The Signal — November 11, 2026

Six governments and ten AI companies have signed a French declaration that commits developers to fail-safe mechanisms for shutting down the models they control. OpenAI and Upwork published a study comparing AI-related freelance projects with other work clients post, and Google DeepMind's new robotics model sorted car parts on Mercedes-Benz production lines in a pilot.

The Paris Call commits AI developers to fail-safe shutdown mechanisms, but the US, Google and Meta didn't sign

France published the Paris Call for AI Safety and Security on November 9, after a meeting hosted by President Emmanuel Macron. The Élysée's signatory list names six governments (Bahrain, Canada, Finland, France, Norway and the Netherlands), ten companies including OpenAI, Anthropic, Microsoft, Mistral AI and Hugging Face, and four civil-society groups, among them Yoshua Bengio's LawZero. US Treasury Secretary Scott Bessent took part in the discussions, but the United States did not sign, and neither did Google, Meta or China, The Next Web reports. Le Monde, cited by The Decoder, counted representatives of eight countries at the meeting, and The Decoder notes that Germany didn't sign. The list remains open to new endorsements.

Under the declaration, developers commit to disclose their models' capabilities, limitations and known risks, to support evaluation by qualified independent evaluators "where appropriate and proportionate to the risk," and to "implement fail-safe mechanisms to shut down model operations within their control, including during training and internal use." Signatories also agree to publish non-sensitive evaluation results in a standard format, to share timely information on serious cyber and AI control incidents, and to work toward a common framework for evaluating frontier models. Governments commit to confidential channels for whistleblowers. OpenAI, one of the signatories, has described one version of a shutdown mechanism for its own research jobs: Control Room, which it says halts model output within 35 seconds of a decision to intervene, as we covered in the November 10 edition.

The declaration is voluntary. It sets no deadlines or sanctions, and the shared testing framework it proposes still has to be developed, according to The Next Web and The Decoder. CeSIA, a French AI safety group that The Next Web says signed the text, argues that it sets no common risk thresholds for deciding whether a model should be developed or released, and doesn't guarantee that independent testers or governments will get access to the systems they need to assess. CeSIA wants governments to have the authority to suspend a model when its risks become too high. "These decisions must rest with public authorities," said Charbel-Raphaël Ségerie, its executive director.

Sources: Élysée · The Next Web · The Decoder

OpenAI and Upwork found bigger budgets and more hours for AI-related freelance work, but similar posted hourly rates

OpenAI's economic research team and Upwork, the freelance marketplace, published a joint report on November 10 comparing what people do in ChatGPT with the work clients post on Upwork. The report counts a job as AI-related if it requests AI skills, requires AI tools or involves improving AI-generated output. Those jobs had posted fixed-price budgets about 52% higher than other jobs in the same work subcategory and posting month. On hired hourly contracts with recorded hours, freelancers logged about 41% more hours on AI-related work in the first 90 days, while posted hourly rates differed by only about 1.5% after the same adjustment. The authors suggest that clients may be commissioning larger projects when the work involves AI, rather than paying more for each hour.

The mix of work is changing too. Content writing's share of Upwork postings fell by roughly 31% between the first half of 2025 and the first half of 2026, and the key takeaways suggest some firms may be using AI for that work instead of hiring writers. The pattern isn't uniform: in 43 of 63 work categories, activities that were relatively common in ChatGPT lost share on Upwork, and in the other 20 they gained. Video and animation made up 1% of sampled ChatGPT conversations but 10% of Upwork job posts. AI-related postings also more often combine tasks that rarely appeared together before, such as AI-generated video with editing and motion graphics. Among postings with at least two tasks, such uncommon pairings made up 8.4% of task pairs in the average AI-related posting and 2.4% in other postings.

The authors say the price and hours comparisons "are descriptive and do not identify a causal effect." The budgets are what clients posted, not what freelancers were paid, and the price and hours data run through April 2026. The key takeaways go further, suggesting AI may be changing the scope of work "without reducing overall labor demand," which they say could help explain why unemployment hasn't risen much; the study's data don't test that. Both organizations also have an interest in the answer, since OpenAI makes ChatGPT and Upwork runs the marketplace being studied.

Sources: OpenAI · OpenAI and Upwork (full report) · OpenAI (key takeaways)

Google DeepMind reports a 96% success rate for Gemini Robotics 2.5 sorting car parts in a Mercedes-Benz pilot

Google DeepMind released Gemini Robotics 2.5 on November 10. It is a vision-language-action model, which maps camera frames, the robot's state and an instruction to motor commands. In a proof-of-concept on active production lines at Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany, Apptronik's Apollo 2 humanoid used the model for kitting: picking metal brackets out of bulk bins and placing them in trays in the order each vehicle needs. A separate DeepMind reasoning model, Gemini Robotics ER, kept track of the workflow. DeepMind reports a 96% success rate. Apptronik chief executive Jeff Cardenas said the models carried over to Mercedes-Benz production "with very little drop-off."

Across 200 tasks, from serving food with tongs to pipetting liquid, DeepMind says the model succeeds 30% to 100% more often than its previous baselines, depending on the type of task, and finishes 20% to 30% faster. Some tasks remain hard. In DeepMind's task-level comparison, as reproduced by AlphaSignal, folding pants and packing a rice cooker each went from 0% to 40%. DeepMind credits its Gemini 4 Argon model with speeding up development by generating thousands of simulated practice runs from real camera scenes. It also released ASIMOV-Jailbreak, an open benchmark of unsafe requests disguised as ordinary ones, on which Gemini 4 Argon resisted 99.5% of attempts; that figure describes Argon, not the new action model.

All of these numbers come from DeepMind. AlphaSignal points out that the release doesn't give the number of attempts, the run duration, how often humans stepped in or what counted as a failure, so the 96% can't yet be compared with other deployments. It covers one task in a pilot, not a reliability rate for regular production. The model is going to DeepMind's robotics partners, with no public API or pricing announced.

Sources: Google DeepMind · AlphaSignal

On the Editor's Desk

We held Meta's Muse Ember, a small model for agent subtasks priced at $0.10 per million input tokens, because many of its headline comparisons are Meta's internal evaluations and its cost chart uses estimates at proposed pricing. We'll look at it again when outside testing appears. OpenAI's Businesses in ChatGPT, a beta that lets companies claim profiles and lets eligible US businesses launch their own agents inside ChatGPT, is on hold until people have used it. We also skipped Google's wider release of Gemini 4 Carbon, which extends paid access to a model first announced in October without new capability claims.