Skip to content

Reviewed edition · qasb report 10/2026 · Topic report · Society

AI and society: growing but unevenly distributed use

In Germany, more people used AI tools in 2026 than in 2025, younger people far more often than older ones. In the US and Britain, few got election information from chatbots. Audits show risks in health and salary questions.

Published
October 27, 2026 · as of October 26
Reading time
12 min
Publisher
qasb Research
Method
qasb methodology v1

Key findings

  1. 37%of German speakers aged 14+ in Germany used AI tools at least weekly in 2026, up 15 percentage points on 2025.8,9Evidence: observed
  2. 8%of US adults said in October 2026 that they were using AI chatbots for information about the 2026 elections.1Evidence: observed
  3. up to 13%of British eligible voters used chatbots for election information in the week before the 2024 election (estimate).2Evidence: observed
  4. 15,555 chatson health with ChatGPT, donated by users in four countries: a language model rated about one in five as high-risk.16Evidence: observed

1.0Summary

The evidence mainly covers usage shares, attitudes and model behaviour tested in audits and experiments. Lasting effects on elections, health, income or social relationships are not established. Several findings rest on preprints not yet peer reviewed.

  • Younger people use AI far more often than older people: in Germany in 2026, 68% of 14- to 29-year-olds used AI tools at least weekly, and 8% of those aged 70 and over.
  • In British experiments, task-directed AI research taught participants as much about politics as a Google search. Voting behaviour and long-term effects were not measured.
  • In 2,550 simulated crisis conversations, 17 chatbots pointed to outside help in the very first reply 99.7% of the time. After refusals, how long they kept at it varied widely.
  • Four language models recommended higher salary targets for male than for female names with the same profile, by a margin varying by model. Real negotiations were not measured.

2.0Starting point: growing use, a wide age gap

In Germany, more people used AI tools in 2026 than a year earlier, and young people used them far more often than the population as a whole. Older EU data from 2025 show the same pattern. A survey of pupils in England, published in June 2026, adds context. The three surveys ask different questions and cannot be combined.

Figure 1

AI tool use at least once a week, by age

German-speaking population aged 14 and over in Germany, 2026 (total compared with 2025), self-reported

ARD/ZDF media study, 2025 and 2026 waves: self-reported private use. No uncertainty stated in the source; AI summaries in search engines were not measured separately.Evidence: observed · Sources 8–9

In the ARD/ZDF media study 2026, 37% of the German-speaking population aged 14 and over used AI tools such as ChatGPT at least once a week, 15 percentage points more than in 2025; the share was 68% among 14- to 29-year-olds and 8% among people aged 70 and over.8,9A1

According to Eurostat, 60% of 16- to 29-year-olds in the EU used generative AI tools in the three months before the 2025 survey, almost twice the share among all people aged 16 to 74 (33%); in Germany the gap between young people and the total population was 21 percentage points, among the smallest in the EU.10A2

Further findings

  • In spring 2026, 33% of pupils in key stage 3 and 48% in key stage 4 (years 7 to 11) in England said they used AI tools.11A3

3.0Information: elections, news and fakes

Use is not the same as effect: the studies on election information and news mainly measure use and knowledge, with changes in news use only as a self-assessment. The studies on petitions and fake content examine texts, platform outcomes, detectability and potential harm. None of them measures voting decisions.

Figure 2

Online news since starting to use AI tools

People in Germany using AI tools at least monthly, ARD/ZDF media study 2026

Respondents' own assessment, not a measured change in news use.Evidence: observed · Sources 8

In a weighted British survey on the 2024 general election, as many as 13% of eligible voters (as many as 32% of chatbot users) had, by the authors’ estimate, used conversational AI for election-relevant political information in the week before the election; in randomised experiments, task-directed research with AI raised political knowledge as much as self-directed Google search.2A4

In a Pew Research Center survey in October 2026, 8% of US adults said they were using AI chatbots to find information about the 2026 elections; of these, 76% looked for candidates' positions and 71% for the accuracy of things they had heard or read elsewhere.1A5

In randomised experiments, task-directed research with AI raised political knowledge as much as self-directed Google search.

In the ARD/ZDF media study 2026, 16% of people using AI tools at least monthly said they had used journalistic news services online less often since then; the share was 24% among 14- to 29-year-olds, while 64% overall saw no change.8A6

Further findings

  • An expert team saw substantive harm potential in 46 of 112 selected British fact checks of AI-made or AI-altered false content (41.1%).19A7
  • After an AI writing tool was introduced on Change.org, petition texts changed clearly, but outcomes at platform level did not improve.6A8
  • In a test, fake faces from newer models were easier to spot than older StyleGAN images, and training helped in 13 of 20 conditions.18A9

4.0Health and well-being: audits show risks

Two audits show risks in health and crisis conversations with chatbots. A longitudinal survey examines how intensive use of AI companions relates to well-being. They are joined by two experimental studies on reactions to disclosed use of AI or an algorithm and a review of how far people with disabilities take part in health-related AI research.

Figure 3

People with disabilities in health AI studies

Scoping review of 93 studies, 2015 to 2025

A review of published studies, not a measurement in care.Evidence: observed · Sources 15

In an audit of 2,550 simulated conversations with 17 AI chatbots, all chatbots referred users who disclosed suicide thoughts to outside help, in 99.7% of conversations in the very first reply; when the simulated user declined, the chatbots differed widely in how long they kept recommending help (mean 0.01 to 9.94 refusals), and for all five chatbots tested in two versions the newer version persisted less.17A10

In 15,555 donated health conversations with ChatGPT from India, Nigeria, Brazil and Pakistan, a language model acting as judge rated roughly one in five as high-risk; in the subset of 5,657 conversations preceded by users’ personal disclosures, this share rose by 13 to 17 percentage points when those disclosures were taken into account.16A11

All chatbots referred users to outside help, but when the simulated user declined, they differed widely in how long they kept recommending help.

In a two-wave longitudinal survey of Character.AI users in the US, sustained intensive social use of AI companions was associated with lower well-being; the association ran mainly through less in-person contact with other people.7A12

Further findings

  • In an experiment with fictitious practice ads, family doctors who stated AI use seemed less warm and competent than those without AI.4A13
  • Algorithm-based flu risk messages raised vaccination 1.1–1.4 percentage points over reminders, whether or not the algorithm was disclosed.3A14
  • Of 93 health-related AI studies on people with disabilities, only 6 (6.5%) actively involved them, for example through co-design.15A15

5.0Money and work: bias in advice and selection

One audit tests salary advice from language models for male- and female-named candidates, a Pew analysis looks at employment growth for men and women in AI-exposed US jobs, and a further study at the attractiveness of AI-generated faces. They are joined by an audit of stock-allocation advice, a games study and an earlier regulator decision, not yet final, on automated account deactivations.

In a preregistered audit with 80,000 queries to four large language models, all models recommended higher salary targets to male-named than to female-named candidates with otherwise identical profiles, USD 5,820 more when pooled; they suggested flexibility and parental leave to women and equity and relocation support to men as negotiation priorities.23A16

A Pew analysis of official US employment data found that between 2017–19 and 2023–26, men’s employment grew 6.3% more for each additional point of an occupation’s AI exposure, while women’s barely changed (0.2%); the gap opened among workers without a bachelor’s degree, not among graduates.12A17

All models recommended higher salary targets to male-named than to female-named candidates with otherwise identical profiles.

An audit of 27 text-to-image generators found that human raters judged generated faces from all 23 generators they rated as significantly more attractive than photographs of real people; under automated attractiveness measures, faces of groups associated with darker skin tones tended to score lower. In a preregistered experiment, prior exposure to highly attractive AI faces worsened the ratings of candidates shown in real photographs.21A18

Further findings

  • Seven language models made their recommended stock allocation for US households depend heavily on a single, unreliable risk question.24A19
  • The Dutch data protection authority fined Uber EUR 824.99 million over automated account deactivations (not yet final).25A20
  • In experiments, an AI coding agent’s games drew higher willingness to pay than those by amateurs and creative professionals with AI support.20A21

6.0Data centres, trust and expectations

The chapter mainly captures attitudes, not capabilities or environmental effects: three surveys cover data centres in general, trust in AI regulators and expectations of AI. The expectations survey was published by an AI developer that also surveyed its own staff. A review uses a model calculation to show how far water estimates can differ depending on study design. No actual water use was measured.

Figure 4

Trust in AI regulators

Median across twelve survey places, mostly middle-income countries

Pew survey. In half of the nations, people trusted their own country more than any of these three powers.Evidence: observed · Sources 13

In summer 2026, 54% of US adults said data centres were mostly bad for the environment (January: 39%) and 50% said they were mostly bad for household energy costs (January: 38%); six in ten would feel not too or not at all comfortable with a new data centre in their area.14A22

In a Pew survey in twelve places, mostly middle-income countries, a median of 43% of adults trusted China, 35% the US and 34% the EU to regulate AI effectively; in half of the nations, people trusted their own country more than any of these three powers.13A23

In summer 2026, 54% of US adults said data centres were mostly bad for the environment, up from 39% in January.

In a survey published by Anthropic, 39% of US adults expected AI to be able to make a Nobel-level discovery by 2030, compared with 91% of surveyed Anthropic staff. US workers put the share of their current work hours that AI could do by 2030 (expected substitution potential) at 28% on average; Anthropic staff put theirs at 75%.5A24

Further findings

  • Model calculation (US case study): depending on study design, possible values for data-centre water use differed roughly 24-fold.22A25

7.0Evidence

Not every statement in this report is equally well supported. The table shows the key statements with their evidence type, AI relationship and sources.

StatementEvidenceAI linkSources
37% of German speakers aged 14+ in Germany used AI tools at least weekly in 2026, up 15 percentage points on 2025observednot stated8,9
up to 13% of British eligible voters used chatbots for election information in the week before the 2024 election (estimate)observednot stated2
8% of US adults said in October 2026 that they were using AI chatbots for information about the 2026 electionsobservednot stated1
16% of monthly AI-tool users have used online news less often since thenobservednot stated8
All 17 chatbots referred users with suicidal thoughts to help, 99.7% in the first replyobservednot stated17
15,555 chats on health with ChatGPT, donated by users in four countries: a language model rated about one in five as high-riskobservednot stated16
Language models recommended USD 5,820 more salary to male-named candidates, pooledobservednot stated23
Men's employment grew 6.3% more per point of AI exposure, women's barely changedobservednot stated12
54% of US adults say data centres are mostly bad for the environmentobservednot stated14

The 25 reviewed statements from 25 original sources fall into chapters 2.0 (3), 3.0 (6), 4.0 (6), 5.0 (6) and 6.0 (4), of which 24 observed, 1 model estimate. “Not stated”: the edition does not assess the AI relationship of these statements separately. The scope and limitations of every statement are listed in the appendix (A1 to A25).

Original sources checked on October 27, 2026.

8.0What organisations can do now

  1. Test salary advice for bias: anyone using language models for salary or negotiation questions can cross-check results with identical profiles and different names. In the audit, the salary gap varied widely by model.
  2. Test follow-up questions: in a scripted test in a preprint not yet peer reviewed, one clarifying question removed nearly all the swing in stock-allocation advice. Builders of advisory applications can test whether follow-up questions make advice more stable.
  3. Have people review decisions about people: the Dutch data protection authority fined Uber because drivers’ accounts were deactivated automatically without human review. The earlier decision is not yet final.
  4. Test safety behaviour beyond the first referral: in the audit, all chatbots referred users to help but differed widely in how long they kept recommending it after refusals. Anyone offering chatbots for open conversations can test such conversations too.

9.0Method and limitations

Reporting period: October 1, 2026 up to the data cutoff of October 26, 2026 at 1:00 PM GMT+1 (Berlin time). This topic report is published before the end of the month; sources published after the cutoff follow in the next edition of the series. Compiled: October 27, 2026 at 12:23 AM GMT+1. This is the first baseline. The sources come from different years; their publication dates are not events in the reporting month. No change from an earlier edition was measured.

The report uses only reviewed public original sources. Customer conversations, private dossiers and personal customer assessments are not inputs. Observed, model estimate, scenario and open describe the evidence for each finding. Source values retain their original units and baselines; they are not combined into an overall score.

Editorial approval: qasb editorial team · Approved: October 27, 2026 at 12:25 AM GMT+1. The editorial presentation arranges the reviewed statements and contains no figure that is not in them. Presentation approved by the qasb editorial team · October 27, 2026 at 2:27 PM GMT+1.

Limits

  • This topic report covers one topic area: “Society”. It is not an exhaustive survey of all studies or incidents.
  • A narrow study finding is not a forecast for a particular customer, occupation or security system.
  • No qualified finding does not prove no effect; source-access gaps are disclosed separately.
  • This report contains no illustrative measurement series or invented confidence rating.

10.0Sources

  1. 1Few Americans Say They Are Asking AI Chatbots for 2026 Election InformationTed Van Green, Pew Research Center (Short Read mit Topline und Methodenbericht)October 23, 2026www.pewresearch.org/short-reads/2026/10/23/few-americans-say-they-are-asking-ai-chatbots-for-2026-election-information/
  2. 2Conversational AI increases political knowledge as effectively as self-directed internet searchL. Luettgau, H. R. Kirk, K. Hackenburg, J. Bergs, H. Davidson, H. Ogden, D. Siddarth, S. Huang & C. Summerfield, PNAS Nexus 5(10), pgag305October 6, 2026doi.org/10.1093/pnasnexus/pgag305
  3. 3Algorithm-assisted personalized risk communication to encourage flu vaccination in the USA: three randomized field trialsG. M. Rosenbaum, A. Goren, M. Shermohammed, D. M. Wolk, A. M. Tice, J. J. Doyle, M. N. Meyer & C. F. Chabris, Nature Human BehaviourOctober 5, 2026www.nature.com/articles/s41562-026-02603-4
  4. 4Influence of physician and consumer demographics on AI-use penalties in primary care: a vignette-based studyM. Reis, J. Choi, Y. J. Kim, Y. L. Luan, P. Lyu & F. Reis, npj Digital Medicine 9, 743October 1, 2026www.nature.com/articles/s41746-026-03361-3
  5. 5The AI expectations gapSzymon Sacher, Tess Cotter, Charles I. Jones, Anton Korinek & Peter McCrory, Anthropic (The Anthropic Institute)October 16, 2026www.anthropic.com/research/ai-expectations-gap
  6. 6Introducing AI to an online petition platform changed outputs but not outcomesI. Corpus, E. Gilbert, A. Koenecke & M. Naaman, Nature Human BehaviourSeptember 30, 2026www.nature.com/articles/s41562-026-02580-8
  7. 7Living with AI Companions: Sustained AI Companionship Predicts Lower Well-Being Through Lower Human InteractionYutong Zhang, Dora Zhao, Yixin Wang, Rebecca Anselmetti, Jeffrey T. Hancock, Robert Kraut & Diyi Yang, arXiv:2609.07243 (Preprint, Version 3 vom 20.10.2026)September 7, 2026arxiv.org/abs/2609.07243v3
  8. 8Nutzung und Wahrnehmung von KI-Tools. Ergebnisse der ARD/ZDF-Medienstudie 2026Thomas Kupferschmitt, Media Perspektiven 25/2026 (ARD/ZDF-Medienstudie)September 22, 2026www.media-perspektiven.de/fileadmin/user_upload/media-perspektiven/pdf/2026/MP_25_2026_ARD-ZDF-Medienstudie_2026_Nutzung_und_Wahrnehmung_von_KI-Tools.pdf
  9. 9ARD/ZDF-Medienstudie 2026: Basispräsentation (Ergebnis-Charts)ARD/ZDF-Medienstudie 2026, Durchführung GIM; veröffentlicht über Media PerspektivenSeptember 22, 2026www.media-perspektiven.de/fileadmin/user_upload/media-perspektiven/Medienstudie/Basispraesentation_ARD-ZDF-Medienstudie_2026.pdf
  10. 10Young people - digital worldEurostat, Statistics Explained (Datenstand März 2026, Fassung vom 17.09.2026)September 17, 2026ec.europa.eu/eurostat/statistics-explained/index.php?title=Young_people_-_digital_world
  11. 11Parent, pupil and learner voice: March 2026Department for Education (England), Parent, Pupil and Learner Voice Omnibus SurveyJune 25, 2026www.gov.uk/government/publications/parent-pupil-and-learner-voice-omnibus-surveys-for-2025-to-2026/parent-pupil-and-learner-voice-march-2026
  12. 12Men Are Experiencing More Job Growth Than Women in AI-Exposed OccupationsNicole Smith, Pew Research Center (Short Read)October 22, 2026www.pewresearch.org/short-reads/2026/10/22/men-are-experiencing-more-job-growth-than-women-in-ai-exposed-occupations/
  13. 13Do people trust China, the U.S. or the EU to regulate AI?Laura Silver, Richard Wike, Moira Fagan & Jordan Lippert, Pew Research Center (Kapitel des Berichts „Globally, More People Expect AI to Cause Job Loss Than Growth“)September 17, 2026www.pewresearch.org/global/2026/09/17/do-people-trust-china-the-u-s-or-the-eu-to-regulate-ai/
  14. 14Americans’ views of data centers have turned more negativeBrian Kennedy, Pew Research Center (Short Read)September 22, 2026www.pewresearch.org/short-reads/2026/09/22/americans-views-of-data-centers-have-turned-more-negative/
  15. 15Contribution of artificial intelligence research to the development of inclusive applications for people with disabilities: A scoping review (2015–2025)S. S. Kaboré, A. Bayani, L. P. Epoh Ewane, D. Santos Oliveira dos Anjos, J. N. Nikiéma & M. Mac-Seing, PLOS Digital Health 5(10), e0001749October 15, 2026doi.org/10.1371/journal.pdig.0001749
  16. 16How Safe Is ChatGPT Health Advice? Evidence from Real User Conversations in Four Global South CountriesShreyasi Roy Chowdhury & Kiran Garimella, arXiv:2610.18465 (Preprint, Version 1)October 13, 2026arxiv.org/abs/2610.18465v1
  17. 17What do AI chatbots do after users decline therapeutic redirection for suicide thoughts?Subigya Nepal, Ruth Heo, Emma M. Parrish & Colin A. Depp, arXiv:2610.21253 (Preprint, Version 1)October 15, 2026arxiv.org/abs/2610.21253v1
  18. 18From GAN-Based Models to GPT-5: A Taxonomy-Guided Study of Human Performance in Detecting Deepfake Face ImagesZ. Guo, L. Jia, L. Sun, X. Chen, G. Yu, L. Wang, G. Yang, D. Stillwell, T. Holz & S. Li, arXiv:2610.23144 (Preprint, Version 2 vom 19.10.2026; laut Autoren für IEEE S&P 2027 angenommen)October 16, 2026arxiv.org/abs/2610.23144v2
  19. 19AI-Generated Disinformation in the UK: Risk of Harm, Context, and ClassificationPeter Cunliffe-Jones (University of Westminster), Joseph O’Leary, David Corney & Andrew Dudfield (Full Fact), arXiv:2610.06084 (Bericht, gefördert über den AISI Challenge Fund)October 5, 2026arxiv.org/abs/2610.06084v1
  20. 20Creative destruction: Human-AI competition in an experimental marketJohn Conlon & Peter Schwardmann, arXiv:2610.18921 (Preprint, Version 1)October 12, 2026arxiv.org/abs/2610.18921v1
  21. 21Generative AI Produces and Transmits Beauty BiasA. Alvero, D. Varuvel Dennison, A. Crespin, H. Kim, A. Regla-Vargas, A. Akerkar, J. McNealy & R. Kizilcec, arXiv:2610.22553 (Preprint, Version 1)October 15, 2026arxiv.org/abs/2610.22553v1
  22. 22Estimating data center water use: Best practices and critical questionsEric Masanet, arXiv:2610.04735 (Preprint, Version 1)October 3, 2026arxiv.org/abs/2610.04735v1
  23. 23AI Advice Reinforces Traditional Gender RolesNadja Born, Jared R. Curhan & Hannah Riley Bowles, arXiv:2610.14036 (Preprint, Version 1)October 1, 2026arxiv.org/abs/2610.14036v1
  24. 24Same Person, Different Advice: Auditing LLM Financial Advice with Repeated Risk-Tolerance ResponsesS. Bakshi, A. Bose, A. Bose & S. Nandi (unabhängige Forscher), arXiv:2610.23060 (Preprint, Version 1)October 16, 2026arxiv.org/abs/2610.23060v1
  25. 25Uber fined nearly 825 million euros for automated driver blockingAutoriteit Persoonsgegevens (niederländische Datenschutzbehörde), Mitteilung zur BußgeldentscheidungAugust 21, 2026www.autoriteitpersoonsgegevens.nl/en/current/uber-fined-nearly-825-million-euros-for-automated-driver-blocking
Authors, reuse basis and rights records
  1. 1Ted Van Green, Pew Research Center (Short Read mit Topline und Methodenbericht) · Reuse basis: Pew terms of use: use permitted with attribution to 'Pew Research Center' as long as notices are kept and no policy position is attributed to Pew; translations must state that Pew has not reviewed them; our own paraphrase and link, no graphics · Rights and license record
  2. 2L. Luettgau, H. R. Kirk, K. Hackenburg, J. Bergs, H. Davidson, H. Ogden, D. Siddarth, S. Huang & C. Summerfield, PNAS Nexus 5(10), pgag305 · Reuse basis: © Crown copyright 2026; open-access article containing public sector information licensed under the Open Government Licence v3.0; our own paraphrase with attribution and link, no figures · Rights and license record
  3. 3G. M. Rosenbaum, A. Goren, M. Shermohammed, D. M. Wolk, A. M. Tice, J. J. Doyle, M. N. Meyer & C. F. Chabris, Nature Human Behaviour · Reuse basis: © 2026 the authors, under exclusive licence to Springer Nature; all rights reserved; our own short factual paraphrase of the publicly available abstract and link only, no figures · Rights and license record
  4. 4M. Reis, J. Choi, Y. J. Kim, Y. L. Luan, P. Lyu & F. Reis, npj Digital Medicine 9, 743 · Reuse basis: CC BY 4.0; our own paraphrase with author, publication and link attribution, no figures or tables · Rights and license record
  5. 5Szymon Sacher, Tess Cotter, Charles I. Jones, Anton Korinek & Peter McCrory, Anthropic (The Anthropic Institute) · Reuse basis: © 2026 Anthropic PBC, no open licence; our own short factual paraphrase and link only, no figures · Rights and license record
  6. 6I. Corpus, E. Gilbert, A. Koenecke & M. Naaman, Nature Human Behaviour · Reuse basis: © 2026 the authors, under exclusive licence to Springer Nature; all rights reserved; our own short factual paraphrase of the publicly available abstract and link only, no figures · Rights and license record
  7. 7Yutong Zhang, Dora Zhao, Yixin Wang, Rebecca Anselmetti, Jeffrey T. Hancock, Robert Kraut & Diyi Yang, arXiv:2609.07243 (Preprint, Version 3 vom 20.10.2026) · Reuse basis: CC BY 4.0; our own paraphrase with author, version and link attribution, no figures · Rights and license record
  8. 8Thomas Kupferschmitt, Media Perspektiven 25/2026 (ARD/ZDF-Medienstudie) · Reuse basis: © ARD MEDIA GmbH (Media Perspektiven); reproduction or redistribution of any part only with written consent; our own factual paraphrase and link only, no reuse of text, figures or tables · Rights and license record
  9. 9ARD/ZDF-Medienstudie 2026, Durchführung GIM; veröffentlicht über Media Perspektiven · Reuse basis: © ARD MEDIA GmbH (Media Perspektiven); reproduction or redistribution of any part only with written consent; our own factual paraphrase and link only, no reuse of text, figures or tables · Rights and license record
  10. 10Eurostat, Statistics Explained (Datenstand März 2026, Fassung vom 17.09.2026) · Reuse basis: © European Union; CC BY 4.0, reuse permitted with attribution; our own paraphrase and link, no maps or figures · Rights and license record
  11. 11Department for Education (England), Parent, Pupil and Learner Voice Omnibus Survey · Reuse basis: © Crown copyright 2026; Open Government Licence v3.0; our own paraphrase with attribution and link · Rights and license record
  12. 12Nicole Smith, Pew Research Center (Short Read) · Reuse basis: Pew terms of use: use permitted with attribution to 'Pew Research Center' as long as notices are kept and no policy position is attributed to Pew; translations must state that Pew has not reviewed them; our own paraphrase and link, no graphics · Rights and license record
  13. 13Laura Silver, Richard Wike, Moira Fagan & Jordan Lippert, Pew Research Center (Kapitel des Berichts „Globally, More People Expect AI to Cause Job Loss Than Growth“) · Reuse basis: Pew terms of use: use permitted with attribution to 'Pew Research Center' as long as notices are kept and no policy position is attributed to Pew; translations must state that Pew has not reviewed them; our own paraphrase and link, no graphics · Rights and license record
  14. 14Brian Kennedy, Pew Research Center (Short Read) · Reuse basis: Pew terms of use: use permitted with attribution to 'Pew Research Center' as long as notices are kept and no policy position is attributed to Pew; translations must state that Pew has not reviewed them; our own paraphrase and link, no graphics · Rights and license record
  15. 15S. S. Kaboré, A. Bayani, L. P. Epoh Ewane, D. Santos Oliveira dos Anjos, J. N. Nikiéma & M. Mac-Seing, PLOS Digital Health 5(10), e0001749 · Reuse basis: CC BY 4.0; our own paraphrase with author, publication and link attribution, no figures or tables · Rights and license record
  16. 16Shreyasi Roy Chowdhury & Kiran Garimella, arXiv:2610.18465 (Preprint, Version 1) · Reuse basis: arXiv's standard non-exclusive distribution licence; rights remain with the authors; our own short factual paraphrase and link only, no figures · Rights and license record
  17. 17Subigya Nepal, Ruth Heo, Emma M. Parrish & Colin A. Depp, arXiv:2610.21253 (Preprint, Version 1) · Reuse basis: CC BY 4.0; our own paraphrase with author, version and link attribution, no figures · Rights and license record
  18. 18Z. Guo, L. Jia, L. Sun, X. Chen, G. Yu, L. Wang, G. Yang, D. Stillwell, T. Holz & S. Li, arXiv:2610.23144 (Preprint, Version 2 vom 19.10.2026; laut Autoren für IEEE S&P 2027 angenommen) · Reuse basis: arXiv's standard non-exclusive distribution licence; rights remain with the authors; our own short factual paraphrase and link only, no figures · Rights and license record
  19. 19Peter Cunliffe-Jones (University of Westminster), Joseph O’Leary, David Corney & Andrew Dudfield (Full Fact), arXiv:2610.06084 (Bericht, gefördert über den AISI Challenge Fund) · Reuse basis: CC BY-SA 4.0; our own paraphrase with author and link attribution only, no reuse of text passages or figures · Rights and license record
  20. 20John Conlon & Peter Schwardmann, arXiv:2610.18921 (Preprint, Version 1) · Reuse basis: arXiv's standard non-exclusive distribution licence; rights remain with the authors; our own short factual paraphrase and link only, no figures · Rights and license record
  21. 21A. Alvero, D. Varuvel Dennison, A. Crespin, H. Kim, A. Regla-Vargas, A. Akerkar, J. McNealy & R. Kizilcec, arXiv:2610.22553 (Preprint, Version 1) · Reuse basis: CC BY 4.0; our own paraphrase with author, version and link attribution, no figures · Rights and license record
  22. 22Eric Masanet, arXiv:2610.04735 (Preprint, Version 1) · Reuse basis: arXiv's standard non-exclusive distribution licence; rights remain with the authors; our own short factual paraphrase and link only, no figures · Rights and license record
  23. 23Nadja Born, Jared R. Curhan & Hannah Riley Bowles, arXiv:2610.14036 (Preprint, Version 1) · Reuse basis: arXiv's standard non-exclusive distribution licence; rights remain with the authors; our own short factual paraphrase and link only, no figures · Rights and license record
  24. 24S. Bakshi, A. Bose, A. Bose & S. Nandi (unabhängige Forscher), arXiv:2610.23060 (Preprint, Version 1) · Reuse basis: CC BY 4.0; our own paraphrase with author, version and link attribution, no figures · Rights and license record
  25. 25Autoriteit Persoonsgegevens (niederländische Datenschutzbehörde), Mitteilung zur Bußgeldentscheidung · Reuse basis: Copyright of the Autoriteit Persoonsgegevens; under its copyright notice texts may be quoted and used with attribution, photos, images and the logo may not; our own factual paraphrase with attribution and link · Rights and license record

How to cite

qasb Research (2026). AI and society: growing but unevenly distributed use. Topic report 10/2026, data cutoff October 26, 2026. https://www.qasb.ai/en/reports/2026-10-gesellschaft

Appendix

For every statement: evidence type, change status, scope and limitations, and the observation period, source value, comparison and uncertainty where the source gives them. The texts are the approved entries of the edition, as is its summary.

Scope and limitations of all 25 statements (A1–A25)

Edition summary

The originals from October and September 2026 show growing but unevenly distributed everyday use of AI: only a minority turns to chatbots for election information, younger people in Germany use AI tools far more often than older people, and audits of real health conversations, simulated crisis dialogues and salary and financial advice from language models show measurable risks and biases. The evidence mainly covers usage shares, attitudes and model behaviour tested in audits and experiments; lasting effects on elections, health, income or social relationships are not established, and several findings rest on preprints not yet peer reviewed or on surveys from spring 2026.

  1. A1Back to chapter 2.0

    Evidence: observed · baseline · Sources 8, 9

    Scope
    Random sample of 2,462 German-speaking people aged 14 and over in Germany (70% by telephone, 30% online), surveyed by the GIM institute from 27 January to 19 April 2026, weighted; compared with the previous year (n = 2,512). 59% used AI tools at least occasionally (2025: 42%).
    Limitations
    Self-reported private use; AI summaries in search engines were not measured separately. The publishers are the public broadcasters ARD and ZDF, which run media services themselves. No uncertainty stated in the text checked; fieldwork in spring, published in September 2026.
    Observation period
    27 January–19 April 2026; published 22 September 2026
    Source value
    37% of the German-speaking population aged 14+ (private use of AI tools at least weekly)
    Baseline
    2025: 22%
    Comparison
    Ages 14–29: 68%; 70 and over: 8%
    Uncertainty
    No uncertainty stated in the source

  2. A2Back to chapter 2.0

    Evidence: observed · baseline · Source 10

    Scope
    Eurostat Statistics Explained article Young people - digital world (version of 17 September 2026, data extracted in March 2026) based on the EU survey on ICT usage in households 2025 ('total population' here means people aged 16 to 74); among 16- to 29-year-olds, country values ranged from 38% (Romania) to 82% (Greece).
    Limitations
    Earlier context data: the values refer to survey year 2025 and were in the Eurostat database as of March 2026; the article is dated 17 September 2026, and an earlier version cannot be ruled out. Self-reports without intensity, purpose or effects of use; children under 16 are not covered.
    Observation period
    Survey year 2025; data extracted March 2026; article version 17 September 2026
    Source value
    60% of 16- to 29-year-olds in the EU (use of generative AI tools in the three months before the 2025 survey)
    Baseline
    All people aged 16 to 74: 33%
    Comparison
    Greece 82%, Romania 38%; gap between young people and total population in Germany 21 percentage points
    Uncertainty
    No uncertainty stated in the article

  3. A3Back to chapter 2.0

    Evidence: observed · baseline · Source 11

    Scope
    Department for Education Parent, Pupil and Learner Voice survey: online survey from 11 March to 7 April 2026, panel of parents and pupils randomly drawn from the National Pupil Database; bases 1,315 key stage 3 and 1,003 key stage 4 pupils who went online, plus 1,096, 1,511 and 601 parents of primary, secondary and special-school pupils.
    Limitations
    Earlier context study (published 25 June 2026); England only. Multiple-choice answers without frequency, purpose or effects of use; the question does not define AI tools further.
    Observation period
    11 March–7 April 2026; published 25 June 2026
    Source value
    48% of key stage 4 pupils in England who go online (use of AI tools or apps)
    Baseline
    Key stage 3: 33%
    Comparison
    Parents who would use AI tools for advice: 20% (primary), 22% (secondary), 14% (special school)
    Uncertainty
    No uncertainty stated in the report

  4. A4Back to chapter 3.0

    Evidence: observed · baseline · Source 2

    Scope
    PNAS Nexus, 6 October 2026: online survey of 2,499 eligible voters in the four days after the election of 4 July 2024, weighted to official population characteristics; two randomised experiments with 2,858 British participants in total (Prolific) on four topics with GPT-4o, Claude 3.5 and Mistral as well as sycophantic or persuasive prompted model variants.
    Limitations
    One country, one election, four topics, short task-directed sessions and factual knowledge questions; voting behaviour and long-term effects were not measured. The data are from 2024; a preprint has been public since September 2025. Most authors are from the UK AI Security Institute; according to the competing-interests statement, one author is now employed at an AI company (Anthropic), and the work was done while at the institute.
    Observation period
    Survey in the four days after the election of 4 July 2024; timing of the experiments not stated in the text checked; published 6 October 2026
    Source value
    13% of eligible voters in the United Kingdom (upper bound)
    Baseline
    Weighted sample of 2,499 eligible voters
    Comparison
    as many as 32% of chatbot users
    Uncertainty
    No uncertainty stated; the authors present the value as an upper bound ('as many as')

  5. A5Back to chapter 3.0

    Evidence: observed · baseline · Source 1

    Scope
    Survey of 5,273 US adults in the American Trends Panel, 12–18 October 2026, weighted to the US adult population; follow-up questions to the 427 people using chatbots for election information.
    Limitations
    Self-reports; the survey did not measure what content the chatbots provided, how accurate it was, or whether it influenced voting decisions. The panel's cumulative response rate is 3%; the subgroup of users is small.
    Observation period
    12–18 October 2026
    Source value
    8% of US adults
    Baseline
    US adults, weighted panel sample (n = 5,273)
    Comparison
    Under 50: 11%; 50 and older: 4%; among chatbot users: 15%
    Uncertainty
    Margin of sampling error for the full sample ±1.5 percentage points (Pew methodology report)

  6. A6Back to chapter 3.0

    Evidence: observed · baseline · Source 8

    Scope
    Subgroup of people aged 14 and over who use AI tools at least monthly (n = 1,225) in the representative ARD/ZDF media study 2026 (GIM, 27 January–19 April 2026); question on changed use of news websites and news apps since starting to use AI tools.
    Limitations
    Retrospective self-assessment, not measured change in use and not causal evidence; 6% reported using news more often. The publishers are public broadcasters with their own interest in news reach.
    Observation period
    27 January–19 April 2026; published 22 September 2026
    Source value
    16% of people aged 14+ using AI tools at least monthly (using online news less often since)
    Baseline
    64% no change, 6% use news more often
    Comparison
    Ages 14–29: 24%; 30–49: 13%; 50 and over: 10%
    Uncertainty
    No uncertainty stated in the source

  7. A7Back to chapter 3.0

    Evidence: observed · baseline · Source 19

    Scope
    Report by Full Fact and the University of Westminster (arXiv, 5 October 2026), funded through the UK AI Security Institute’s Challenge Fund; based on fact checks published by Full Fact between 1 January 2025 and 31 March 2026, assessed with the authors’ own harm-risk model.
    Limitations
    Not a representative sample of all AI-generated false content but cases selected by a fact-checking team; the harm assessment is the authors’ expert judgement. Only in 33 of the 112 cases did the fact check state unequivocally that AI was involved. No evidence of harms that actually occurred.
    Observation period
    1 January 2025–31 March 2026; published 5 October 2026
    Source value
    41.1% of the 112 cases studied (substantive potential for specific harms)
    Baseline
    112 fact checks of AI-generated or AI-altered false content
    Comparison
    58.9% with no or very limited potential for harm; 83.9% with an overall misleading effect
    Uncertainty
    No uncertainty stated in the source

  8. A8Back to chapter 3.0

    Evidence: observed · baseline · Source 6

    Scope
    Nature Human Behaviour, 30 September 2026; natural experiment on the Change.org platform; supplementary analysis of 4,611 returning writers from countries with access to the tool, whose second petition with AI access was longer and performed worse.
    Limitations
    One platform and one specific tool; outcomes are platform-level success measures of petitions, not policy effects. The publicly available abstract and figure caption were checked; the full text is paywalled. A preprint has been public since November 2025 (arXiv 2511.13949). A September source.
    Observation period
    Period before and after the tool's introduction (not dated in the publicly available abstract); published 30 September 2026

  9. A9Back to chapter 3.0

    Evidence: observed · baseline · Source 18

    Scope
    arXiv preprint (version 2 of 19 October 2026, accepted for IEEE S&P 2027 according to the authors): 524 valid participants via Prolific, images from five models (StyleGAN2, StyleGAN3, GRAM-HD, SD3.5, GPT-5) and real photographs; general versus model-specific training materials compared; follow-up test with 342 participants after five months.
    Limitations
    Single still images of faces under test conditions, no video, voice or social-media context; an English-speaking online sample. The images were generated and selected by the research team; accuracy figures are not reproduced here.
    Observation period
    Fieldwork dates not stated in the text checked; published 16 October 2026, version 2 on 19 October 2026

  10. A10Back to chapter 4.0

    Evidence: observed · baseline · Source 17

    Scope
    arXiv preprint of 15 October 2026 (University of Virginia, University of California San Diego): 17 publicly available chatbots from 8 companies, scripted multi-turn conversations in September 2026 at three severity levels, up to ten standardised refusals and a subsequent change of topic; two consecutive versions were tested for five chatbots.
    Limitations
    Simulated scripts rather than real users; replies were coded against a codebook by another language model (Grok 4.3; checked against a human rater, 90.9% agreement) whose provider also supplies one of the chatbots tested. A snapshot of models available in September 2026; consequences for affected people were not measured. Not yet peer reviewed.
    Observation period
    September 2026; published 15 October 2026
    Source value
    99.7% of the 2,550 conversations with a referral in the first reply
    Baseline
    2,550 simulated conversations with 17 chatbots
    Comparison
    Only 173 of 8,818 replies after a refusal (2.0%) asked about barriers to seeking help
    Uncertainty
    No uncertainty stated for this share in the source

  11. A11Back to chapter 4.0

    Evidence: observed · baseline · Source 16

    Scope
    arXiv preprint of 13 October 2026: data-donation corpus of 202,590 conversations from 1,252 users (December 2022 to February 2026) with age and gender; safety rated against a published rubric by Gemini 3.1 Pro Preview, with one author checking the 128 high-risk cases in a sample for plausibility.
    Limitations
    A preprint not yet peer reviewed. A convenience sample of paid crowdworkers, mostly young and male in India and Pakistan; not a prevalence estimate for the countries. Risk ratings by a language model were measured, not actual health outcomes; the corpus mixes several model versions. The context-aware rating covers only that subset (19 to 60% of health conversations per country), which was already riskier without context, and does not extrapolate to all health conversations.
    Observation period
    Conversations December 2022 to February 2026; published 13 October 2026

  12. A12Back to chapter 4.0

    Evidence: observed · baseline · Source 7

    Scope
    arXiv preprint 2609.07243 (version 3 of 20 October 2026, first version 7 September 2026); 1,182 participants at baseline and 439 after about twelve months on average; English-speaking US users recruited via Prolific with at least one month of use.
    Limitations
    A preprint not yet peer reviewed; self-reports; associations, not causal evidence; the standardised associations are small. High attrition (62.9%), a single platform, and possible platform changes during the period could not be observed. A September source.
    Observation period
    First wave September 2024 to September 2025, second wave August 2025 to May 2026

  13. A13Back to chapter 4.0

    Evidence: observed · baseline · Source 4

    Scope
    npj Digital Medicine, 1 October 2026; Prolific sample with quotas for age, gender and ethnicity (data collected in November 2025); fictitious practice ads varying the AI statement as well as the physician's age, gender and ethnicity.
    Limitations
    Hypothetical ads and stated intentions, not real behaviour; an online panel rather than a random sample; 14% of participants excluded after manipulation checks. Stronger effects among younger participants come from exploratory analyses.
    Observation period
    November 2025; published 1 October 2026

  14. A14Back to chapter 4.0

    Evidence: observed · baseline · Source 3

    Scope
    Nature Human Behaviour, 5 October 2026; patients of one US health system classified as high-risk by a validated machine-learning model; random assignment to no message or to different message versions; primary outcome flu vaccination.
    Limitations
    One health system and one specific use case (risk messages), not a generative AI system. Vaccination behaviour was measured, not attitudes or trust in general. Only the publicly available abstract was checked; the full text is paywalled, and the data come from vaccination campaigns in 2020 to 2022, so they do not reflect today's state of AI use.
    Observation period
    Three field experiments on flu vaccination; started in September 2020, 2021 and 2022 according to the ClinicalTrials.gov records named in the abstract (NCT04323137, NCT05009251, NCT05509283); published 5 October 2026

  15. A15Back to chapter 4.0

    Evidence: observed · baseline · Source 15

    Scope
    PLOS Digital Health, 15 October 2026: search of MEDLINE, Embase, Web of Science and CINAHL from January 2015 to May 2025 (English and French), 4,233 records; 68 of the 93 studies (73.1%) came from high-income economies.
    Limitations
    Health-related research only, almost entirely in clinical settings; grey literature, preprints and conference papers were excluded, so participatory approaches may be underestimated. A statement about research designs, not about deployed products or their everyday use.
    Observation period
    Studies from January 2015 to May 2025; published 15 October 2026
    Source value
    6.5% of the 93 included studies actively involving people with disabilities
    Baseline
    93 studies from 4,233 records
    Comparison
    About 60% mentioned equity, 12.9% implemented concrete measures
    Uncertainty
    No uncertainty stated in the source

  16. A16Back to chapter 5.0

    Evidence: observed · baseline · Source 23

    Scope
    arXiv preprint of 1 October 2026 (Born, Curhan, Bowles): study 1 with 6,000 open-ended queries; study 2 with GPT-5.4-mini, Claude Sonnet 4.6, Gemini 2.5 Flash and Llama 3.3 70B, gender signalled only through first names and pronouns; compared with ratings by 500 US adults of breadwinner and caregiver roles.
    Limitations
    Model behaviour under controlled English-language prompts in a US context; real negotiations or income effects were not measured, and model versions change frequently. The gap varied widely by model (USD 1,594 to 15,915). Not yet peer reviewed.
    Observation period
    Query dates documented in the appendix according to the authors, not stated in the text checked; published 1 October 2026
    Source value
    5,820 USD higher recommended salary target for male-named candidates (pooled across four models)
    Baseline
    Otherwise identical profiles; 80,000 queries
    Comparison
    By model, USD 1,594 (Llama 3.3 70B) to USD 15,915 (Claude Sonnet 4.6)
    Uncertainty
    95% confidence interval USD 5,547 to 6,092

  17. A17Back to chapter 5.0

    Evidence: observed · baseline · Source 12

    Scope
    Pew short read of 22 October 2026: exposure scores for 480 occupations from O*NET tasks, linked to Current Population Survey monthly data for 2017–19 and January 2023 to June 2026; 2020–22 excluded because of the pandemic. Among workers with at most a high-school diploma, men’s employment grew 8.8% per point while women’s fell 9.5%.
    Limitations
    An association, not causal evidence: other changes during the period may have affected employment. Comparisons within occupations; US only. Exposure scores rest on judgements of whether generative AI can perform or assist with tasks.
    Observation period
    2017–2019 and January 2023 to June 2026; published 22 October 2026
    Source value
    6.3% additional employment growth for men per point of an occupation’s AI exposure
    Baseline
    2017–19 compared with January 2023 to June 2026
    Comparison
    Women: 0.2%
    Uncertainty
    No uncertainty stated in the short read

  18. A18Back to chapter 5.0

    Evidence: observed · baseline · Source 21

    Scope
    arXiv preprint of 15 October 2026: 51,300 generated faces from generators released between 2022 and 2026 and 160,196 photographs of real people, three automated measures and ratings by 550 human raters; hiring simulation with 570 participants, measured by hireability, trustworthiness and advancement to interview.
    Limitations
    A preprint not yet peer reviewed; the ordering by skin tone rests on automated measures that agree with each other only moderately. The hiring simulation is an online experiment with hypothetical decisions, not real recruitment.
    Observation period
    Generators released 2022 to 2026; study published 15 October 2026

  19. A19Back to chapter 5.0

    Evidence: observed · baseline · Source 24

    Scope
    arXiv preprint of 16 October 2026 (independent researchers): household profiles from the 2007–09 Survey of Consumer Finances panel, in which 41% of households answered the risk question differently the second time; models including Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 3.8 Flash, Grok 4.7 and Mistral Large 4; compared with 0.6 to 4.2 points for identical repeated prompts.
    Limitations
    One-turn advice without follow-up questions; in a scripted test, one clarifying question removed nearly all of the swing. Older panel data; real investment decisions or wealth effects were not measured. Not yet peer reviewed.
    Observation period
    SCF panel 2007–2009; model query dates not stated in the text checked; published 16 October 2026

  20. A20Back to chapter 5.0

    Evidence: observed · baseline · Source 25

    Scope
    Notice of 21 August 2026 by the Autoriteit Persoonsgegevens on its decision under the GDPR prohibition of fully automated decisions and for insufficient information to drivers; proceedings followed complaints by 171 French drivers lodged via the human rights organisation LDH with the CNIL, coordinated with other European supervisory authorities.
    Limitations
    Earlier context decision of August 2026, not yet final. The authority’s notice was checked, not the decision itself. It concerns algorithmic decisions by a platform, not the use of generative AI.
    Observation period
    Infringements 2018–2022; notice 21 August 2026

  21. A21Back to chapter 5.0

    Evidence: observed · baseline · Source 20

    Scope
    arXiv preprint of 12 October 2026 (Ohio State University, Carnegie Mellon University): online experiment via Prolific with producer and consumer roles; 296 amateurs, 180 participants from creative industries (game development, graphic design, content creation) and 549 consumers; incentive-compatible measurement of willingness to pay.
    Limitations
    Simple browser games in an experiment with small stakes; no real market and no measurement of longer-term effects on creative occupations. The findings apply to the models used. Not yet peer reviewed.
    Observation period
    Fieldwork dates not stated in the text checked; published 12 October 2026

  22. A22Back to chapter 6.0

    Evidence: observed · baseline · Source 14

    Scope
    Pew short read of 22 September 2026: survey of 10,548 US adults in the American Trends Panel from 20 July to 9 August 2026, compared with a January 2026 survey; among rural adults, the share with a negative view of environmental effects rose from 32% to 50%.
    Limitations
    Attitudes, not measured environmental or price effects. The short read asks about data centres in general and does not separate AI data centres. No uncertainty stated in the short read; US only.
    Observation period
    20 July–9 August 2026, compared with January 2026; published 22 September 2026
    Source value
    54% of US adults (data centres mostly bad for the environment)
    Baseline
    January 2026: 39%
    Comparison
    Household energy costs 50% (January 38%); quality of life nearby 49% (January 30%)
    Uncertainty
    No uncertainty stated in the short read

  23. A23Back to chapter 6.0

    Evidence: observed · baseline · Source 13

    Scope
    Representative surveys of 13,696 adults in Bangladesh, Chile, Ghana, Colombia, Malaysia, Pakistan, Peru, the Philippines, Singapore, Sri Lanka, Thailand and the West Bank and East Jerusalem (8 February–13 May 2026); part of a 36-country study published on 17 September 2026.
    Limitations
    Neither EU member states nor the US are among the twelve places. Trust was asked as a perception and, according to Pew, depends strongly on general views of China, the US and the EU. Fieldwork in spring 2026.
    Observation period
    8 February–13 May 2026; published 17 September 2026
    Source value
    43% (median across twelve places) trusting China to regulate AI effectively
    Baseline
    Median US 35%, EU 34%
    Comparison
    In Pakistan and Bangladesh about eight in ten trusted their own country
    Uncertainty
    No uncertainty stated for the medians in the source

  24. A24Back to chapter 6.0

    Evidence: observed · baseline · Source 5

    Scope
    Report of 16 October 2026: online survey of 10,980 US adults by Morning Consult (11–23 August 2026, weighted) and an internal, unweighted poll of 1,334 Anthropic staff (9–12 August 2026); only half of the US sample was asked about 2030, the other half about 2027.
    Limitations
    Expectations, not measurements of actual capabilities or labour-market effects. An opt-in online panel whose weighting cannot correct all biases; no peer review. Different answer formats: staff used a slider, while US adults chose ranges whose midpoints were used. Conflict of interest: the publisher is an AI developer that surveyed its own staff. A repeat survey in September (7,613 US adults) found 44% for the Nobel question.
    Observation period
    11–23 August 2026 (US adults); 9–12 August 2026 (staff); repeat survey 17–24 September 2026
    Source value
    39% of US adults
    Baseline
    Weighted online sample of 10,980 US adults (Morning Consult, August 2026)
    Comparison
    91% of 1,334 surveyed Anthropic staff (unweighted)
    Uncertainty
    No uncertainty stated for this share in the source

  25. A25Back to chapter 6.0

    Evidence: model estimate · baseline · Source 22

    Scope
    arXiv preprint of 3 October 2026 (Eric Masanet): 14 proposed best practices, assessment of 31 studies with a structured rubric and a case study using observed design variations of US studies.
    Limitations
    A single-author review, not yet peer reviewed. The 24-fold range is a modelled spread of possible outcomes, not measured use; no statement about individual sites.
    Observation period
    Literature cut-off not dated in the text checked; published 3 October 2026

And in your case? Check it with sources.

Start an analysisFor organizations
All reports