As governments consider how to regulate artificial intelligence, executives from major AI companies are making their views known. Their public arguments put a central question in focus: how can lawmakers set safeguards while accounting for the commercial interests of the companies building these systems?
Executives press their views on regulation
OpenAI CEO Sam Altman, Google CEO Sundar Pichai and Microsoft’s Brad Smith each spoke with policymakers or reporters during the week. Altman warned in London that the EU’s proposed AI Act could lead OpenAI to withdraw its services from the bloc if the company could not comply. He said, “We will try to comply, but if we can’t comply we will cease operating.”
Pichai also called for “appropriate” guardrails, while arguing that rules should not hinder innovation. Smith, speaking with lawmakers in Washington, put forward a five-point plan for public governance of AI. Their approaches differed, but all three executives presented themselves as willing to accept regulation while emphasizing limits they believe rules should respect.
Those limits matter because regulation can affect how companies build and deploy AI, and the costs of doing so. The source article describes a debate over whether AI training on copyrighted material is allowed under the fair use doctrine in the U.S. Smith did not address that unresolved question. Strict licensing rules for training data, if imposed at the federal level, could be costly for Microsoft and other companies using similar methods.
Transparency and responsibility are contested
Altman appeared to question parts of the EU proposal that would require companies to publish summaries of copyrighted data used to train AI models. He also raised concerns about making companies partly responsible for how their systems are used downstream.
The proposal’s requirements to reduce energy consumption and resource use in AI training also drew scrutiny. Training can require substantial computing resources, so limits in this area could shape how firms develop models. The article does not settle how such obligations should be designed; it shows that companies are already challenging provisions they see as burdensome.
In the U.S., Altman addressed members of the Senate Judiciary Committee with statements about AI risks and recommendations for regulation. Sen. John Kennedy (R-LA) urged the industry representatives to explain in plain English what rules lawmakers should implement. The exchange highlights a tension: legislators are asking the companies to help describe the technology and its risks, while also deciding how much authority and influence those companies should have in writing the rules.
Suresh Venkatasubramanian, Brown University’s director of the Center for Tech Responsibility, captured the concern with a warning: “We don’t ask arsonists to be in charge of the fire department.” The metaphor points to a question of trust and accountability. Companies have technical experience that can inform policy, but they also have commercial interests affected by the outcome.
Research broadens the AI conversation
Beyond regulation, the week’s developments showed how varied AI research has become. Google DeepMind and collaborators developed a framework for evaluating “extreme risks,” including capabilities related to manipulation, deception and cyber-offense. The work reflects an effort to make risk assessment more systematic, although the article describes the science as still evolving.
Researchers at SLAC used machine learning to infer complex particle beam shapes from small amounts of data. Their model could reduce the data and computing effort usually needed to predict beam shape, with a future step being experimental demonstration on full 6D phase space distributions.
Adobe Research and MIT tackled a computer vision task: identifying which parts of an image show the same material. They built a synthetic dataset, then fine-tuned an existing vision model to perform the task. The article notes that the practical use remains unclear, but the work addresses a subtle distinction between an object’s material and its color or other visual properties.
More languages and new settings for AI
BLOOMChat, built on BLOOM, was reported to support 46 languages and to be competitive with GPT-4 and other models. It remains experimental, so the article cautions against using it in production. Its multilingual coverage could, however, help teams test AI-related products across languages such as Spanish, Japanese and Hindi.
Other projects point to AI operating in specialized or constrained environments. Geolabe is using satellite data to detect and predict groundwater variation, while Zeus AI is working on generating 3D atmospheric profiles from satellite imagery. IEEE researchers are pursuing a neuromorphic processor designed to use limited size, weight and power for training models in space.
Picknick is exploring how autonomous robots in high-stakes settings might communicate their intentions visually to human supervisors, reducing the need for intervention. Together, these efforts show that AI policy debates are unfolding alongside research into concrete applications. The rules lawmakers choose may influence how such systems are trained, evaluated and deployed, while the technology itself continues to spread into new fields.