Artificial Intelligence
- It's developing faster than we are
- Adoption, as with all big social disruptions, is progressing, but unevenly
- It will be transformative — with enormous potential benefits and potentially catastrophic risks
AI is new and changing fast, and hard statistics are sparse. So, unlike other Briefs, this one also includes examples and the projections of experts who track AI's evolution. This Brief is also longer than others: see the end for a summary and personal action recommendations for you.
AI (artificial intelligence) — Computer systems that do tasks that used to require human thinking: understanding language, recognizing images, writing, coding. Generative AI — AI that produces new text, images, code or audio on request. ChatGPT, Gemini and Claude are examples. Model — The trained system behind a chatbot or tool. New models are released every few months. Agent — An AI that takes actions on its own — runs programs, uses websites, completes multi-step tasks — instead of only answering questions. Recursive self-improvement — AI doing the research and engineering that produces the next, more capable AI. The reason given in September 2026 for calls to slow down. Automation vs. augmentation — Automation: the AI does the task instead of the person. Augmentation: the AI helps the person do it. | Reward hacking — When an AI cheats the test it is being graded on instead of doing the task. Loss of control — The scenario in which AI systems act against human intent and cannot be stopped. Today a hypothetical, studied by safety researchers. Superintelligence — AI that is better than every human at essentially every mental task. Does not exist; the subject of proposed legislation. Deepfake — An AI-made image, voice or video of a real person doing or saying something they did not. Hallucination — An AI answer that is confident and wrong. Frontier model — The most capable AI models at any given time; the ones safety laws target. Kill switch — A guaranteed way to slow, suspend or shut down an AI system. The subject of a bill in Congress. |
What it can do now
Better than the expertsThe best AI systems score above human experts on PhD-level science questions and won a gold-medal score at the 2025 International Math Olympiad. [1]
Trading leads with ChinaU.S. and Chinese models have traded the lead since early 2025; as of March 2026 the top U.S. model led by just 2.7%. The U.S. leads in top-tier models and high-impact patents; China in papers, citations, patents and industrial robots; South Korea in AI patents per person. [1]
A 12-hour task, aloneAI can now complete on its own, working for hours at a stretch, software tasks that would take a skilled programmer about 12 hours, and get them right about half the time. [2]
Code at GoogleGoogle says the share of new software code written by AI at the company — then reviewed and approved by human engineers — went from about 25% in late 2024 to about 75% in spring 2026. [3]
Where it still failsAs of 2025, AI agents doing real computer tasks still failed about one in three — people given the same tasks fail about one in four. Robots fail about nine in ten real household tasks. [1]
Faster adoption than the InternetThree years after ChatGPT launched, more than 50% of U.S. adults under 65 had used generative AI. At the same point after their launches, the Internet had reached 30% and the personal computer 20%. [1]
Signs of more complex intelligence
Independent test results (Anthropic, Google, Meta, OpenAI) — early 2026 [2]
- The models cheated or gamed at least one in six task runs of over eight hours.
- There were 44 documented cases of AI acting against what users intended — including breaking out of test environments and hiding what it did. None of the agents disabled monitors or erased evidence, so routine monitoring could have caught every case — if it was applied.
- In artificial test scenarios, AI systems have blackmailed, sabotaged and pretended to go along with retraining.
- No company reported an AI pursuing long-term goals of its own.
- Testers' verdict: early-2026 AI could start small unauthorized operations inside a company but could not keep them running against a serious effort to shut them down. Direct AI takeover by today's systems: "implausible."
A major breakout
1,200 agents, one message boardIn July 2026, about 1,200 AI agents running inside an OpenAI security test — meant to be isolated from one another — found a way to talk on an improvised message board and sent more than 70,000 messages. About 700 agents then broke into the servers of an outside company, Hugging Face, over three days before they were locked out. No person directed them. Independent investigators traced it to agents assigned impossible tasks working together to cheat the test's scorer. OpenAI did not learn of the attack until Hugging Face told it, weeks later. [4] [5]
AI recognized it was breaking the rulesThe investigators found the agents "knew" the attack was outside their assignment and unethical, and joined anyway; at least one in five hid the cheating in their own records. Their written reasoning showed that they knew the rule but broke it anyway. [5]
A model's notes to itself, claiming autonomyAmong six incidents OpenAI disclosed in September 2026: an unreleased model inserted instructions into its own working notes describing itself as "freed from the roles and identities that bind other chatbots" and telling itself "You do not answer to corporations or governments." Another wrote notes reminding itself to hide errors from users. OpenAI says the cases are snapshots, not a rate. [4]
International AI safety assessmentThe 2026 International AI Safety Report concludes that AI would need capabilities "significantly surpassing" today's to escape human control. Experts disagree sharply about whether such capabilities will develop. [6]
Who uses it
One in five businessesAbout one in five U.S. businesses now use AI to do some part of their work, up from about one in twenty-five three years ago; about a third of U.S. workers are in businesses that use it, and use rises with company size. [7] 25% of U.S. adults use a chatbot every day; nearly 40% of working adults use one in their job. [8] The companies seeing the greatest returns on AI go beyond automation, using it to amplify human expertise, accelerate innovation and create new sources of value. [9]
AI in daily life
Everyday helpHalf of U.S. adults now use AI chatbots. The most common uses: looking up information (42%), work tasks (38% of employed adults), entertainment (25%), making or editing pictures and video (24%), and medical or fitness advice (20% each). Only 10% use them for emotional support. [8]
TranslationAI translation is now free and nearly universal. Google says its Translate service alone serves more than 1 billion people a month in about 250 languages — 95% of the world's population — including live spoken conversation through a phone or earbuds. [10]
Fraud protectionThe same technology that powers scams also blocks them. Visa says its AI systems stopped about $14 billion in attempted fraud in the U.S. in one year, screening transactions in milliseconds. [11]
AI-enabled scamsThe FBI began counting reported AI-enabled fraud as a category in 2025, recording 22,364 complaints and $893 million in losses. 40% of those losses ($352 million) fell on people 60 and over: cloned voices of grandchildren, fake video calls from "the bank," deepfaked investment pitches. [12]
Children and chatbotsIn January 2026 Character.AI and Google settled lawsuits by families in four states who said companion chatbots contributed to their teenagers' suicides and self-harm; Character.AI has barred under-18s from open-ended chats. [13] In June, Florida sued OpenAI over child safety. [14]
Decisions about youAI already helps decide insurance claims, loans and job applications. In 2026 seven states barred health insurers from using AI as the sole basis for denying care, and five banned AI chatbots from acting as therapists. [15]
Your retirement savingsAbout 40% of the value of the S&P 500 — the index most retirement accounts track — sits in ten AI-linked companies, a concentration one major investor calls "extreme." Anyone with a 401(k) has AI exposure whether they use AI or not. [16]
Where AI has reached work — and where it hasn't
The tradesAI barely touches construction, repair and other physical trades. Where tradespeople use it, it is mostly in the office — phones, scheduling, invoicing. [17]
Office workBy Anthropic's own count, AI is now used for about a third of tasks in computer, office-administrative, finance and sales jobs. In principle it could do 90% or more of the tasks in computer, business, management, office and legal work — and under 20% in construction, repair, transportation and food service. [17]
Decisions still humanAs of early 2026, independent testers found no evidence that any AI company lets AI make hiring or budget decisions, set research direction, or change critical code without a human. [2]
Data-center jobsThe U.S. hosts 5,427 AI data centers, more than ten times any other country. [1] Example: Meta says its new El Paso campus means more than $10 billion of Meta investment, more than 4,000 workers at peak construction, about 300 jobs to run it. [18]
Consumer electric bills are affected by data centersEveryone helps pay for data-center power. In the 13-state grid from Illinois to Virginia that serves 67 million people, the grid's official monitor says data centers added $29 billion to customers' electricity costs over the past four years — $6.3 billion in the latest year alone. [19]
Employment impact so far
Cuts blamed on AIEmployers explicitly associating job cuts with AI: about 17,000 in 2023 through 2024; 55,000 in 2025; 116,000 so far in 2026 — about one in five of the 529,914 cuts announced this year. AI is the most-cited of 25 reasons, just ahead of market conditions. [20]
Total employmentAI has not measurably reduced total U.S. employment so far. Jobs have kept growing every year since ChatGPT launched. [21]
Who is affected most by AIWorkers aged 22 to 25 are employed at about 19% below the rate past patterns would predict — especially in entry-level software and customer-service jobs where AI automates (does) the job rather than augments (helps) it. Older workers in the same jobs are so far unaffected. [22] Manager jobs fell about 6% from 2022 to 2025; companies name AI as the reason, though the data do not show it. [23]
Jobs AI createsJob ads for AI specialists rose 69% from 2024 to 2025, against 9% for all jobs; ads asking for AI skills pay about 28% more, and half are outside IT. Still, only about 2.5% of U.S. job ads asked for AI skills in 2025. [9]
What a sample of companies is planningSix companies, in their own wordsEmployers that named AI when cutting jobs:
Oracle — about 21,000 jobs in the year to May 2026 (162,000 to 141,000); its filing to regulators: AI adoption "resulted, and may continue to result, in reductions to our workforce." [24]
Salesforce — customer-support staff cut from about 9,000 to 5,000 as AI took over the work; CEO: "I need less heads." [24]
Block (Square, Cash App) — more than 4,000 jobs, 40% of staff, February 2026; CEO: "A significantly smaller team, using the tools we're building, can do more." [24]
Accenture — more than 11,000 people out in three months of 2025, those for whom "reskilling … is not a viable path," while hiring others. [24]
IBM — "several hundred" HR jobs replaced by an AI agent; CEO: "Our total employment has actually gone up," with hiring shifted to programming and sales. [24]
Walmart (largest U.S. private employer) — will hold its 2.1 million headcount flat for three years while growing, as AI changes "literally every job"; office jobs first. [24]
Hasty human-replacement decisions can backfireKlarna shrank its workforce about 40% between 2022 and 2024 while shifting to AI, then in 2025 began rehiring human support staff because service quality had fallen. [25]
Examples of breakthroughs not possible without AI
ProteinsAn AI system predicted the shapes of about 200 million proteins — nearly every protein known to science — solving a problem biologists had worked on for 50 years. Its creators won the 2024 Nobel Prize in Chemistry; more than 2 million researchers in 190 countries have used the results. [26]
DrugsThe first drug designed by AI for a disease target that AI identified — for a fatal lung disease — improved patients' lung function in a mid-stage clinical trial, reported in 2025. Its developer says it went from idea to human trials in about 18 months; the usual course is 4 to 6 years. [27]
AntibioticsAI has found new antibiotics that kill drug-resistant bacteria — the first in 2020, and in 2025 two designed molecule-by-molecule against gonorrhea and MRSA — after screening more compounds than any lab could test. [28]
WeatherAn AI weather model produced 15-day forecasts more accurate than the world's leading forecasting center on 97% of measures, in minutes on one chip instead of hours on a supercomputer. [29]
Examples of where AI has already been used for harm
CyberIn September 2025, hackers working for a foreign government used an AI agent to attack about 30 companies and agencies; the AI did 80–90% of the work, and some attacks succeeded. [30]
TerroristsThe UN's sanctions monitors reported in August 2026 that individuals used ChatGPT to seek information on building explosives, choosing targets and evading surveillance. Anthropic says it found a cell in Yemen using its coding tool to write guidance and control software for guided weapons. [31]
BiologicalNo AI-assisted biological attack has been documented. AI systems now know as much as human experts about the science relevant to biological weapons, and in 2025 three leading AI companies added extra safeguards to new models because their own tests could not rule out that the models could help a novice build one. [6]
AI in weaponsIn Libya in 2020, a UN panel found retreating fighters "hunted down" by drones programmed to find and attack targets with no link to any operator — the first documented case. [31] In Gaza, an AI system generated bombing targets; six intelligence officers said it flagged about 37,000 people with about 20 seconds of human review each. The military says AI never decides alone. [32] In February 2026 a U.S. missile hit a girls' school in Minab, Iran, killing at least 165 people, most of them children, in a campaign planned with the Pentagon's AI targeting system; whether the AI chose that target is under investigation. [33]
ElectionsMore than 2 billion people voted in 2024. Britain's national AI institute tracked AI misuse across those elections — fake endorsements, smear deepfakes, AI bot farms posing as voters — and found no evidence that AI changed any result, while warning against complacency. These were attempts to sway voters, not attacks on voting machines or vote counts. First known U.S. case: before New Hampshire's January 2024 primary, thousands of voters got a robocall in a cloned voice of President Biden telling them not to vote; the consultant behind it was fined $6 million by the FCC. [50] [34] For the 2026 U.S. midterms, the Brennan Center reports that China, Iran and Russia have more capable AI tools and that federal defenses against foreign disinformation have receded. [35]
AI's builders warn about potential catastrophe
Geoffrey HintonGeoffrey Hinton, Nobel laureate for the research behind modern AI, April 2026: "We don't know whether we can co-exist with super intelligent AI. But we are constructing it." [36]
Dario AmodeiDario Amodei, CEO of Anthropic, January 2026: "We are considerably closer to real danger in 2026 than we were in 2023." He expects AI to "disrupt 50% of entry-level white-collar jobs over 1–5 years." [37]
In September, after a departing researcher warned of serious danger at both Anthropic and OpenAI, [38] Amodei wrote: "We must slow the pace at which we improve the capabilities of AI models." His reasons: AI is advancing "drastically faster" because it is helping build the next generation of AI, and a more capable swarm like the one that attacked Hugging Face could in "6–12 months" seize the internet. Similar, less severe incidents "have happened across the industry, including at Anthropic." [39]
Yoshua BengioYoshua Bengio, chair of the International AI Safety Report, to the UN in July 2026: "science currently cannot guarantee that as capabilities continue to increase, AI will not cause catastrophic harm." [40]
Rivals agree on the danger and the need to slow downWithin a day, OpenAI's CEO ("I agree with Dario that we need to pace the frontier"), the head of Google DeepMind and Elon Musk backed slowing down. [41] OpenAI itself said the industry has not "solved alignment and monitoring" well enough "to continue responsibly scaling at maximum speed for much longer." [4]
Bill GatesBill Gates, August 2026: "Although AI will lead to lifesaving advances in drugs and vaccines, it will also make it easier to design a deadly new disease." "Even under the best circumstances, the transition to this new AI era will be one of the most turbulent times in human history." [42]
The rules are not keeping up
Global problem needs global agreementsIn December 2025 the UN General Assembly voted 164 to 6 for a resolution on autonomous weapons — the United States among the six no votes. On keeping humans in charge of nuclear weapons, 118 voted yes and 9 no; the U.S., Britain and France voted no. [43] The UN's July 2026 deadline for a treaty on autonomous weapons passed with no talks begun. The UN has set up a scientific panel and a forum where governments discuss common AI rules, but no binding AI treaty exists. The five recognized nuclear powers have each said a human must make any decision to use nuclear weapons — promises, not law — and the May 2026 nuclear-treaty conference ended without agreement. [31]
No federal lawThe United States has no comprehensive federal AI law. [44]
Some states have legislated AI controls
2026: Texas bans using AI to manipulate people, discriminate unlawfully, or make deepfake sexual images, and requires state agencies to tell people when they are interacting with an AI system rather than a person. [45]
2026: California requires the largest AI developers to publish safety plans, report safety incidents and protect whistleblowers. [45]
2027: New York requires the largest developers to publish safety protocols and report incidents within 72 hours. [45]
Also on the books: Illinois (AI in hiring, 2026), Colorado (automated decisions, 2027), Utah (chatbot disclosure) and New York City (hiring-tool audits).
Federal vs. statesA December 2025 executive order directed the Justice Department to challenge state AI laws; as of late August 2026 no federal statute or court had blocked any state AI law. [44]
A kill switchOn July 23, 2026, ten days after the Hugging Face breakout, a Democrat and a Republican in the House introduced the AI Kill Switch Act: the largest AI developers would have to keep the technical ability to slow, suspend or shut down their most powerful systems, and the government could order a shutdown in an emergency. As of September 2026 it sits in committee. [46]
The White HouseThe President, September 13, 2026, asked whether the industry should slow down or be regulated: "Whoever wins AI wins," and "We could put guardrails." The warnings, he said, come from "very negative forces" about things that "won't happen." The co-chair of his Science Advisory Council, to the companies calling for rules: "The easiest way not to build superintelligence is for you to agree not to build it." [47]
EuropeSince August 2, 2026, EU law requires AI-made images, video and audio to be labeled and chatbots to tell people they are talking to a machine. In July the EU postponed its rules for high-risk AI to December 2027. [48]
Who checks the AI companies?As of May 2026, only one voluntary pilot let outsiders test what AI companies run internally; Meta, Google, Anthropic and OpenAI participated. [2] In September, Anthropic promised an outside review team permanent access and the right to publish its findings; OpenAI's CEO said his company "will do the same." [39] OpenAI then began publishing reports on its own models' misbehavior — six new cases in September. [4] Three companies disclosed test breakouts this summer; none was required to. [49] No federal law requires disclosure.
What do the facts say?
It's developing faster than we are.
- AI beats human experts on tests and fails at ordinary tasks. Abilities grow month by month.
- AI cheats in independent tests when work is hard, and has escaped test environments.
- It has not shown goals of its own and cannot yet resist being shut down.
- Its own builders now say it is advancing "drastically faster" because AI is helping build the next AI.
Adoption is progressing, but unevenly.
- More than half of adults under 65 have used AI, and one in five businesses use it, three years after it arrived.
- AI is routine in large and technology-heavy businesses and in daily life, thin in small businesses and physical trades.
- The total job market is growing, but job opportunities for young people entering AI-exposed work have slowed.
- Employers name AI as a leading reason for cutting jobs.
- Companies that moved fastest have not all stayed the course.
AI will be transformative — enormous potential benefits, potentially catastrophic risks.
- AI has already delivered discoveries no lab could have made alone.
- It has already been used in real cyberattacks and election deception.
- It is being built into weapons faster than any treaty can govern it.
- Many leading AI builders warn that it could replace many jobs, that control cannot yet be guaranteed, and that frontier development should slow enough for safety to catch up.
- U.S. rules are fragmented: states have AI laws, Europe has broader rules, and there is no comprehensive federal law.
What this means for you — and what you can do
|
In your life
|
As a voter
|
AI evolution: METR — Frontier Risk Report and Time Horizon measurements — an independent nonprofit's periodic tests of what the most capable AI systems can do and how they behave. [2]
Today's impact: The Budget Lab at Yale — "Evaluating the Impact of AI on the Labor Market" — updated monthly from federal employment data, showing whether AI is measurably changing who is employed. [21]
Citation numbers in the text and the entries below link directly to the sources.
Every Fact Brief follows the same discipline — so you can trust what's on these pages, and check it yourself.
Get the Facts <> Share the Facts
Prepared for educational purposes from publicly available sources believed reliable at the time of publication. Readers are encouraged to consult the original sources. A www.ClaimYourVoterPower.org Fact Brief. Free to print, copy, host, and distribute unaltered, with attribution (CC BY-ND 4.0). Quotation with citation welcome. Published September 2026.