Artificial intelligence has become part of everyday work faster than the institutions responsible for deploying and governing it have adapted. That is a central tension in the 2026 report from the One Hundred Year Study on Artificial Intelligence, or AI100, which examines both the gains from generative AI and the unresolved questions surrounding its economic and societal impact.

Titled AI Goes Mainstream: The Scaling Era and the Rise of Generative AI, the study is the initiative’s third report. AI100 is administered out of Stanford’s Institute for Human-Centered Artificial Intelligence and brings together researchers across disciplines. The 2026 panel, chaired by Oxford’s Michael Wooldridge, offers an assessment of a technology whose expanding capabilities have made questions about value, reliability and accountability more immediate.

The Next Advances May Depend on More Than Scale

The panel describes recent progress as the product of “the scaling era of AI.” Larger training datasets, more powerful chips and substantial investment enabled models capable of performing a wide range of tasks. Yet the researchers argue that extending that approach faces constraints, including access to high-quality data and the resources required to train frontier systems.

In its technical analysis, the report describes a more complicated relationship between size and useful performance. Data curation, efficient training and additional computation when a model answers a question are becoming important sources of improvement. Smaller systems can sometimes outperform larger ones. The authors do not establish a definitive end to scaling; they describe a field exploring additional routes to capability.

For enterprise leaders, this suggests that model selection requires more than following a leaderboard. A useful evaluation would examine performance on the organization’s work, the context available to the system and its operating cost. That is an implication of the findings, rather than a prescription for choosing a particular vendor or architecture.

Investment Has Outpaced Evidence of Broad Returns

The report draws a distinction between spending on AI and the value ultimately produced by it. Its authors describe an infrastructure boom with consequences for capital markets, electricity demand and freshwater supplies. But they argue that the effects beyond investment and valuations are not yet clearly visible in aggregate economic data or corporate returns.

“AI is not yet paying its way,” the panel wrote. That broad assessment should not be read as evidence that no individual deployment delivers a return. It raises a different question: whether demonstrated benefits, across enough activities and organizations, will justify the scale of the commitments being made.

At the same time, Stanford HAI’s accompanying overview highlights substantial opportunities in science. It describes research timelines being compressed and the panel’s expectation that AI-enabled laboratories will become widespread within a decade. Those are prospects requiring validation, not guaranteed commercial outcomes. Organizations pursuing them will need ways to distinguish faster generation of ideas from discoveries that withstand scrutiny.

Capability Does Not Resolve Reliability or Trust

The report’s conclusions caution that “current AI systems remain jagged.” Strong performance on one task can coexist with errors, bias or weak reasoning on another. The panel also identifies challenges involving privacy, copyright and systems that exploit shortcuts instead of accomplishing their intended objectives.

Stanford HAI’s summary emphasizes cybersecurity as an area where AI strengthens both attackers and defenders. It also describes the growing difficulty of distinguishing authentic recordings from synthetic images, audio and video. These developments put pressure on established ways of verifying information and deciding whom to trust, extending the consequences beyond the technology department.

The practical implication for organizations is to evaluate a system in the setting where it will be used. A convincing demonstration does not establish that a model can safely carry out a consequential workflow. Leaders need to decide what evidence is sufficient, which decisions require human judgment and how errors will be detected and addressed.

Governance Must Include the People Affected

The authors argue that AI’s direction should be treated as a “collective choice rather than an inevitability.” They describe public concerns about employment, education, well-being and the environment, alongside limited public influence over development and deployment. Their argument places participation within the governance discussion, rather than treating acceptance as something to secure after decisions have been made.

The panel also finds fragmented international oversight and disagreement among experts about future capabilities. Its conclusions acknowledge opportunities to improve health, scientific discovery and access to knowledge while warning about displacement, declining skills and unevenly distributed benefits. Stanford HAI’s overview further notes that geopolitical competition complicates efforts to reach enforceable international agreements.

For business leaders, the report provides a reason to connect technical experimentation with explicit decisions about accountability and value. The relevant questions are practical: whose work will change, who will verify results and how will the organization know that an improvement has occurred? Progress in model capability makes those decisions more urgent; it does not make them on an organization’s behalf.

Peter High is President of Metis Strategy, a business and IT advisory firm. He has written three bestselling books, including his latest Getting to Nimble. He also moderates the Technovation podcast series and speaks at conferences around the world. Follow him on Twitter @PeterAHigh.