noticias ia para empresas regulacion ia openai

The EU AI Act now reaches text: within weeks, what ChatGPT and Codex write in the European Union will carry an invisible watermark, across all plans. Schneider Electric is paying $22.6 billion for PTC, the software used to design many of the machines you buy, as industrial AI concentrates in fewer hands. And a study from Stanford, Berkeley, and Microsoft explains why only 11% of companies accurately predict their AI spending: the cheap model turns out to be more expensive in one out of three comparisons. This follows Monday's radar, October 5.

ChatGPT will hide an invisible mark in what it writes in the EU: OpenAI adds the mark, but the label is on you

OpenAI explained on Monday how it will comply with Article 50 of the European AI Act, which since August 2 has required AI-generated content to carry a machine-readable mark: in the coming weeks, text from ChatGPT and Codex will carry an invisible watermark for users in the European Union, across all plans; in the API, it is an option, disabled by default, for customers worldwide. The technique, which it calls textGrain, leaves a statistical signal in word choice. With a 1% false positive rate, its detector finds it in 80% of 200-token passages and 95% of 400-token ones, but if one in four words is changed for a synonym, detection drops from 92% to 17%. That's why, for now, the detector will only be available to researchers and organizations approved on a case-by-case basis, and OpenAI warns: the mark does not identify the user and, if missing, does not prove the text is human. Anthropic announced the same in August for Claude, worldwide. For your company, three things. The provider applies the mark; the visible label is yours in two cases set by the same article: if you publish fake but realistic images, audio, or video, and if you publish AI-generated texts to inform the public on matters of public interest, unless a person has reviewed them and someone assumes editorial responsibility. Violating Article 50 can cost up to 15 million euros or 3% of global turnover; for an SME, whichever is lower. If you offer customers a chatbot or text generator built on the API, check with your lawyer whether you're the one who has to mark it, because the option ships turned off. And do not judge candidates or providers with an AI detector: with these numbers, it proves nothing. The practical fix is a simple log: which pieces the AI creates, who reviews them, and under whose name they are published. AI text that goes out under your company name without anyone reviewing it is not UNATTENDED: it is unsigned. Source

Schneider Electric pays $22.6 billion for PTC: if you design with Creo, Windchill, or Onshape, your vendor is changing hands

Schneider Electric, the French electrical equipment and industrial automation group, announced on Monday the acquisition of PTC for $205 per share in cash: $22.6 billion (about 20.1 billion euros) for the equity and $23.7 billion including debt, with a 42% premium over the previous close. It is the largest acquisition in its history. PTC makes design software (CAD) and product lifecycle management (PLM) software for more than 30,000 industrial customers and booked about 2.4 billion euros in revenue in 2025. Schneider already owns AVEVA, is still closing its purchase of Norway's Cognite, an industrial AI specialist, and wants to combine design data with factory-floor data to give AI agents the "reliable context" they need; it expects 250 million euros in annual savings and about 800 million in additional sales. The deal is expected to close in the third quarter of 2027. For your company: if your engineering team works with Creo, Windchill, or Onshape, the vendor that holds your product know-how will change hands within a year, and the cross-selling Schneider seeks will come to you as bundle offers. Before renewing, three things: if you renew before closing, lock in the price for two or three years; verify that you can export your models and bills of materials in open formats, such as STEP, without asking for permission; and ask in writing what happens to your data if the product is integrated into another platform, as Schneider promises to keep everything "open by design." The bigger picture: industrial AI is being bought in layers—design, factory, data—and whoever holds your product data will decide which AI they sell you. A process that you cannot take out of your provider's platform is not INTEGRATED: it is locked in. Source

Only 11% of companies accurately predict their AI spending: the cheap model costs more in a third of comparisons

The Wall Street Journal reported on Monday that, in a survey of 396 companies, only 11% were able to accurately forecast their AI spending. Part of the explanation lies in a study by Stanford, Carnegie Mellon, Berkeley, and Microsoft Research: they measured eight reasoning models across twelve types of tasks, and in 32% of comparisons between two models, the lower-priced one ended up costing more, up to 28 times more. An example from the study itself: the price per token for Gemini 3 Flash is 80% cheaper than for GPT-5.4, but across the set of tasks, it turned out to be 38% more expensive. The culprit is the tokens the model spends "thinking" and the loops an agent goes through: for the same question, one model may use ten times more reasoning tokens than another, and running the same question again can cost up to 9.7 times more from one run to the next. "Price alone should not be used to infer which model is actually cheaper," says Lingjiao Chen, one of the authors. For your company: the price per million tokens on the provider's website does not tell you what your process will cost. Before choosing a model, take 20 real tasks from your business—a day's invoices, a week's queries—run them through the two candidates, and compare the total cost per completed task, not per token. Repeat the test a couple of times and budget a range, not a single figure. And ask your provider or integrator for the cost per request on the invoice, not just the monthly total. This is a step beyond the spending cap of Copilot's pay-as-you-go: the cap prevents a surprise, but it doesn't tell you which task is expensive. A process where you don't know the cost per task is not UNATTENDED: it is unbudgeted. Source

Cohere sets spending caps per person and per agent on its platform: this is the checklist you should demand from any provider

Cohere, the Canadian company that signed a merger with Germany's Aleph Alpha in September, unveiled North 2 on Monday, the new version of its platform for building and using AI agents in the enterprise. Its admin console now tracks token use per user and per agent, and lets you set limits per person or team, a company-wide cap, and alerts before you hit it. Each agent can carry its own guardrails—filtering personal data in questions and answers, and detecting attempts to sneak in instructions—and a policy that limits it to authorized actions; it runs multi-step processes on its own but asks a person to sign off on designated steps. It supports models from other vendors and runs in the cloud, on your own servers, or fully offline. For your company: even if you don't buy Cohere, this is the list you should put in front of any agent provider, from Microsoft to your integrator. Can I see spending per agent and per person? Can I set a cap for each one? Does it alert me before the limit? Which steps require human sign-off? What personal data does it filter? Can I change models without rebuilding the agent? Can it run on my servers if the data can't leave the building? Every "no" is a control the provider keeps, and as we said in Monday's flash, your agents' rules and logs have to be yours. An agent you can't cap or make ask for sign-off isn't UNATTENDED: it's running loose. Source

What to watch tomorrow?

When the watermark shows up in ChatGPT accounts in Spain and whether Google, Microsoft, or Mistral explain how they will mark their text; and the release of textGrain as open source, which OpenAI has promised so others can build on it.