This interview with Microsoft CTO Kevin Scott argues that AI must become a democratized platform, like the steam engine, not a tool for a few companies. He believes successful platforms let others build more value than the platform owner itself. He also suggests data should be treated as labor with compensation. He is bullish on Microsoft Azure, Office 365, and Teams, seeing AI boosting their productivity.
Kevin Scott, Microsoft's CTO, discussed on the Lex Fridman podcast the company's extensive work across cloud computing, SaaS services, operating systems, hardware (such as high-end PCs and headsets), research (ranging from economics to AI), GitHub, LinkedIn, search advertising, and gaming (including
Kevin Scott, Microsoft CTO, former Senior Vice President of Engineering at LinkedIn and head of mobile advertising engineering at Google. This episode's main thread: Microsoft's historical legacy as a platform company and its positioning in the AI era, along with how technology addresses societal challenges. The most weighty judgment in the entire episode: Kevin Scott argues that AI must become a "platform" rather than a "product"—a successful platform should create far greater economic value for those building on it than for the platform owner itself, and Microsoft has historically operated on this logic.
Kevin Scott argues that AI must ultimately be "democratized" like the steam engine, becoming a ubiquitous foundational component rather than being monopolized by a few companies.
Scott cites Bill Gates' classic assertion: "A successful platform is measured by the economic value it creates for those building on it, far exceeding the value created by the platform owner itself." He applies this framework to AI: AI should not be a proprietary tool for a handful of companies, but rather "a platform for others to build businesses, achieve creative goals, start ventures, and solve problems at work and in life."
Historical analogy: Scott compares AI to the steam engine — in the late 18th century, as the first machine to replace human labor at scale, the steam engine initially delivered value only to a few with capital, but eventually became "fully democratized" as a component of the modern world. AI must follow the same path.
Data support: Scott notes that global internet penetration has now reached 50%, connecting approximately 3.5 billion people. The democratization of AI means "not just tens of thousands of people having interesting tools, but millions of people having access to them."
Falsification condition: If AI infrastructure and decision-making power become concentrated in "a few companies in a few cities," Scott's thesis would be disproven — he believes this would be detrimental to both economic fairness and innovation.
Scott introduces Microsoft's "Data Dignity" research, conducted in collaboration with economist Glenn Weyl and VR pioneer Jaron Lanier, which seeks to establish a mechanism for evaluating and compensating data contributions.
Core Mechanism: Scott likens data to "raw materials consumed by AI machines," but notes the current lack of a transparent market to assess the value of these data contributions. He distinguishes between two types of data contributions:
Real-World Example: Scott mentions companies like Scale AI (as well as numerous similar firms in China) that are already engaged in data annotation. Workers are compensated for providing labeled data — a tangible version of "data as labor."
Uncertainty: Scott acknowledges this is "very difficult," because the value of data only emerges when aggregated, and there is a lack of transparency in tracking how data is used and how value flows.
Reader's Note: Scott speaks here from the dual perspective of a Microsoft executive and a collaborator. The "Data Dignity" framework itself carries an undertone of defending the data collection practices of large technology platforms.
Scott argues that vertical social networks (such as LinkedIn and Xbox), due to their clear community objectives, have fewer dimensions of content safety issues compared to general-purpose platforms, making AI decision-making easier.
Comparative Analysis:
| Dimension | Xbox Gaming Social | |
|---|---|---|
| Core Objective | Connecting people with opportunities (job seeking, mentorship, sales leads) | Playing games, entertainment |
| Content Boundaries | Professional identity, professional communities | Game-related interactions |
| Safety Focus | Inappropriate/offensive comments, dangerous content, illegal content | Anti-bullying (distinguishing between "playful teasing" and "genuine bullying") |
Policy Call: Scott emphasizes the need for policymakers to clarify "where the line is," because "in a democracy, you don't want a group of people making decisions unilaterally." However, some platforms are currently forced to make unilateral decisions because policy-making is "not fast enough."
Historical Perspective: Scott uses the history of written communication to illustrate the issue—from papyrus 3,000 years ago, the printing press 500 years ago, the offset printing press in the late 19th century, to the PC 50 years ago. It took humanity 3,000 years to establish mechanisms such as journalistic ethics, editorial integrity, and peer review, while the digital age has connected 3.5 billion people globally within half a century. "We basically have 3,000 years of work to do, but we may only have the next ten years."
Microsoft has clearly drawn a red line for the use of facial recognition technology and has called on governments to establish a democratic regulatory framework, while leveraging technology itself to address the problems it creates.
Microsoft’s Position: Scott cites Brad Smith’s blog post from fall 2018, stating that Microsoft believes certain uses of facial recognition “should not be permitted” and require regulation. In the short term, in the absence of regulation, Microsoft has drawn its own red lines — for example, adopting an “overly cautious” stance toward certain law enforcement applications, preferring to err on the side of over-restriction.
Bias Issues: Scott notes that models may learn “strange things,” such as “all doctors are male.” Microsoft is using synthetic visual data generated by GANs to train facial recognition models in order to eliminate bias — this represents a “super good use” of deepfake technology.
Deepfake Countermeasures: Scott proposes a “chain of provenance” approach — using encryption and social networks to establish a complete “chain of custody” for content, allowing users to verify the source of content (e.g., “this really came from the White House”). He acknowledges that technical detection may always lag behind generation technology, but the provenance approach is feasible.
Uncertainty: Scott admits that “we are not ready” to address the challenge deepfakes pose to “the nature of truth.”
Scott summarizes three keys to leading tens of thousands of engineers: forward-looking infrastructure investment, a clear mission, and "story" as the core tool for large-scale coordination.
Three Key Elements:
1. Forward-Looking Infrastructure Investment: This includes not only technical infrastructure such as storage systems and cloud APIs, but also development processes, tools, and collaborative culture. Scott warns: "At a small scale, you can improvise, but the greatest pain comes from not forming a clear enough opinion early on, and then accumulating technical and cultural debt."
2. Clear Mission: Not a "cute slogan," but something that provides clear direction. Scott cites Yuval Harari's core argument in Sapiens — that story is the essential tool for coordinating large-scale human activity. Once the Dunbar number (approximately 150 people) is exceeded, failure without a shared goal becomes "catastrophic."
3. The Power of Story: Scott views currency, constitutions, and laws all as "stories" — "If we don't believe in them, they are nothing."
| Position | Analyst View | Key Data |
|---|---|---|
| Microsoft Azure Cloud | Bullish (as platform infrastructure) | Specific scale not disclosed |
| Microsoft Office 365 | Bullish (AI integration continuously enhances productivity) | Fluid Framework (collaborative editing framework) under development |
| Microsoft Teams | Bullish (as collaboration hub) | Increasing Office usage occurs within Teams |
| HoloLens | Bullish (has generated real-world impact) | "Selling large volumes of devices," primarily targeting technical workers such as factory maintenance personnel |
| Bullish (exemplar of vertical social network) | Specific data not disclosed | |
| Xbox/xCloud | Neutral (content security challenges) | Specific data not disclosed |
| Scale AI | Neutral (mentioned as a data labeling case) | Specific data not disclosed |
1. Kevin Scott: A successful platform should create far more economic value for those building on it than for the platform owner itself. — Citing Bill Gates' classic assertion as the core philosophy of Microsoft's AI strategy.
2. Kevin Scott: AI must ultimately be "democratized" like the steam engine, becoming an ubiquitous foundational component. — Historical analogy: The steam engine evolved from a capital-intensive tool into a common component of the modern world; AI must follow the same path.
3. Kevin Scott: Data should be treated as "labor," requiring transparent market mechanisms to evaluate and compensate data contributions. — The "Data Dignity" project, in collaboration with Glenn Weil and Jaron Lanier, aims to enable users to "potentially even make a living from the data they create."
4. Kevin Scott: Content safety issues in vertical social networks are easier to resolve than on general-purpose platforms, as they have clearer community goals. — LinkedIn (connecting people with opportunities) and Xbox (entertainment) each have well-defined boundaries, allowing AI to make better automated decisions within those scopes.
5. Kevin Scott: It took humanity 3,000 years to establish journalistic ethics and peer review mechanisms, but the digital age connected 3.5 billion people globally in half a century; we may have only a decade to solve content governance issues. — A sense of urgency from a historical perspective.
6. Kevin Scott: The solution to deepfakes is not technical detection, but chain-of-custody tracing—using encryption and social networks to establish a complete provenance chain for content. — "You can click on something and see a verified chain of custody showing that this came from an identity you trust."
7. Kevin Scott: Three elements of large-scale engineering leadership—forward-looking infrastructure investment, a clear mission, and "story" as a coordination tool. — Citing Yuval Harari: Stories are the essential tool for coordinating groups larger than Dunbar's number (approximately 150 people).
8. Kevin Scott: Technological optimism is a self-fulfilling prophecy—"We are the stories we tell ourselves; if we become too pessimistic, we may fail to prepare for the best possible future." — A call to maintain hope, citing the Green Revolution as a historical precedent for technology solving global problems.