Microsoft has spent years telling the world that artificial intelligence will remake work, medicine, and government. On a Monday morning at Asia Society, its president finally had to explain who’s supposed to clean up the mess if he’s wrong.
The Hon. Dr. Kevin Rudd AC – twice prime minister of Australia, now on his second tour as Asia Society’s president and CEO – opened with a joke about seniority. Asia Society: founded 1956. Microsoft: barely 50. Then the joke ended and the audit began. For the next hour, Rudd worked through Microsoft’s AI architecture layer by layer – infrastructure, models, data, apps – and at every layer asked some version of the same question: who’s accountable when this breaks?
Lulu C. Wang, Asia Society’s vice chair and global trustee, set the stakes before either man sat down, telling a room that included former president Josette Sheeran and trustee emerita Denise Tso that 75 years of institutional survival rests on one bet: conversation is worth having precisely when the stakes are highest and the outcome least certain.
Layer one: infrastructure. Data centers in more than 40 countries, originally built for video streaming and cloud storage, now repurposed to train frontier models. Layer two: the models – Microsoft’s own, plus OpenAI’s, Anthropic’s, and a fast-growing bench of Chinese competitors. Layer three, which Smith calls the ‘IQ layer’: your actual data – emails, docs, spreadsheets – the stuff that makes a model useful to you specifically, and has to be locked down exactly as hard as it gets used. Layer four: applications. Copilot. The thing most people think is AI, when really it’s just the visible tip.
It’s a clean org chart. It’s also a liability map with responsibility sliced four ways, so no single layer answers for the whole system. That got tested when Rudd pushed into the sharpest terrain of the morning: agents that, in Smith’s words, ‘break out, break in, cheat, lie’ – and what happens when a system starts improving itself faster than the company that built it, or any regulator watching it, can keep up.
Rudd asked for the highlight reel first. Smith wouldn’t play along with forecasting – ‘there’s no such thing as a crystal ball’ – and offered a spec sheet instead: wildfire cameras replacing watchtowers across California and the Australian bush; ambient AI that drafts a doctor’s notes during the appointment so the doctor can move to the next patient; radiology tools that catch lung cancer faster than a trained eye; a WhatsApp legal-translation tool built for Malawi, where roughly 700 lawyers cover the entire country and child marriage remains routine. All of it, he said, runs through an internal team called the AI for Good Lab, eight years old, working mostly with nonprofits.
The smallest example landed hardest: a self-built agent that runs every morning at 5 a.m. summarizing the 12 most important unanswered emails from the day before. He still prints it out. Asked where his own message would’ve ranked, he laughed: number three.
On safety, Smith reached for a hundred-year-old analogy. Around 1905, cars got fast enough to kill people at intersections. Nobody solved that by demanding car companies engineer the danger out of the vehicle – society built traffic lights and guardrails on the road instead. His argument: hyperscalers owe AI the same thing – monitoring agents, capping what they can spend so they don’t ‘burn up your AI bill without you knowing about it,’ building in kill switches at multiple levels – rather than dumping the entire safety job on the labs training the models.
Reasonable argument. Also convenient for a company that makes most of its AI money at the infrastructure layer, not the lab bench. Smith basically said as much, arguing Washington fixates on the handful of firms building frontier models and ignores everyone else in the pipeline.
On recursive self-improvement – a model training itself instead of being trained by people – Smith didn’t dodge it. ‘That makes most people pretty nervous,’ he said, calling for real rules on when it’s allowed and how a system heading off-course gets caught before it accelerates. He cited two data points: Anthropic opening its systems to third-party evaluators, and Accenture rolling out a business line built around human oversight of AI deployments. Watching the watchers, he said, is already becoming its own industry.
His best line came right after: nobody boards a plane without government inspection behind it, and nobody hesitates at a dairy case stocked with a dozen kinds of milk, because a shared health standard sits underneath every carton. ‘Do we really think,’ he asked, ‘that the most powerful technology on Earth is likely to be less regulated than a carton of milk?’
Before jobs, Rudd asked what parents in the room needed to hear. Smith opened with something close to an admission: the industry got ‘a little too exuberant’ about social media and phones in classrooms over the last 15 years, and the mental-health bill for kids came due. The response: a binding deal signed two weeks earlier with the American Federation of Teachers – 10 principles on safety, guardrails, privacy, and transparency to parents. The detail he kept circling back to: the tool finishes a task and stops, rather than keeping a kid glued to the screen instead of a teacher. It became legally binding across Microsoft’s school contracts on November 1, made public two hours after signing so competitors could be measured against it.
Rudd had his own gripe – as ‘an old-fashioned Australian country boy,’ he’s bothered every time a chatbot calls itself ‘I.’ An ‘I’ is a person, he argued, not software. Smith agreed: ‘AI is an it, not an I,’ while conceding the industry hasn’t settled the terminology fight at all.
Rudd’s toughest question was about employment, and Smith’s first pass – dignity in work, a New York Times piece on retirees who keep volunteering – didn’t satisfy him. Rudd brought up a panel in Deer Valley where he’d watched AI executives run ‘a thousand miles’ from the question of where the next decade’s jobs actually come from.
Smith’s answer was a story about horses. He argued the combustion engine helped cause the Great Depression: fewer horses meant less demand for oats and hay, farmers switched to cash crops, overproduction tanked prices, farmers defaulted, rural banks failed, and the collapse spread until the whole system cracked – a shock nobody saw coming because nobody was tracking the horse population. Rudd, who steered Australia through 2008 with no existing playbook, used it to name what the conversation was circling: structural adjustment, and what government owes the people caught inside it – tying it straight to populism on both the left and right.
Smith’s own proposals stayed deliberately half-formed: the roughly 1,100 community colleges already positioned to retrain workers, and a note that employer investment in job training climbed from 1980 to 2000, as PCs entered offices, then flattened. He floated rethinking payroll taxes – a tax on human labor at the exact moment AI makes that labor easier to replace – while stopping short of backing a tax on AI usage itself, calling it premature but not off the table.
Microsoft’s Community First Infrastructure Initiative, launched in January, is meant to smooth friction between data-center build-outs and the towns absorbing them. Smith named five original commitments – electricity, water, taxes, jobs, local investment – plus a sixth that’s surfaced only in the last six months: noise. His case study was Quincy, Washington, host to Microsoft data centers for 20 years: poverty cut in half, population outgrowing Seattle’s, the best public high school building in the state, a new police station, fire station, and aquatic center – and, he joked, the traffic-light count going from one to two. He predicted state and national rules will eventually lock in electricity-rate protections and water-use limits, calling that outcome, despite corporate instincts to resist regulation, the thing that actually buys public trust.
Rudd’s last question came with a grin: ‘It’s Washington. You’re Donald Trump. I’m Xi Jinping.’ Smith kept it modest – sustain the dialogue the two leaders opened in Beijing in May, add technical experts, agree on what counts as a frontier model, trade best practices. Longer term, he wants China and America’s traditional allies at the same table, not a deal struck bilaterally and left there.
Asia Society billed this as four topics, roughly equal weight: artificial intelligence, technology policy, cybersecurity, and geopolitics. What filled the hour leaned hard into topic one and mostly skipped the other three.
Cybersecurity never got specific – Rudd’s opening on agents ‘autonomously attacking infrastructure’ was as sharp a question as the morning produced, and Smith’s answer swerved back to governance and kill switches instead of the actual threat landscape. Technology policy in the normal sense – antitrust, chip export controls, the competition scrutiny Microsoft itself is under – wasn’t mentioned once. Geopolitics showed up only in the last sixty seconds, as a joke instead of a real conversation about Taiwan or chip supply chains. Human rights, digital safety beyond schoolkids, and immigration – all listed under Smith’s own portfolio in his official bio – didn’t come up at all.
None of that makes the hour a bust. What actually happened – on jobs, structural adjustment, and the small-town politics of hosting a data center – went deeper than the flyer promised. What was advertised mostly waited for a session that never showed.
What earned the sold-out sign wasn’t a single headline-grabbing line. It was watching two guys pull in different directions and land somewhere in the middle – Rudd pushing toward the bigger structural and geopolitical stakes, Smith pulling it back to what Microsoft is already doing about them. Seventy-five years into a bet that conversation still matters, this one didn’t settle who’s accountable for the machine. It made clear the people closest to answering that question are still figuring out the vocabulary – one small town, one vanished horse population at a time.
A version of this question was on my own list had the floor opened to the audience: why not turn the tool loose on its own mess – ask the AI itself to model which jobs its disruption creates, and route displaced workers toward them, rather than leaving that mapping to policy debates that move at legislative speed? There was no QandA session at this event, so it went unasked. But it’s worth sitting with, because it exposes the soft spot in Smith’s own framework. He described a system capable of drafting doctors’ notes, screening radiology scans, and translating legal rights into WhatsApp messages for women in Malawi – plainly capable, in other words, of pattern-matching at scale. Turning that same capability on the labor market itself, forecasting where the ‘new jobs’ he kept promising will actually materialize, was never proposed by either man on stage. Whether that’s an oversight or a tell – a company more comfortable describing AI’s power in the abstract than pointing it directly at its own economic fallout – is exactly the kind of question a live QandA exists to press. This one never got the chance.