Can AI deepfakes influence elections?
AI makes it cheap to create realistic fake images, audio and video, and harm from such content is well documented. Evidence that it has swayed elections is so far limited, and Sweden's 2026 election saw no coordinated foreign influence campaign.
What has been observed
Generative AI can produce text, images, audio and video that are hard to tell from the real thing. The International AI Safety Report 2026 cites studies in which people took AI-written text for human writing 77% of the time, and mistook AI-generated voices for real speakers 80% of the time. Many tools are free or cheap, need no technical skill and can be used anonymously.
The clearest documented harm is not political but personal. Studies cited in the report show that 96% of deepfake videos online are pornographic, and the victims are mostly women and girls. The report also notes that systematic data on how common and how serious these harms are is still lacking, partly because victims often do not report them.
Influence and elections
In laboratory experiments, AI systems are often at least as persuasive as non-expert humans. Evidence of their effect in the real world remains limited, according to the International AI Safety Report, and some research suggests that the cost of reaching people and the general difficulty of changing minds put a ceiling on the impact.
Sweden offers a concrete example of both sides. In September 2026 the Swedish Psychological Defence Agency (MPF) described an AI-driven bot network it calls Holiday: about 1,200 fake accounts on X that published more than 50,000 posts, comments and shares in 2025 and 2026, taking opposing sides on political issues to deepen divisions. MPF assesses that a state actor bought the network from a company in East Asia. X suspended the accounts in the summer of 2026. After the general election on 13 September 2026, MPF and the National Cyber Security Centre reported no major cyberattacks or coordinated foreign information influence against the election.
Why it is hard to solve
Detection tools struggle to keep up with the quality of generated content, and it is often difficult to trace a fake back to whoever made it. Some content, such as sexual images of children, causes harm even when everyone knows it is AI-generated, so detection alone is not enough.
What is being done
Since August 2026 the EU AI Act requires deepfakes, and AI-generated text published to inform the public, to be clearly labelled. Providers of generative AI must make their output identifiable, and chatbots must tell users they are talking to a machine. A voluntary code of practice on labelling supports these rules. From 2 December 2026, AI systems that generate non-consensual sexual content or child sexual abuse material are banned in the EU.
Researchers see watermarking and content provenance records as promising, alongside media literacy. In Sweden, MPF works with other authorities and with platforms to map and stop influence operations, as in the Holiday case.
What you can do
- Before sharing a striking image, clip or quote, check who first published it and whether established news outlets report the same thing.
- Be extra careful with emotionally charged content close to elections; that is when fakes spread fastest.
- Report fake accounts and manipulated content to the platform.
- If someone makes sexual deepfakes of you or your child, save the evidence and report it to the police and the platform.
Sources
- International AI Safety Report 2026 – International AI Safety Report (chair: Yoshua Bengio), 2026
- AI Act – regulatory framework for AI – European Commission, 2026
- Botnätverket ”Holiday” användes mot Sverige – kunde stängas ner efter kartläggning av MPF – Myndigheten för psykologiskt försvar (MPF), 2026
- Inga större informations- eller cyberangrepp mot svenska valet – Myndigheten för psykologiskt försvar (MPF), 2026
Data updated: 2026-11-19