Radically different
If you’re following AI news, you’re probably getting whiplash. AI is a gold rush. AI is a bubble. AI is taking your job. AI can’t even read a clock. The 2026 AI Index from Stanford University’s Institute for Human-Centered Artificial Intelligence, AI’s annual report card, comes out today and cuts through some of that noise.
Despite predictions that AI development may hit a wall, the report says that the top models just keep getting better. People are adopting AI faster than they picked up the personal computer or the internet. AI companies are generating revenue faster than companies in any previous technology boom, but they’re also spending hundreds of billions of dollars on data centers and chips. The benchmarks designed to measure AI, the policies meant to govern it, and the job market are struggling to keep up. AI is sprinting, and the rest of us are trying to find our shoes.
All that speed comes at a cost. AI data centers around the world can now draw 29.6 gigawatts of power, enough to run the entire state of New York at peak demand. Annual water use from running OpenAI’s GPT-4o alone may exceed the drinking water needs of 12 million people. At the same time, the supply chain for chips is alarmingly fragile. The US hosts most of the world’s AI data centers, and one company in Taiwan, TSMC, fabricates almost every leading AI chip.
The data reveals a technology evolving faster than we can manage. Here’s a look at some of the key points from this year’s report.
In a long, heated race with immense geopolitical stakes, the US and China are almost neck and neck on AI model performance, according to Arena, a community-driven ranking platform that allows users to compare the outputs of large language models on identical prompts. In early 2023, OpenAI had a lead with ChatGPT, but this gap narrowed in 2024 as Google and Anthropic released their own models. In February 2025, R1, an AI model built by the Chinese lab DeepSeek, briefly matched the top US model, ChatGPT. As of March 2026, Anthropic leads, trailed closely by xAI, Google, and OpenAI. Chinese models like DeepSeek and Alibaba lag only modestly. With the best AI models separated in the rankings by razor-thin margins, they’re now competing on cost, reliability, and real-world usefulness.
The index notes that the US and China have different AI advantages. While the US has more powerful AI models, more capital, and an estimated 5,427 data centers (more than 10 times as many as any other country), China leads in AI research publications, patents, and robotics.
As competition intensifies, companies like OpenAI, Anthropic, and Google no longer disclose their training code, parameter counts, or data-set sizes. “We don’t know a lot of things about predicting model behaviors,” says Yolanda Gil, a computer scientist at the University of Southern California who coauthored the report. This lack of transparency makes it difficult for independent researchers to study how to make AI models safer, she says.
Despite predictions that development will plateau, AI models keep getting better and better. By some measures, they now meet or exceed the performance of human experts on tests that aim to measure PhD-level science, math, and language understanding. SWE-bench Verified, a software engineering benchmark for AI models, saw top scores jump from around 60% in 2024 to almost 100% in 2025. In 2025, an AI system produced a weather forecast on its own.
“I am stunned that this technology continues to improve, and it’s just not plateauing in any way,” says Gil.
However, AI still struggles in plenty of other areas. Because the models learn by processing enormous amounts of text and images rather than by experiencing the physical world, AI exhibits “jagged intelligence.” Robots are still in their early days and succeed in only 12% of household tasks. Self-driving cars are farther along: Waymos are now roaming across five US cities, and Baidu’s Apollo Go vehicles are shuttling riders around in China. AI is also expanding into professional domains like law and finance, but no model dominates the field yet.
These reports of progress should be taken with a grain of salt. The benchmarks designed to track AI progress are struggling to keep up as models quickly blow past their ceilings, the Stanford report says. Some are poorly constructed—a popular benchmark that tests a model’s math abilities has a 42% error rate. Others can be gamed: when models are trained on benchmark test data, for example, they can learn to score well without getting smarter.
AI companies are also sharing less about how their models are trained, and independent testing sometimes tells a different story from what they report. “A lot of companies are not releasing how their models do in certain benchmarks, particularly the responsible-AI benchmarks,” says Gil. “The absence of how your model is doing on a benchmark maybe says something.”
Within three years of going mainstream, AI is now used by more than half of people around the world, a rate of adoption faster than the personal computer or the internet. An estimated 88% of organizations now use AI, and four in five university students use it.
It’s early days for deployment, and AI’s impact on jobs is hard to measure. Still, some studies suggest AI is beginning to affect young workers in certain professions. According to a 2025 study by economists at Stanford, employment for software developers aged 22 to 25 has fallen nearly 20% since 2022. The decline might not be pinned on AI alone, as broader macroeconomic conditions could be to blame, but AI appears to be playing a part.
Employers say that hiring may continue to tighten. According to a 2025 survey conducted by McKinsey & Company, a third of organizations expect AI to shrink their workforce in the coming year, particularly in service and supply chain operations and software engineering. AI is boosting productivity by 14% in customer service and 26% in software development, according to research cited by the index, but such gains are not seen in tasks requiring more judgment. Overall, it’s still too early to understand the bigger economic impact of AI.
Around the world, people feel both optimistic and anxious about AI: 59% of people think that it will provide more benefits than drawbacks, while 52% say that it makes them nervous, according to an Ipsos survey cited in the index.
Notably, experts and the public see the future of AI very differently, according to a Pew survey. The biggest gap is around the future of work: While 73% of experts think that AI will have a positive impact on how people do their jobs, only 23% of the American public thinks so. Experts are also more optimistic than the public about AI’s impact on education and medical care, but they agree that AI will hurt elections and personal relationships.
Among all countries surveyed, Americans trust their government least to regulate AI appropriately, according to another Ipsos survey. More Americans worry federal AI regulation won’t go far enough than worry it will go too far.
Governments around the world are struggling to regulate AI, but there were some minor successes last year. The EU AI Act’s first prohibitions, which ban the use of AI in predictive policing and emotion recognition, took effect. Japan, South Korea, and Italy also passed national AI laws. Meanwhile, the US federal government moved toward deregulation, with President Trump issuing an executive order seeking to handcuff states from regulating AI.
Despite this federal action, state legislatures in the US passed a record 150 AI-related bills. California enacted landmark legislation, including SB 53, which mandates safety disclosures and whistleblower protections for developers of AI models. New York passed the RAISE Act, requiring AI companies to publish safety protocols and report critical safety incidents.
But for all the legislative activity, Gil says, regulation is running behind the technology because we don’t really understand how it works. “Governments are cautious to regulate AI because … we don’t understand many things very well,” she says. “We don’t have a good handle on those systems.”
Nature Medicine, Published online: 13 April 2026; doi:10.1038/s41591-026-04382-x
Author Correction: Real-time surveillance system for patient deterioration: a pragmatic cluster-randomized controlled trial
Background: An increasing number of rehabilitation technologies are being developed to support upper limb rehabilitation after stroke, with smart textile solutions for surface electromyography (sEMG) emerging as a promising approach. Early end-user involvement is crucial for developing user-friendly and clinically valid rehabilitation tools. Objective: This study aims to refine and evaluate the prototype design and usability of a smart textile biofeedback system for self-administered upper limb training after stroke. Methods: The training system includes a knitted smart textile sleeve with integrated electrodes over the forearm muscles, an sEMG unit, and tablet-based biofeedback software. An iterative co-design process was followed, including initial testing, demonstration sessions with end users (9 clinicians and 10 individuals with stroke), and a final evaluation of the co-design process. Participants’ experiences were gathered through semistructured interviews, analyzed using content analysis, and the User Experience Questionnaire. The co-design team included experts in stroke rehabilitation, textile engineering, biomedical engineering, software development, and human factors, as well as a research partner with lived experience after stroke. Results: The perspectives of the end users and the expert team were collectively integrated into prototype refinements of the sleeve and training software to meet the needs of the intended target group. The experiences of end users formed 2 main categories: “This could be an exciting new training tool for stroke rehabilitation” and “The tool works well, but some changes could enhance independent training.” End users found the smart textile sleeve and biofeedback system easy to use and saw potential for integrating it into their training routines. Both end-user groups rated the system as attractive, stimulating, and novel. Conclusions: The results of this study establish a necessary ground toward the development of a smart textile sEMG biofeedback system for self-administered upper limb training after stroke. Findings from the co-design process support the continued development and evaluation of the system as a self-administered upper limb training tool for individuals living with stroke.
<img src="https://jmir-production.s3.us-east-2.amazonaws.com/thumbs/09868209cfa3cec73d3f73d20963ee99" />
Background: Stairclimbing wheelchairs offer enhanced mobility for users navigating multilevel environments, yet limited research addresses the ergonomics of lever propulsion-based stair climbing mechanisms. Comprehensive ergonomic assessment integrating both subjective user feedback and objective biomechanical analysis is essential for optimizing assistive device design for comfort and usability. Objective: This pilot study aims to assess the ergonomic design of a transformable stair-climbing wheelchair through a dual-methodology approach, evaluating plane surface movement accessibility and quantifying muscle activation patterns during lever-propelled stair-climbing operations using surface electromyography (sEMG). Methods: This 2-part study involved anthropometric measurements from 20 male participants to establish design parameters using 5th and 95th percentile values. Part A assessed plane surface movement with 9 participants (7 healthy, 2 with paraplegia) navigating a simulated urban course featuring a 5° ramp, a 90° turn, and narrow passages across 3 trials. Task completion times and subjective ride easiness ratings were recorded. Part B used a Taguchi-based fractional factorial design to evaluate 3 ergonomic factors, including torso angle (λ), lever distance (L), and lever orientation (ψ), across 7 healthy participants. Maximum voluntary contraction (MVC) was measured for 4 muscles, including biceps brachii long head (BBL), triceps brachii long head (TBL), brachioradialis, and posterior deltoid (PDT). Results: In Part A, the ramp and 90° turn proved most challenging due to the wheelchair’s 65 kg weight and large turning radius (~1450 mm). Driving control scored highest (6/10), while comfort scored lowest due to the tilted seat design. In Part B, a straight torso (λ=0°) consistently reduced muscle strain, particularly for brachioradialis. A lever distance of approximately 50 mm and a neutral to slightly supinated orientation (ψ=0°-30°) optimized muscle effort. Interaction effects revealed high strain configurations (λ=45°; L=100 mm; ψ=−30°) exceeding 75 MVC, while optimal settings reduced strain to approximately 50 MVC. Conclusions: Optimal ergonomic parameters of λ=0°, L=37.5 mm, and ψ=15° are recommended to minimize fatigue and enhance user comfort. Design improvements should prioritize weight reduction, compact form factor for maneuverability, and adjustable seat tilt. The modular wheelchair design permits customization for diverse user populations. Future research should include larger, gender-diverse participant groups and real-world validation studies.
<img src="https://jmir-production.s3.us-east-2.amazonaws.com/thumbs/c6e4d2f2e2e8a4f16fcea3fc762bfa0d" />
Background: Wellness tourism increasingly incorporates short yoga sessions for stress management, yet evidence on their immediate physiological and emotional responses remains limited. Objective: This exploratory pilot study aimed to generate preliminary data on acute physiological and emotional responses to a single 24-minute yoga session (breathing, postures, and meditation) compared with quiet sitting in novice practitioners. Methods: In this single-center, randomized waitlist-controlled exploratory pilot study (1:1 allocation), 19 university-affiliated adults with self-reported daily stress and no prior guided yoga experience were allocated to an immediate 24-minute yoga session (n=10) or quiet sitting (n=9). The primary outcome was change in salivary cortisol from baseline to postintervention. Secondary outcomes included salivary alpha-amylase, heart rate (Apple Watch [Apple Inc], subset n=12), and real-time emotional indices (stress, calmness, and concentration) measured by a commercial single-channel electroencephalogram (EEG) device (Kansei Analyzer [Dentsu ScienceJAM], subset n=8). Data were analyzed using change scores relative to the breathing phase, with between-group comparison performed by independent tests or multivariable linear regression adjusted for age, sex, and baseline values as appropriate. Results: No significant between-group differences were observed for salivary cortisol or alpha-amylase. In the yoga group, heart rate increased during postures and remained mildly elevated during meditation relative to the breathing phase, consistent with mild physical exertion. Exploratory individual EEG analyses suggested reduced stress and increased concentration during breathing and meditation phases in some participants. Conclusions: In this small exploratory pilot study, a single 24-minute yoga session did not significantly alter salivary stress biomarkers at the group level. However, it induced measurable cardiovascular activation and heterogeneous EEG-derived emotional responses. These hypothesis-generating findings demonstrate marked interindividual variability, suggesting the need for larger validation studies with improved control conditions.
<img src="https://jmir-production.s3.us-east-2.amazonaws.com/thumbs/4888ec268eb7bb97167592e681af43d2" />