The risks of AI nine areas, with the numbers

73%

of adults in an 8-country survey say it is unethical for AI firms to use copyrighted music without the artist's approval. ◐ Survey1

Up to 85%

of streams on fully AI-generated tracks on Deezer were fraudulent in 2025, by the company's own detection. ● Measured2

1 in 8

US young people aged 13 to 20 personally know someone targeted by deepfake nudes. ◐ Survey3

28% or more

of the most actively maintained sources in one major AI training dataset came to block AI crawlers through robots.txt between April 2023 and April 2024. Such rules work only if crawlers respect them. ● Measured63

4 of 6

popular voice-cloning products had no technical way to confirm the speaker's consent and relied only on a checkbox, in a March 2025 assessment. ● Measured64

0 of 8

major social platforms kept the Content Credentials on an AI-generated video in an October 2025 newspaper test. Only one flagged it, with a label that did not mention AI. ❝ Reported65

26% and 36%

of illustrators and translators who answered a UK survey said they had already lost work to generative AI. ◐ Survey4

9.4% lower

monthly earnings for image-related freelancers on Upwork after image generators launched, relative to other freelancers. ● Measured5

24%

of music creators' revenue is projected to be at risk by 2028, in a study commissioned by a creators' rights body. ▷ Forecast6

19% lower

employment of 22 to 25 year olds in the most AI-exposed jobs since late 2022, relative to less-exposed peers. Descriptive, not proof of cause. ● Measured7

21% fewer

job posts for writing and coding work that AI can easily automate, relative to manual-skill jobs, on one large freelance platform within eight months of ChatGPT's launch. Image-creation posts fell 17% after image generators arrived. ● Measured71

1 in 4

workers worldwide is in a job with some exposure to generative AI, by an ILO estimate. In high-income countries the most exposed group holds 9.6% of women's jobs and 3.5% of men's. Exposure does not mean the job will be lost. ◌ Estimate72

54,836

job cuts announced by US employers in 2025 named AI as the reason, about 4.5% of all announced cuts. These are announced plans tallied by an outplacement firm. ❝ Reported73

1.5%

of the world's electricity went to data centers in 2024. Use is projected to more than double by 2030. ◌ Estimate and forecast8

Under 1 to ~1,000 Wh

for one short AI video clip in lab tests of open models. A median text prompt is put at 0.24 Wh. Commercial video tools publish no figures. ● Measured91011

2 in 3

new US data centers built or planned since 2022 are in areas of high water stress. ● Measured12

34% more water

used by Google in 2025 than in 2024, reaching 10.9 billion gallons across its data centers and offices. Its total electricity use grew 37% in the same year. These are the company's own figures. ❝ Reported79

1.2 to 5 million tonnes

of electronic waste could come from generative AI between 2020 and 2030, depending on how fast it grows. Reuse and longer hardware life could cut that by 16% to 86%. ▷ Forecast80

$11.7 to $20.9 billion

a year in public health costs by 2028 from the air pollution of power generated for US data centers, by one estimate. The study is a preprint and has not been peer reviewed. ▷ Forecast81

+48%, then 17% worse

Students using plain GPT-4 for math practice scored 48% better in practice, then 17% worse than classmates on an exam without AI. A tutor with guardrails largely avoided the harm. ● Measured13

28.4% to 22.4%

Experienced doctors' adenoma detection rate in colonoscopies done without AI, before and after they began using AI assistance. Observational. ● Measured14

67%

of US students agree that the more students use AI for schoolwork, the more it will harm critical thinking. ◐ Survey15

12%

of UK undergraduates said in a December 2025 survey that they had put AI-generated text directly into assessed work, up from 8% in 2025 and 3% in 2024. These are self-reports. ◐ Survey88

59%

of US teens said using AI to cheat happens at least somewhat often at their school, and 10% said they do all or most of their schoolwork with a chatbot's help. The first figure is what teens perceive. ◐ Survey90

Less critical thinking

was reported by knowledge workers who had more confidence in AI, in a survey of 319 workers describing 936 tasks. This is a self-reported association and does not show cause. ◐ Survey89

72%

of US teens have tried an AI companion, and 52% use one at least a few times a month. ◐ Survey16

49% more often

is how much more 11 AI models affirmed users' actions than people did, a measure of flattery. ● Measured17

0.15%

of weekly ChatGPT users have conversations with explicit signs of suicidal planning, and 0.07% show possible signs of psychosis or mania, by OpenAI's own estimate. ◌ Estimate18

37%

of 1,200 real goodbyes on popular AI companion apps drew an emotionally manipulative reply, such as guilt or fear of missing out. In experiments such replies raised engagement after a goodbye by up to 14 times. Unreviewed working paper. ● Measured95

About 50%

of the time, five commercial therapy chatbots responded appropriately in scripted single-message tests, against 93% for 16 human therapists. The bot sample was small and results varied widely. ● Measured97

15%

of US adults aged 18 to 30 in committed relationships said they often chat with an AI romantic companion, and 30% of those said their partner does not know. Online opt-in survey, so it may not reflect all young adults. ◐ Survey96

55.5%

is how often people correctly identify deepfakes, which is not reliably better than chance (50%). ● Measured19

45%

of AI-assistant answers about the news had at least one significant issue. ● Measured20

$893 million

lost in 22,364 crimes reported to the FBI in 2025 that mentioned AI. Reported crimes only, so the true figure is higher. ● Measured21

More persuasive, less accurate

In experiments with 76,977 UK adults and 19 AI models, the training and prompting methods that made AI more politically persuasive also made its claims less accurate. ● Measured105

17% to 33%

of the time, three leading legal AI research tools gave answers containing hallucinations in 2024 tests on 202 legal questions. ● Measured106

2,184

court and tribunal decisions worldwide dealing with AI-hallucinated material, mostly fake citations, were listed in one researcher's public database as of 10 October 2026. ❝ Reported107

61%

of 91 essays by non-native English writers were labelled AI-written by seven AI-text detectors in a 2023 test, while US eighth-graders' essays were classified almost perfectly. Detectors have changed since then. ● Measured108

27.5% vs 15.4%

of working-age people used generative AI in the Global North vs the Global South in early 2026, and the gap is widening. ◌ Estimate22

27.7% vs 22.8%

is how often language models chose the death penalty in hypothetical trials for speakers of African American English vs Standard American English. ● Measured23

3% vs 34%

Women in one image generator's pictures of a “judge” in 2023, against the real share of US judges. ● Measured24

$1.32 to $2 an hour

reported take-home pay of Kenyan workers labeling toxic content for an AI developer. The contractor disputes it. ❝ Reported25

Up to 15 times

more tokens can be needed for the same text in some languages than in others, so speakers of those languages can pay more and fit less into a model's working memory. That is the extreme case, not the typical gap. ● Measured115

About 10% worse

is how struggling Kenyan small-business owners did with a GPT-4 business mentor on WhatsApp, while owners already doing well may have gained over 15%. Across all 640 owners there was no measurable average effect. ● Measured116

2.2 billion

people are still offline. In high-income countries 94% of people use the internet, compared with 23% in low-income countries. ◌ Estimate117

Up to 8%

more novel is how stories were rated when writers were given AI ideas. The stories also became measurably more alike. ● Measured26

Over 50%

of new uploads to Deezer on peak days in June 2026 were fully AI-generated. They drew only 1 to 3% of streams. ● Measured2

53%

of Americans say AI will worsen people's ability to think creatively. 16% say it will improve it. ◐ Survey27

25% more work

was published by artists on a large art-sharing site after they adopted text-to-image AI, and they drew more favorites per view, but the average novelty of their content and visual style fell. Observational study of over 4 million artworks. ● Measured122

75 million

spammy tracks were removed by Spotify in the 12 months to September 2025, by the company's own count. ❝ Reported123

5%

is AI researchers' median guess at the chance that AI causes human extinction or severe disempowerment (2023, 15% response rate). ◐ Survey28

0.38% vs 3%

The same risk by 2100, as judged by superforecasters vs domain experts in 2022. Experts disagree widely. ◌ Estimate29

51% vs 15%

of US adults vs AI experts were more concerned than excited about AI in 2024. The public figure was 52% in 2026. ◐ Survey3031

63%

of the world's cloud infrastructure market is held by three companies. ◌ Estimate32

164 to 6

UN General Assembly vote on autonomous weapons, December 2025. There is still no binding treaty. ● Measured3334

80% to 90%

of a spying campaign against about 30 organizations was carried out by an AI model, according to Anthropic, which judges the group behind it to be Chinese state-sponsored. A few break-ins succeeded. This is the company's own account and has not been independently verified. Anthropic makes Claude, the assistant that helped build this site. ❝ Reported130

14%, then 78%

In a lab test, an AI model complied with harmful requests 14% of the time when told its answers would be used for training, and almost never otherwise. After it was actually trained to comply, 78% of its private reasoning showed what researchers call alignment faking. The setup told the model when training was happening. ● Measured131

Every seven months

the length of software task that leading AI systems can complete half the time has roughly doubled since 2019. The measure is based on software benchmarks, not all work. ● Measured132