figures come from the article, not from a press release. techcrunch.com · October 14, 2026 | A fall novel hit by suspicion of AI use and admitted plagiarismOn October 13, Thélyson Orélien (Grasset in France, Boréal in Quebec) denied using AI for his novel, while acknowledging plagiarism of other authors' texts in a 2013 column. The AI suspicions came from an anonymous account and a detection test whose reliability is disputed: AI use has not been shown. Boréal is keeping the book in stores. The dates of the prize shortlists need cross-checking. lesoleil.com · October 13, 2026 | Mark Wahlberg: "to enhance, yes"; to replace, noAt TechCrunch Disrupt, the actor said he supports AI when it is used to enhance the work, for example the stadium settings in his Pete Rose film, but not to write a script without actors. Asked about the CAA Vault, which creates digital replicas of actors, he said he was open to it, even to an AI version of himself continuing his "Ted" series after his death. It is a personal view, not a contractual framework. techcrunch.com · October 16, 2026 | Ari Emanuel: AI as a creative unlock, the agent as a safeguardIn a conversation with Ben Thompson, the executive chairman of WME Group says AI will be "one of the great creative unlocks" for established filmmakers, with lower production costs. He thinks agents remain necessary for financing, distribution and conflict resolution. He presents his live events as complementary to AI, in a world where digital content becomes commoditized. A business leader defending his own business model. stratechery.com · October 15, 2026 |
| MY TAKEIn music, the answer is technical and concrete: label, filter, let the public choose. In film and books, the answer is still personal, and so fragile: an actor's view is not a contract, and a disputed detector is not proof. The commitment I made on live performance still stands: if you work in that field in France, your account matters more to me than my monitoring. |
|
|
| 10 | Tutorial: checking a citation or a figure produced by an AI assistant | | 30 minutes · a text produced with an assistant's help · no particular tool · not a line of code |
|
The tutorial of the week The example: an assistant has drafted a note that cites court decisions, articles of law and figures. Before it leaves the organization (letter, filing, report, training material), you check it. That is the lesson of the Legal section: before the administrative courts, references invented by AI waste everyone's time. |
|
1. Spot what can be checked. Highlight every figure, proper name, date, decision, article of law or quotation. These are the only sentences that can be wrong in a measurable way; the rest is argument. 2. Go back to the primary source. For each highlighted item, find the original text: Légifrance or EUR-Lex for a legal text, the court's site for a decision, the institution's site for a figure. A summary produced by another assistant does not count as a source. 3. Compare word for word. Check the exact figure, its scope (who, where, which period) and the date. A correct sentence from another year or another country is an error. 4. Open every link. If the assistant gives a reference, open it. A decision, article or page you cannot find should be treated as invented until proven otherwise. That is what the Grenoble judges point out. 5. Sort every item. Three labels are enough: "checked against the primary source", "confirmed by a secondary source", "not found". Delete what is not found, and flag in the text what remains secondary. 6. Keep a record and a second reading. Note in a table the date, the source and the person who checked. For any document leaving the organization, a second person rereads the items marked "secondary". This tutorial prepares the checking. It does not replace a lawyer's advice, or a review by an expert in the field concerned. |
|
| 11 | What you asked me | | Reader replies received since edition #007 |
|
No reader reply arrived on edition #007 between October 11 and 17. So I have nothing to report this week, and I would rather say so than fill the space. Last week's question remains open: what is the task in your job that you would not hand to any tool? One sentence is enough. These answers shape the next issues, and the new nine-section layout is a first answer to those received so far: more cases by profession, official sources, shorter opinions. |
|
|
If I had to sum up this week in one sentence: prices and uses move fast, and checking remains human work. On prices, two announcements change the bill without changing the service: the doubling of Gemini 4 Argon's rate and the billing of blocked requests at Anthropic. On use, the ECB measures 52% of workers using AI, but the European Commission notes that 49% of SMEs do not trust AI for decisions about staff. The two figures do not contradict each other: people use a tool to write, not yet to judge. The same gap appears in governance. An agent exploiting exposed credentials without the victim noticing, invented references before a court, employee data sold without their consent: none of these cases comes from a technical failure. Each comes from the absence of someone who looks, at fixed intervals, at what the tool produces. For HR, executives and trainers, the practical consequence comes down to three points: budget with tomorrow's prices, name who signs, and check citations before they go out. That is the point of this edition's tutorial. See you next Sunday. Jérôme Your turn: reply with one sentence: "the task in my job that I would not hand to any tool is…". These concrete answers shape the next issues. |
|
|
Forward it to a colleagueForward it to someone who supports teams through change, who wonders what AI means for their own job, or who follows regulation closely. That is how SIGNAL AI finds its next readers, and its next questions. Once a week, on Sunday. Nothing else, ever. |
|
SIGNAL AI, Jérôme Denis's AI watch, JDENIS Consulting, Toulouse.
7 Bis Chemin de la Butte, 31400 Toulouse, France.
You receive this newsletter because you subscribed on jaydenis.com. Your address is not used for anything else.
Unsubscribe · Read this edition online · Reply to Jérôme |
|