Available evidence puts English at about 92.65% of GPT-3's training data, 89.7% of Llama 2's and nearly 90% of Claude 2's. The models now shaping global culture were built almost entirely on one language's words.
Nayeem & Rafiei, 2026
Corroborated660 million people use one AI companion app, XiaoIce, alongside Chai (4 million active users) and Replika (2.5 million): software built in one place, woven into daily emotional life everywhere.
De Freitas et al., 2024
Only 20% of the world's population speaks English at home, yet nearly half of the training data for major AI models is in English.
UN News, 2025
73% of AI experts expect the technology to improve how people do their jobs; 23% of the US public agree. A 50-point gap over the same technology.
Stanford HAI, AI Index, 2026
Sketches, translations, creative tools and softer arguments: what AI adds when cultures meet.
Researchers analysed 2.6 billion crowdsourced sketches of everyday concepts from 236 countries and territories. Cross-cultural similarities drawn from the sketches align 32% more closely with known cultural distances than text-based measures do.
Pera et al., 2026
Single-sourceGoogle says AI techniques let it add over 110 new languages to Google Translate in a year, spoken by half a billion people, for a total of 249, including 60 African languages.
UN News, 2025
Company-reportedA study of over 4 million artworks from more than 50,000 users found text-to-image AI raised creative productivity by 25% and the chance of a work being favourited by 50%.
Cited study, via AI Companionship review (arXiv), 2025
Single-source1,500 people were paired to debate gun control online. They took up 2,742 of the AI chatbot's suggestions to restate or soften their posts, about two-thirds, and felt more heard and understood.
BYU, on Argyle et al. (PNAS), 2025
Peer-reviewedA survey of Indonesia's language resources covers Bahasa Indonesia and 726 regional languages, most of them small, and lists the lexical resources being built to keep them in use.
Riza (ACL Anthology), 2008
Peer-reviewedA 2025 study reported that 91 percent of educators observe enhanced learning when their students use creative AI.
Cited study, via AI Companionship review (arXiv), 2025
Single-sourceAbout 29% of Gen Z users already use generative AI in their daily communication, a share the authors expect to keep growing.
Tsuria & Tsuria, Religions (MDPI), 2025
Peer-reviewed58% of employees worldwide used AI at work regularly or semi-regularly in 2025; in India, China, Nigeria, the UAE and Saudi Arabia the share passed 80%.
Stanford HAI, AI Index, 2026
CorroboratedNo verified figure exists for how much cultural exchange happens off the platforms: informal sharing, diaspora networks, in-person teaching. Every estimate we found extrapolated from platform data, so none is given.
We could not establish whether exposure comes before tolerance or after it. The correlation is documented; no dataset we found settles the direction, so no causal figure is claimed.
Bias that varies by model, words that leak from chatbots into speech, and opinion that moves both ways at once.
On the same cross-lingual religious-stereotype probe, GPT-4o-Mini picked the anti-stereotype answer in about 92% of cases; Gemini-2.5-Flash managed about 68%, with more invalid replies.
Shankar et al., 2026
Single-sourceWords ChatGPT favours, such as “delve” and “meticulous”, are turning up more often in spoken language, according to an analysis of more than 700,000 hours of videos and podcasts.
Scientific American, 2025
Single-sourceOptimism is rising alongside nervousness: more respondents now say AI's benefits outweigh the drawbacks, and South-East Asia leads on expecting change.
Stanford HAI, AI Index, 2026
CorroboratedAround half of Replika's users describe their relationship with their AI companion as romantic.
De Freitas & Keller 2022, cited in De Freitas et al., 2024
Single-sourceFrance's public LLM arena, compar:IA, has collected over 600,000 prompts, 89% of them in French: offered a choice, people use AI in their own language.
Termignon et al., 2026
Single-sourceNo reliable count exists of AI-generated cultural works entering circulation. Providers that publish volumes use incompatible definitions of a work.
No verified measure exists of how much local adaptation survives inside a licensed international format. The industry treats it as commercially sensitive, and we found no independent audit.
Shared models, vanishing languages, convincing fakes and companions standing in for people.
BERT was downloaded 24 million times in one month, and GPT-3 powers hundreds of deployed applications: the shared infrastructure researchers warn can make outcomes uniform across unrelated decision-makers.
Bommasani et al. (NeurIPS), 2022
Peer-reviewedFewer than 300 first-language speakers of Mvskoke remain, and nearly all are over 60: a generational cliff no new tool outruns on its own.
Mvskoke roadmap (ACL ComputEL), 2024
Peer-reviewedParticipants identified text-to-speech deepfakes of political speeches in only 72% of observations, closer to the 50% of random guessing than to a perfect score.
Groh et al., Nature Communications, 2024
Peer-reviewedAI and robotics are moving into sewing, cutting and fabric inspection; estimates suggest as many as 60 percent of jobs in Bangladesh's garment sector could be lost to automation by 2030.
Center for Global Development, 2025
Single-source72% of US teens have tried an AI companion chatbot and 52% use one regularly: a substitute for everyday contact that arrived within a couple of years.
Robb & Mann 2025, cited in Chu et al., 2026
Single-sourcePeople whose jobs were one standard deviation more exposed to AI were 45 percent less likely to believe in God; countries with many robots saw religiosity fall 3 percent a decade.
Jackson & Waytz, via Chicago Booth Review, 2025
Single-sourceMore than 88% of the world's languages, spoken by about 1.2 billion people, have been overlooked by language technology: the tools mediating culture serve a small share of its tongues.
Joshi et al. 2020, cited in NLP progress review (ACL), 2024
Corroborated700 companies, including over 30% of the Fortune 100, rely on a single vendor for algorithmic resume screening: the same concentration that can make outcomes uniform.
Bommasani et al. (NeurIPS), 2022
Peer-reviewedAs of 2017, algorithmic smart replies made up 12% of all messages sent through Gmail, nudging everyday writing towards the phrasing the system found safest.
Hohenstein et al., Scientific Reports, 2025
Peer-reviewedWith five generative-AI ideas to hand, writers' stories became more alike: the rise in similarity equalled 10.7% of the total range. Individual stories improved; the pool narrowed.
Doshi & Hauser, Science Advances, 2026
Peer-reviewedOf Indonesia's regional languages, 386 are spoken by 5,000 people or fewer, 233 by 1,000 or fewer and 52 by 100 or fewer.
Gordon 2005, cited in Riza (ACL Anthology), 2008
Peer-reviewedUncertain baselines, concentrated speakers and the questions research has barely started to ask.
AI companions are sold as a cure for loneliness, yet the size of the problem is itself uncertain: estimates of chronic loneliness in the US run from 30% to 60%, depending on the survey.
Holt-Lunstad et al.; Ipsos, cited in De Freitas et al., 2024
Single-sourceThirteen of Indonesia's regional languages have a million or more speakers and account for 69.91% of the population; the other 713 share just 41.4 million speakers.
Lauder 2004, cited in Riza (ACL Anthology), 2008
Peer-reviewedA systematic review of socio-cultural AI research found only ten percent of the abstracts analysed touch policy at all: governance is the field's least-studied corner.
Feher & Katona, Futures, 2021
Peer-reviewedAcross 25 countries a median 53% trust the EU to regulate AI effectively, against 37% for the United States and 27% for China.
Pew 2025, via Stanford HAI, AI Index, 2026
CorroboratedAsked for Black African doctors treating white children, the image model always drew the children as Black, and in 22 of over 350 images drew the doctors as white.
NPR, on Alenichev & Kingori, 2025
Single-sourceThe exchange is real and so is the concentration. Both sides of this page are sourced to the same standard: every figure carries its origin and a confidence label, and where no verified number exists the page says so. What the numbers cannot tell you is which losses were chosen and which were simply allowed.
Several figures are cited second-hand, as their source states them: Joshi et al. via an ACL review, Robb & Mann via Chu et al., De Freitas & Keller and Holt-Lunstad et al. via De Freitas et al., Gordon and Lauder via Riza, Pew via the AI Index, and two studies via an arXiv review of AI companionship.
