Stanford AI100 Five-Year Report: AI Has Become Part of Daily Life, but Society Is Still Learning How to Cope with It

Five years ago, when discussing AI, many questions centered on whether it could “write, program, or participate in scientific research.” Today, these capabilities have made their way into the workplace, and more challenging questions have emerged: Who is responsible for errors? How should efficiency gains be distributed? And what new skills do people need to develop?

The third AI100 report, released on November 10, 2026, brings together the changes from 2021 to 2026 into a single overview. Titled *AI Goes Mainstream: The Scaling Era and the Rise of Generative AI*, the report was authored by an interdisciplinary research team led by Michael Wooldridge of the University of Oxford. AI100 is a long-term research project managed by Stanford HAI that conducts an assessment every five years; it is not the same report as the annually published AI Index.Introduction to the Launch of Stanford HAIThis explains the background.

Scale Still Matters, and the Path to Progress Is Widening

The report identifies computing power and data scale as key drivers of progress in generative AI over the past five years, while also noting that the supply of high-quality data and the cost of cutting-edge training are becoming limiting factors. The research team believes that future progress is likely to stem more from computational efficiency during inference, context and memory, algorithmic efficiency, and the combination of tools and agents. This is an assessment of the technological roadmap and should not be simplistically interpreted as meaning that “large models have reached their limit.”Key Findings of the ReportThis hierarchy has been preserved.

In the context of enterprise selection, this means that the evaluation should focus on the entire workflow. The score a single model achieves on a test does not necessarily reflect its ability to read the correct files, invoke the appropriate tools, handle missing information, and deliver verifiable results. If procurement decisions are based solely on comparing specifications or rankings, there is a high likelihood of overlooking the factors that truly determine the cost of implementation.

Capacity growth and reliability issues coexist

The AI100 report does not characterize current systems as stable, reliable, general-purpose experts: the same system may excel at one task while exhibiting factual errors, biases, or common-sense mistakes in a related task. The research team also noted divergent views on the timeline for AGI and did not reach a consensus prediction that it “will inevitably be achieved within five years.”Conclusions Section of the ReportThere has been explicit discussion on this matter.

This is a very practical reminder for everyday users. For example, when organizing interview transcripts, you can have the AI summarize the main themes first, but you’ll still need to listen to the recordings yourself to verify key quotes; when writing an analysis, you can have the AI suggest a structure, but you’ll still need to refer back to the original sources to verify the data supporting your conclusions. The review process should be designed with the consequences of errors in mind, rather than assuming the output is reliable simply because the article flows smoothly or the interface looks professional.

As the pace of scientific research accelerates, validation cannot be skipped.

The report focuses on how AI drives scientific discovery, but it also raises a related question: How should new discoveries proposed by machines be verified? This adds a necessary second part to the concept of “automated scientific research.” Generating hypotheses, designing experiments, and interpreting results are distinct stages, and bias can be introduced at any one of them.

Taking materials research as an example, one can envision AI helping to screen candidate formulations, thereby reducing the need for researchers to engage in trial-and-error. However, there remains a gap between the shortlist of candidates and materials capable of stable production—a gap involving experimental reproducibility, equipment requirements, cost, and long-term performance. This example is an analysis of the process and does not constitute a report announcing a specific materials breakthrough. The efficiency that truly warrants measurement should encompass the entire process, from proposal to final validation.

Employment issues also include how new employees can grow

The report discusses the automation of knowledge work, work quality, and the pressures faced by entry-level employees. One often-overlooked paradox is that many tasks that appear repetitive also serve as exercises for new hires to develop their judgment. If all foundational work is delegated to systems, organizations will need to redesign their training pathways; they cannot wait until senior employees leave to realize that their experience has not been passed on.

A more practical approach is to shift the focus of training from “how many drafts are completed” to “whether one can explain the rationale behind revisions.” New hires can use AI to produce a first draft, but they must identify gaps in evidence, recognize flawed assumptions, and explain why they accept or reject a particular suggestion. Only then can the time saved be transformed into a learning opportunity, rather than simply reducing the workforce.

Who gets to decide how AI is deployed?

The Max Planck Society's introduction to the report also highlighted issues such as a concentration of cutting-edge development, insufficient public participation, and fragmented governance.This InterpretationReaders should be aware that the impact of AI has extended beyond the technology sector.

For schools, companies, and public institutions, it’s helpful to start by addressing several specific questions: Which decisions must be approved by a human? How can affected individuals file a complaint? Is it possible to trace errors after they occur? And is it realistic to opt out of a particular system? The value of AI100 lies in bringing these day-to-day institutional issues to the forefront alongside technological progress. While uncertainty remains about what the next five years will bring, mechanisms for accountability, review, and participation can be incorporated into deployment plans right now.