False and misleading information: evidence summary and map
A · Strong evidence Commissioned by the UK Government Office for Science, the Sheffield and Edinburgh team maps 228 review-level studies plus 1,239 recent primary studies and a consultation with 11 experts including Sander van der Linden and Stephan Lewandowsky, though as a rapid scoping review it does not grade evidence quality.
Summary based on the full text·Original: about 150 minutes to read
In short
A 136-page evidence summary and map (May 2026) commissioned by the UK Government Office for Science, led by Harriet Baird, Hui Zhang, Lesley Uttley, Kalina Bontcheva, Thomas Webb and Joanna Wright (University of Sheffield) with Linda Bauld (University of Edinburgh). It covers how false and misleading information should be understood and measured, what harm it does, and which interventions work. Its conclusions: the problem is now permanent and multi-topic, generative AI multiplies it across all topics, and no single intervention is enough. Prebunking has the strongest individual-level evidence, while voluntary platform self-regulation is judged insufficient.
Key numbers
- 228review-level studies (2020-2026) mapped by the team
- 28.6% to 97.1%jump in the share of months with AI-generated false content in EDMO briefs after March 2023
- 3.0 to 5.9average active false-information topics per month, 2021 versus 2025-26
- 16%AI-generated content's share of tracked false and misleading content by December 2025
As reported by the source.
English·PDF·opens usher.ed.ac.uk
What it reports
- Across 56 months of European Digital Media Observatory briefs (June 2021 to January 2026), no category fell below 71% monthly prevalence; health appeared in 100% of months, the Ukraine conflict in 92.9%, immigration 89.3%, climate 83.9%.
- Active topics per month rose from 3.0 (2021) to 4.2 (2022), 5.1 (2023), 5.7 (2024) and 5.9 (2025-26).
- AI-generated false content went from 28.6% to 97.1% of months after March 2023 and reached 16% of tracked content by December 2025; the first political deepfake audio was recorded in September 2023.
- The EU External Action Service reported 259% growth in AI-enhanced foreign information manipulation in Europe in 2025.
- Generative AI is also used to build networks of fake websites meant to poison other models' training data.
- 61% of impact reviews concern health misinformation; evidence on climate, democracy, economic decisions and marginalised groups is thin.
- A meta-analysis of 42 studies (42,530 participants) found inoculation improves people's ability to tell true from false content but does not significantly change intention to share misinformation.
- Organisational harms (reputation, market swings, workplace productivity and engagement) are documented but least studied.
How they got it
Between December 2025 and March 2026 the team searched Scopus and PsycInfo for reviews with key terms in titles (228 found), ran a supplementary search for primary studies from 2025 (1,239), added 22 grey-literature items, 19 expert suggestions and the EDMO fact-checking database. Google Gemini generated initial themes, followed by human thematic coding of all 228 references. Eleven experts took part in consensus workshops.
Limits
- As a scoping review it did not extract effect sizes or rate evidence quality; the authors say rigorous systematic reviews are still needed.
- Title-only search terms and a six-month timeline mean relevant work will have been missed.
- The EDMO trend analysis counts monthly topic presence from fact-check briefs, not the volume or reach of false content.
Why it matters for companies
The report treats false information as a permanent operating condition, which for a company means reputation and crisis plans built for continuous exposure rather than one-off incidents. AI-generated audio and video of named people is now routine, so verification steps for executive voice or video instructions belong in fraud and communications controls. For internal communication the evidence favours prebunking: explaining a manipulation technique before staff meet it works better than correcting afterwards, though it needs repeat reminders. The finding that workplace effects are under-researched means companies should not expect off-the-shelf benchmarks.
Summary based on the full text. Read the original for the full argument and data.Open PDF on usher.ed.ac.uk