Epistemic AI, deepfakes and trust
Every day we trust things we cannot check for ourselves: a map, a news report, a doctor's chart. That trust works when sources are usually right and mistakes can be traced and corrected. AI changes both sides of that bargain. It can produce fluent answers, images and voices that are wrong or invented, and it can make true evidence easier to dismiss as fake. It can also check facts at a scale no newsroom could manage, point readers to better sources and help people reason together. The question in this area is whether our shared sense of what is real grows sturdier or more brittle.
Wonder
Truth and common ground
Tools help us check what is true and find where we agree.
About 20%
drop in belief in a chosen conspiracy theory after a conversation with an AI, among 2,190 believers. It lasted at least two months. Measured
22.8% to 38.6%
Share of small groups whose members all took the same side, before and after AI-mediated discussion, in one UK experiment. Measured
70%
of the time, crowd raters preferred an AI system's suggested source over the existing one, for the Wikipedia citations it judged most likely to be unverifiable. Measured
24% to 28%
more accurate predictions from people who could consult an advanced AI forecasting assistant, against a control group using a weaker model, in an experiment with 991 people. One outlier question strongly affected the results. Measured
It takes Systems built for accuracy and agreement instead of attention. No AI has yet stopped a war.
Worry
Truth and trust
55.5%
is how often people correctly identify deepfakes, which is not reliably better than chance (50%). Measured
45%
of AI-assistant answers about the news had at least one significant issue. Measured
$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. Measured
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. Measured
17% to 33%
of the time, three leading legal AI research tools gave answers containing hallucinations in 2024 tests on 202 legal questions. Measured
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. Reported
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. Measured
What is being done
Europe now requires AI to say what it is.
Since 2 August 2026, the EU AI Act has required chatbots to tell people they are talking to an AI, generative AI providers to mark their output in a machine-readable way, and those who publish deepfakes to disclose them. Systems already on the market have until 2 December 2026 to meet the marking duty.
For artists and musicians
Most listeners cannot tell, and most want to be told.
In an eight-country survey of 9,000 adults commissioned by Deezer, 97% failed a blind test to pick out fully AI-generated songs, and 80% wanted fully AI-made music labelled. The test used only three tracks.
Worth knowing
A correct machine fact-check did not help readers.
In a randomized experiment, an AI model correctly flagged about 90% of false headlines, yet showing people its fact-checks did not improve their ability to tell true from false. Human-written fact-checks did help.
Worth knowing
Supply may matter less than demand.
Some researchers argue that misinformation's reach is limited mainly by demand, driven by partisanship and distrust, more than by how much of it is produced. In a 2026 survey across 48 countries, about 10% of people had used AI chatbots for news in the past week.
An open question
Will widespread AI labels restore trust, or teach people to doubt everything that carries no label?
This page does not list everything. Last reviewed October 11, 2026.
- Measured counted or experimental data
- Survey what people report or believe
- Estimate a judgment or model, often a company's own
- Forecast a projection of the future
- Reported a company's, agency's or journalist's own account