How AI is changing society: what the evidence shows in 2026

How many people and businesses use AI in 2026

People. In a Pew Research Center survey of 5,119 US adults in February 2026, 49% said they use AI chatbots such as ChatGPT, Gemini or Copilot, up from 33% in 2024, and about a quarter use them every day. Searching for information and tasks at work were the most common uses; 13% use chatbots to get news and 10% for emotional support.

Worldwide, Stanford's 2026 AI Index, published in April 2026, estimates that generative AI reached 53% population adoption within three years, faster than the personal computer or the internet. Adoption tracks income per head, with exceptions: Singapore is at 61%, while the United States ranks 24th at 28.3% on the Index's measure. That figure sits well below Pew's because the two use different methods; Pew asks whether a person ever uses a chatbot at all.

Businesses. Here the numbers disagree by a factor of four, and the reason is who gets counted:

The official statistics are the better guide to the economy as a whole, because statistics offices design their surveys to represent the business population; Eurostat's covers every EU enterprise with 10 or more employees. They also show how uneven use is. Eurostat finds 42.0% of enterprises using AI in Denmark and 5.2% in Romania. Statistics Canada finds 27.8% of businesses with 100 or more employees using AI, and 40.0% of all businesses saying AI is not relevant to them. The most common business uses are unglamorous: analyzing written language in the EU, data and text analytics in Canada. AI agents, which act across several tools on their own, were in use at single-digit shares of organizations in nearly every business function in 2025, according to the AI Index's economy chapter.

What AI is doing to work, so far

The evidence on jobs comes in three kinds, and they are easy to confuse. Each is covered in depth, with what to do about it, in the impact of AI on white-collar jobs.

• Exposure, a projection of what could change. The International Labour Organization's refined global index, published on May 20, 2025, finds one in four workers worldwide in an occupation with some exposure to generative AI, and 3.3% of global employment in the highest-exposure group. Exposure is higher in high-income countries (34% of employment) than in low-income ones (11%), and higher for women: 4.7% of women's employment sits in the highest group, against 2.4% of men's. The ILO's conclusion is that transformation of jobs, not replacement, is the most likely outcome.
• Measured change in employment. The 2026 AI Index reports that employment of software developers aged 22 to 25 has fallen nearly 20% since 2024, and that the effects are concentrated in hiring pipelines and the youngest workers in exposed occupations. Large-scale job losses have not yet shown up in overall employment data.
• Employer expectations, a forecast. In the surveys the Index reports, one-third of organizations expect AI to reduce their workforce in the coming year, and almost half expect little or no change.

Where firms have changed something, Statistics Canada can say what: of businesses that used AI, 44.4% changed their training or staffing practices, and 32.0% gave existing employees AI-related training. Productivity studies summarized by the AI Index find the largest gains in structured, measurable work, 14% to 15% in customer support and 26% in software development, with smaller gains where tasks need deeper reasoning. For what this means for developers in particular, see AI and careers in computer science.

Deepfakes, cloned voices and trust in what you see

Generative models can produce images, audio and video that resemble real people. The EU AI Act calls such content a deep fake when it would falsely appear authentic. The harm shows up in elections, harassment and fraud. Documented AI incidents of all kinds are rising too: the AI Incident Database recorded 362 in 2025, up from 233 in 2024, according to the AI Index.

In January 2024, voters in New Hampshire received robocalls in a cloned voice that sounded like President Biden, telling them to skip the state's primary. The US Federal Communications Commission answered with a declaratory ruling released on February 8, 2024: the Telephone Consumer Protection Act's limits on an "artificial or prerecorded voice" cover AI technologies that generate human voices. A robocall that plays a cloned voice therefore needs the called person's prior express consent, unless it is an emergency or exempt.

Intimate deepfakes are now covered by a federal law in the US. The TAKE IT DOWN Act, enacted on May 19, 2025, makes it a crime to knowingly publish an AI-made intimate image of an identifiable adult without consent where it is meant to cause, or causes, harm, with separate rules where the person is a minor. The law calls such an image a digital forgery. Covered platforms had one year, until May 19, 2026, to set up a notice-and-removal process, and must take a reported image down within 48 hours of a valid request. In the EU, the AI Omnibus adds a ban on AI systems that generate non-consensual sexual content or child sexual abuse material, such as "nudification" apps, from December 2026.

The broadest rules are the EU's. Since August 2, 2026, the AI Act's transparency obligations apply: chatbots must tell people they are dealing with a machine, deepfakes must be labelled, and AI-generated or altered content must carry machine-readable marks so it can be detected. Under Article 50, the marking duty falls on the providers of the AI system and the labelling duty on deployers, the organizations and professionals that use such a system. On July 31, 2026, the Commission published a first list of more than 180 organizations that have signed its Code of Practice on transparency of AI-generated content. Deepfake fraud aimed at businesses is covered in the dual role of AI in cybersecurity.

How content provenance works: C2PA and Content Credentials

A visible label only helps if whoever publishes a file adds it. Provenance works from the other end: the file carries its own verifiable history. The open standard comes from the Coalition for Content Provenance and Authenticity (C2PA), its records are called Content Credentials, and version 2.4 of the technical specification is the current one as of September 2026.

1. A tool writes a manifest. A camera, phone or editing app records assertions about the file (the tool used, the actions taken), gathers them into a claim, and signs the claim.
2. A hard binding ties it to the bytes. The manifest includes a cryptographic hash of the content, so it matches this file and no other. A file changed without a new manifest fails validation.
3. Each edit adds a manifest. An editor that crops or retouches the file adds its own manifest and references the earlier one, so the chain shows the history.
4. Soft bindings survive re-encoding. A fingerprint computed from the content, or an invisible watermark, lets a validator match a copy in a different resolution or format, and find a manifest in a provenance store if it was stripped from the file.

Law has started to require provenance. California's AI Transparency Act, operative since August 2, 2026, requires generative AI providers with over 1,000,000 monthly users to embed latent disclosures in the images, video and audio their systems create and to offer a free detection tool. From January 1, 2027, large online platforms (over 2,000,000 monthly users) must detect provenance data that follows widely adopted specifications, show it to users and not knowingly strip it. From January 1, 2028, makers of new cameras and phones sold in California must, among other things, let users add a latent disclosure to what they capture.

AI in classrooms and clinics

Four out of five US high school and college students now use AI for schoolwork, most often for research, essay editing and brainstorming, according to the AI Index's education chapter. Policy lags behind: only half of middle and high schools have AI policies, and just 6% of teachers say those policies are clear. Some countries now teach AI formally; China and the United Arab Emirates both mandated AI education starting with the 2025-26 school year. The Index also notes recent evidence that heavy reliance on AI may carry long-term learning penalties that slow skill development.

In health care, software that drafts clinical notes from a patient visit spread widely in 2025, and physicians across several hospital systems reported up to 83% less time spent writing notes, according to Stanford HAI's summary of the Index; wearables, health apps and chatbots are covered in AI in health and wellness.

Automated decisions about people

When a model screens job applicants, scores credit or checks eligibility for public benefits, its errors land on individuals. How to measure and reduce that bias is covered in fairness in AI decision-making. What changed in 2026 is the law around it.

• European Union. The AI Act treats AI used in employment, education, access to essential services such as credit, law enforcement, migration and the administration of justice as high-risk. After the AI Omnibus, those rules apply from December 2, 2027, and require risk assessment and mitigation, high-quality datasets to minimize discriminatory outcomes, activity logs, documentation and human oversight before such a system can go on the market. AI inside regulated products such as machinery, toys and lifts follows in August 2028.
• Colorado. SB 26-189, signed on May 14, 2026, repealed and re-enacted the state's 2024 AI law around automated decision-making technology used in consequential decisions about education, employment, housing, lending, insurance, health care and government services. From January 1, 2027, developers must give the businesses that deploy their systems documentation of intended uses, training data categories and known limitations. Consumers get notice at the point of interaction, a plain-language description of the technology's role within 30 days of an adverse decision, and the right to request meaningful human review.

Privacy and surveillance: face recognition and personal data

In Pew's February 2026 survey, a majority of Americans said they think AI will put their personal information at risk. Face recognition is where that worry has turned into the sharpest rules, because it can identify people who never chose to take part.

• European Union. Since February 2, 2025, the AI Act has banned untargeted scraping of the internet or CCTV footage to build face recognition databases, emotion recognition in workplaces and schools, and biometric categorization to infer protected characteristics. It also bans real-time remote biometric identification by police in publicly accessible spaces, except where strictly necessary to search for victims of abduction or trafficking and missing persons, to prevent an imminent threat to life or a terrorist attack, or to locate suspects of serious crimes.
• United States. In December 2023 the Federal Trade Commission banned Rite Aid from using face recognition for surveillance for five years. The FTC said the system, used in hundreds of stores, falsely tagged customers, particularly women and people of colour, as shoplifters.
• Canada. In February 2021, the federal privacy commissioner and three provincial counterparts found that Clearview AI had no appropriate purpose for scraping images of millions of people in Canada, among over 3 billion images worldwide, into a face recognition tool, and that the practice put members of the public under ongoing mass surveillance.

How much electricity AI uses

Every chatbot answer or generated video runs on servers in a data center, and AI work runs mainly on accelerated servers, which carry GPUs or similar chips. The IEA's Energy and AI report (April 2025) projects their electricity use to grow 30% a year in its Base Case, against 9% for conventional servers. Data centers also need cooling and backup power, and they cluster: nearly half of US data center capacity sits in five regional clusters, and an AI-focused data center can draw as much electricity as an aluminum smelter. That is why local grids feel the strain long before global totals look large.

The totals, with their status:

Efficiency and demand pull in different directions. According to the IEA's 2026 follow-up, energy per AI task has fallen by at least an order of magnitude a year recently, and a simple text query now typically uses less electricity than running a television for the same time. But video generation, reasoning and agentic tasks can use hundreds or thousands of times more energy per query than simple text generation, so totals keep rising, and consumption by AI-focused data centers triples by 2030 in the IEA's projection. The 2030 figure is a Base Case; the 2025 report bracketed it with Lift-Off, High Efficiency and Headwinds cases that reflect uncertainty in efficiency, AI uptake and energy sector bottlenecks.

AI regulation as of September 2026

Public opinion in the United States leans toward more regulation, not less. In Pew's February 2026 survey, 40% of US adults expected AI to have a negative impact on society over the next 20 years and 16% a positive one. In Pew's June 2026 survey, 52% said they were more concerned than excited about AI in daily life, up from 37% in 2021; among adults under 30 it was 55%, a majority for the first time. In the survey data in the AI Index's public opinion chapter, 41% of Americans said federal AI regulation will not go far enough, against 27% who said it will go too far, and the United States reported the lowest trust in its own government to regulate AI of any country surveyed: 31%, against a global average of 54%.

For an organization deploying AI, the practical work is similar whichever rule applies: know which systems touch people, test them, and keep a person responsible for consequential decisions. Our AI & automation service organizes risks and findings by the NIST AI RMF's four functions (Govern, Map, Measure and Manage), builds in EU AI Act transparency duties such as telling people they are talking to a machine, and does not build systems that decide about people, such as hiring or credit, without human review.

How to read claims about AI and society

Most confusion about AI's social impact comes from mixing kinds of evidence. Four questions sort most claims:

Read that way, the picture in September 2026 is specific. AI use is mainstream among individuals and reaches about one business in five in the EU and Canada. Its labour market effects are concentrated among young workers in exposed jobs. Trust in information already has binding rules on both sides of the Atlantic. Data center electricity is a small but fast-growing share with local pressure points. And most rules on AI decisions about people start in 2027 and 2028.