What Do the WEF, MIT, and the EU Parliament Actually Disagree About on AI?
I’ve now read the Caribbean AI Summit agenda more times than anyone alive, including the people who wrote it, and I still get the room numbers wrong (Business is Ballroom B, Technical is 209, Society is 208, and I’m going to have that tattooed somewhere by October 9). That’s most of what co-chairing has meant this week, plus learning where the coffee is.
But reading the same agenda 40 times does something useful, which is that you start to notice who’s going to be in the building at the same time and what they’d say to each other if you put them at the same lunch table (I’m a D&D Dungeon Master, so seating interesting people at one table and seeing what happens is more or less my hobby). And what I keep noticing is that the 3 names the summit led with don’t agree on AI, and they disagree in the interesting way, where everybody is serious and everybody has done the work.
So let me tell you where I think the fault lines are, with the honest caveat that I haven’t asked them (Edward, you co-chair the thing, you have their email addresses, and you could have just asked them instead of writing 1,400 words guessing what they think, but apparently that’s not how you do things).
Start with Kay Firth-Butterfield, who was Head of AI at the World Economic Forum and now runs Good Tech Advisory. She closes the Technical track on Day 2 at 4:45 with “The Global State of AI: Governance, Risk & What’s Next.” Her book from January, Coexisting With AI (a better title than any of my 15 books, which all start with some version of “Look Smarter Than You Are”), argues that organizations have been prioritizing how fast they deploy over how mature their governance is, and she put it plainly in an interview about the book: “We need to understand what is truly at stake before we allow machines to shape outcomes that affect people’s lives.” That’s a before-you-ship position, if I’m summarizing her fairly (and she can correct me in person on the 10th).
At exactly the same time, in a different room, Idoia Salazar closes the Business track with “Ethics & EU Regulation: the Global Playbook.” She founded OdiseIA, she’s on the expert team for AI regulation at the European Parliament’s AI Observatory, and she works on Spain’s AI regulatory sandbox, which means she spends her actual working days on the question of how you write a rule for something that keeps changing shape. Europe’s answer has been the AI Act, and lately the AI Act’s own dates have been moving. In May, EU lawmakers reached a provisional deal to push the rules for stand-alone high-risk AI systems from August 2, 2026 to December 2, 2027, mostly because the technical standards you’d need to comply with them weren’t ready yet. (I ran interRel for 25 years, so I’m in no position to judge anybody else’s project dates.)
And then there’s Fabio Duarte from MIT’s Senseable City Lab, who’s in the opening keynote block on Day 1 with “AI for Urban Biodiversity.” His lab builds vision models that identify insects in city gardens, running on cheap low-power hardware, so a city can actually count what’s living in it before it makes a planning decision. Fabio’s whole posture, as far as I can read it from the outside, is to go measure something nobody was measuring and let the data tell you what the question should have been.
Everyone in the building will agree that AI should be safe before the first coffee break, and so will you. Put those 3 side by side, though, and I think where they’d actually argue is timing, meaning when do you know enough to decide (as a Douglas Adams fan, I’m obligated to point out that the answer is not 42), and I don’t think any of them is wrong, which is exactly why I want to hear them argue. I went to Shimer College, a Great Books school that ran entirely on the Socratic method, so smart people disagreeing politely across a table is basically my comfort zone.
If you run finance and you’d rather not fly to Puerto Rico to hear academics disagree (although we have great beaches, and I’ll gladly point you at one), this is still your argument. The CFOs I talk to are living inside it right now. Do you write the AI policy before the pilot, or after? And do you wait for the regulators to tell you what “high-risk” means in December 2027, or decide for yourself in October 2026? The Business track has a Microsoft session on Day 1 at 11:45 called “Why Your AI Pilot Never Made It to Production,” and I’d bet a good dinner (vegetarian, since I’m the one paying if I lose) the answer involves somebody picking wrong on one of those 2 questions.
I’ll tell you where I land, with the warning that I’m a guy whose website gets redesigned every Saturday at 5AM by a team of AI agents, so my risk tolerance is not a reliable benchmark for a public company. I’m an evidence-first person. I didn’t trust any of my agents until I’d watched them get things wrong in the wild, and the things they got wrong were never the things I’d have written a policy about in advance (my meeting-prep agent once briefed me thoroughly on the wrong human being, and then did it again, and my image agent still puts me in a baseball cap instead of my fedora about 1 time in 8, and I’d never have thought to write a rule for either one, and Edward, you realize you’re now making Kay’s argument for her). So my bias runs toward Fabio’s garden, where you watch for a while and write the rule afterward.
That said, Kay’s point is the one that keeps me honest, because “watch it in the wild” is a perfectly fine plan for a website and a terrible plan for a loan decision. Fabio’s garden can get a beetle wrong and nobody calls a lawyer, which isn’t true of your credit model. And the December 2027 date is the most underrated planning fact in finance right now, because 14 months is roughly one budget cycle and one bad audit, and anybody who waits for the final text before starting to document what their models do is going to be writing that documentation in a hurry.
So here’s a prediction, with a date on it so you can come back and laugh at me. When Europe’s high-risk rules finally apply on December 2, 2027, the companies that are ready will be the ones who started writing down what their AI got wrong sometime around now, for their own reasons, and then found out Europe wanted roughly the same list. If I’m wrong, I’ll write the follow-up post explaining why, and it’ll be long, and it’ll have a lot of parentheses in it.
The other thing I’ll say about this lineup is that people from the World Economic Forum, MIT, the EU Parliament, Carnegie Mellon, Penn, UC Berkeley, and SMU (my own data science alma mater, so I’m contractually obligated to mention them) said yes to a first-year conference on an island, which still slightly amazes me, and I think it’s because the crowd down here has gotten interesting enough to be worth the flight. I’ve lived in San Juan since early 2023, and I’ve watched the Puerto Rico AI Community go from a meetup to more than 1,000 different people through the door (JD Torres and I are giving the welcome at 9:00 on Day 1, so come early and say hello), and the summit is us inviting the rest of the world over to see it.
The bad news is that Kay and Idoia close the summit at the same time in different rooms, so you’ll have to pick one, and so will I. I genuinely haven’t decided yet, and I’d love your vote! The summit is October 9 and 10 at the Puerto Rico Convention Center, every session is subtitled live in English and Spanish, and tickets are on the summit site. If you’re coming, find me between sessions (I’ll be the one in the fedora). If you can’t make it, write me at Edward@Roske.AI and I’ll tell you how it went.
Asking good questions, Edward