62,000 workers just got a seat at the table on AI 🌿

Most stories about AI at work are stories about AI done to workers. This week there is a different one.

Kaiser Permanente and the Alliance of Health Care Unions have agreed what MIT Sloan describes as the first comprehensive deal by a major US healthcare employer to build worker input into every stage of AI decisions. It covers 62,000 workers.

The structure is concrete. A 10-person AI Task Force, five from labour and five from management, oversees the process. More than 3,500 unit-based teams will identify where AI should and should not be used in their own work. Peer advisers will help colleagues adapt. The model borrows from how Kaiser and its unions handled the rollout of electronic health records 20 years ago.

The people closest to the work usually know best where a tool helps and where it creates risk. A nurse or technician can spot a bad fit long before a dashboard does. Putting them in the room at the start is cheaper than repairing trust afterwards. MIT Sloan's authors note that similar agreements are starting to appear in hospitality and entertainment, which suggests this could become a template.

This is a governance structure. It tells us who gets a say, not whether the decisions will be good. Task forces can stall, and 3,500 teams is a lot of meetings.

The test is simple. In a year, can Kaiser point to a specific AI tool its workers chose not to use, and one they asked for? If yes, this is the real thing.

Google DeepMind published SynthID Bio in Nature, a method that watermarks AI-designed protein sequences. In lab tests, watermarked binders performed as well as unmarked ones, and the mark survived synthesis. It gives gene-synthesis screeners a way to trace a design's origin. DeepMind says robustness against deliberate tampering is unsolved. This is DeepMind's own announcement, though peer-reviewed.

Source: Google DeepMind

GE HealthCare and Mass General Brigham are exploring generative AI that lets oncologists search both structured data and clinical notes in radiation therapy workflows. The encouraging part is the earlier result: Mass General showed the company's workflow software cut the time from intake to treatment from 30 days to 8. The new project is exploratory, with no results yet.

Source: MedTech Dive

OpenAI is working with America's SBDC network to train about 150 advisors and run hands-on workshops for at least 1,000 small businesses. The idea is reach: owners already trust their local advisor. The usage figures are OpenAI's own, and the programme is modest in scale, but the on-ramp design is sensible.

Source: OpenAI

Six AI giants pledged. Who checks the checkers?

Anthropic, OpenAI, Google, Meta, xAI and Nvidia have signed the "Joint Commitment on Frontier Responsibilities." The pledge promises robust internal controls on cyber, biological, and chemical risks, board-level oversight, independent external audits, and regular meetings on shared standards. A day later, the FTC opened a broad investigation into the safety practices of OpenAI and Anthropic. The timing makes this week a useful case study in what a voluntary pledge is worth.

Our PM + Risk Manager lens

A product manager reads this as a commitments list, and the useful question is which ones can be measured. "Board-level oversight" and "independent audits" are specific enough to check: either the board has a standing safety agenda, or it does not, and either an auditor's report exists, or it does not. "Robust internal controls" is not. The pledge is strongest where it names a mechanism and weakest where it names an adjective. If you are a customer or partner of any of the six, you can now ask for those artefacts in a procurement review, and that is a real change in leverage. Watch whether audit scope and findings are shared publicly.

A risk manager sees the familiar structural weakness. The pledge is voluntary, the companies drafted it, and the companies hire the auditors. In financial services, we learned that this arrangement can work, but only with auditor independence, defined standards, consequences for failure, and some form of disclosure. None of those are in the headline terms. The shared-standards meetings could help, or they could become a forum for converging on the minimum. The FTC investigation is a reminder that outside pressure, not self-regulation alone, is likely to determine what the commitments mean in practice.

The next 12 to 24 months

The things to watch are concrete. Do audit reports become public, even in summary? Do any signatories publish what an audit found, including something uncomfortable? Does a shared standard emerge that is stricter than any one company would have chosen alone? A start is not a win, but it is something to hold them to, and that is the reason it earns a place here.

Sources: Al Jazeera · Fortune

InstructMesh, from researchers at MIT CSAIL, Google and Northeastern, generates 3D models from text and lets you repair flaws in plain language. Nearly 80% of popular Thingiverse models had structural problems, and novices found and fixed issues about 90% of the time. Demonstrated uses include custom items and medical devices such as knee braces.

Why it matters: lowering the skill bar for making usable physical objects, from assistive gear to household fixes, is a quiet access win. Fair warning: this is a research prototype for UIST in November, so you cannot download it today.

→ Read more: MIT News

If your employer wrote an AI rulebook tomorrow, what is one decision you would want your team, not your manager, to make?

Hit reply. We read every response.