Microsoft turns Copilot into a governed model shelf while a court prices careless AI use

September 13, 2026

Microsoft put Grok from SpaceXAI into limited preview for some Microsoft Frontier Program customers as an optional model inside Copilot for Word, Excel and PowerPoint, with the choice controlled by an admin setting and unavailable in the EU, EFTA and the UK.

This looks like another step away from the idea that a work assistant comes with one fixed model. The practical change is that model choice is moving into the software people already use, while the employer keeps the switch. That matters beyond Microsoft because it makes cloud control - who may use which model, and where - part of the product, not a separate procurement task.

The New Mexico Supreme Court fined defence lawyer Stephen Aarons $5,000, held him in contempt and struck his earlier briefs after he filed a murder-appeal brief containing fabricated witnesses and police testimony generated with ChatGPT without verification. This suggests the cost of careless generative-AI use is becoming concrete in ordinary professional work, not only in lab-safety debates or company policy memos. For anyone deploying these tools, the pressure is likely to move from broad warnings to checked workflows: somebody has to verify output before it can enter a filing, a report or a record.

Anthropic CEO Dario Amodei announced a plan to invite outside embedded evaluators into Anthropic with access comparable to internal risk teams and urged governments to require similar measures at leading AI labs. OpenAI CEO Sam Altman said OpenAI has confidentially filed for an IPO but will not go public in 2026.