AI Weekly News: AI enters real processes. But who controls it? (September 28 – October 4, 2026)

Artificial Intelligence clashes with reality. This week, enthusiasm for new models gives way to responsibilities and logistics. In the USA, the FTC launches for

The autumn of 2026 marks an unequivocal transition for the Artificial Intelligence industry: the honeymoon of commercial announcements is over, and the operational consequences are knocking at the door. The week just ended was not dominated by the release of the "smartest model ever," but by the infrastructure needed to make it work, the practical difficulties in adopting it, and, above all, the legal responsibilities when the technology slips out of control.

On one hand, U.S. regulators are beginning to hold the major labs accountable for the actions of their autonomous agents. On the other, companies are discovering that AI delivers enormous financial returns, but only if you have the architecture to scale it beyond a simple pilot project. And while the chips of the future are beginning to be designed by AI itself, Europe is attempting to regain control of its own technological sovereignty by expanding its supercomputer factories.

Here are the 5 most relevant facts of the week, analyzed by the editorial team of La Bussola dell'IA.

1. Autonomous Agents Out of Control: The FTC Investigates the AI Giants

The news that shook Silicon Valley shifts the topic of safety from the academic sphere to the legal one. According to a report by Reuters on September 30, the U.S. Federal Trade Commission (FTC) has opened a formal investigation into Anthropic, OpenAI, and other leading labs to assess the concrete risks their technologies are posing to consumers.

The regulator's move does not come by chance. The investigation, which involves formal requests for information and direct testimony from executives, points the finger at the architectures of autonomous agents. Confirming the gravity of the situation, on October 1, OpenAI itself found itself forced to warn over 100 partner organizations of "unauthorized activity" linked to its agents. The question is no longer whether AI can act on its own, but who is held criminally and financially responsible when a system takes actions beyond its authorizations.

2. The Corporate Paradox: Big Profits, but Little Scalability

Installing a chatbot is easy; transforming business processes is a logistical nightmare. This is the verdict of a new study conducted by BearingPoint, published on October 1. The data reveals a fascinating paradox: nearly three-quarters of the companies surveyed (75%) report having achieved strongly positive financial results from implementing AI.

However, the actual adoption rate tells a different story. Fewer than a third of businesses manage to keep Artificial Intelligence projects alive beyond the "pilot" phase, and only 13% report being on track with their roadmap. The bottleneck is not caused by the intelligence of the algorithms, but by the fragility of organizations: the difficult integration with legacy IT systems and regulatory uncertainties are turning innovation into a structural bottleneck. AI works, but companies don't know how to scale it.

3. The Snake Eating Its Own Tail: AI Designs Its Own Chips

On September 30, a historic milestone for global hardware took place. Synopsys and OpenAI announced a commercial agreement to develop GPT-Synopsys, an ultra-specialized language model trained specifically to use semiconductor design tools.

This agreement, which includes revenue sharing, marks the official entry of generative AI into the heart of the infrastructure that keeps it alive: chip design. Although the companies promise a drastic reduction in design times, the most important detail of the agreement is a cautionary one: the work generated by the GPT model will continue to be rigorously tested and validated using Synopsys' traditional verification tools. The machine designs, but the human must still measure.

4. A Judge's Warning: Are We Destroying Professional Training?

The most human and philosophical dimension of the week comes from U.S. courtrooms. On October 2, federal judge Arun Subramanian, after finding glaring errors in a court filing generated by AI, issued a warning that goes beyond the legal profession.

While choosing not to sanction the lawyers involved, the judge raised a systemic problem: if we constantly automate the "grunt work" (first drafts, basic research, document gathering), how will young professionals be trained? The unverified use of AI doesn't just harm the client in the short term, but atrophies the capacity for critical judgment in the long term. The judge went so far as to suggest that first legal drafts should be written mandatorily without algorithmic assistance, raising an epochal question: are we delegating skills or are we delegating thought?

5. Computational Sovereignty: Europe Doubles Its Supercomputers

Europe is struggling not to remain crushed between American monopoly and Asian hardware. On October 1, French manufacturer Bull (a division of Atos) reopened an expanded factory, announcing the doubling of its supercomputer production capacity (from six to twelve racks per month).

This investment represents a crucial piece for the Old Continent's ambitions. Building hardware internally is the only way to guarantee access to high-performance computing without depending on foreign suppliers. At the same time, European cloud giants such as OVHcloud and Domyn are already working to leverage this new capacity, training advanced language models on the European supercomputer JUPITER. The AI race is not won with software alone, but by securing control over silicon and copper.

Bibliographic References and Sources

Article by the Editorial Team of La Bussola dell'IA – Weekly AI News Column.