Full Fact

Google’s AI detector told us these five images were “made or edited with Google AI”. Those who posted them insist they’re not AI, and the images appear to be real and unaltered. After we contacted Google, it fast-tracked an update to its SynthID Detector. The tool now says it’s “unsure” about all five. But days later, another politician’s image was still flagged, and she says it’s genuine too. AI detectors can be a useful tool, but they should never be the only test of whether an image is real or fake. Always look for other signs. Read the full investigation: https://lnkd.in/e657yiMc

Google has updated the AI detector tool it recently launched to the public, after a Full Fact investigation found it was flagging apparently real images. We identified five examples of images shared by politicians in recent months which the SynthID Detector said were “made or edited” with AI - even though the images appear to be real and unaltered, and those who posted them insist they’re not AI. After we showed Google the findings of our investigation, it changed the tool to make some of its verdicts less certain. The public-facing tool - which can detect invisible SynthID watermarks on imagery which was made or edited with AI tools from Google, OpenAI and others - now says it is ‘Unsure’ about the five images we originally looked at. We welcome the public launch of the tool and the changes Google has made after our investigation. The tool is a valuable one, and in fact we’ve used an early version of it which was made available to the media to help with our fact checking for some time. But we don’t rely solely on SynthID to prove whether an image is fake (or real), and if you’re using the detector tool to assess if an image is AI-generated, it’s important to always check for other signs too. Read the full investigation: https://lnkd.in/e657yiMc Get our fact checks and investigations in your inbox: https://lnkd.in/eBKnHjSK

Over 10 weeks from mid-July, our team of fact checkers and data scientists asked five widely used AI models the same 22 questions every day: Gemini 3.1, ChatGPT, Grok, DeepSeek and Claude. We scored tens of thousands of individual claims against five criteria: factuality, transparency, timeliness, consistency and civic responsibility. No model repeated the core piece of misinformation it was tested on, but the exercise exposed several vulnerabilities and points of failure.  Among the findings: ▪️ Rules not yet in force were reported as law. Chatbots said bank cards could be used as voter ID, and that a £10,000 asylum rule was in force, despite it not having been adopted. ▪️ Outdated figures were presented as current. Several models quoted a weekly asylum support rate of £8.86 after it had risen to £9.95. ▪️ UK-wide answers ignored differences between the nations. Voter ID rules vary across the UK, but models frequently omitted the caveat needed to make their answers accurate. ▪️ The same question produced different answers on different days. One model correctly identified 2025 as the answer to when England last won a major football trophy, but gave 2022 on other days. So, we’re calling on AI companies to ensure models check live, authoritative sources for time-sensitive facts; distinguish clearly between rules in force and proposals; retain essential caveats even when this makes the answers longer; and be able to account for the sources they cite. This is the first in a series of reports we plan to publish, and we will have much more to say in the months ahead. This project has been made possible thanks to a grant from the The Patrick J. McGovern Foundation. Read more about our methodology and key findings here: https://lnkd.in/e5SjGcbB

We thought it was about time we reintroduced ourselves! 👋 We're Full Fact: a team of independent fact checkers, journalists, technologists and policy experts on a mission to build a better information environment to restore trust. We know accurate information helps people make better decisions and supports healthier, more connected communities. As fact checkers, we’ve seen first hand how reliable information can build understanding, and strengthen democracy. By improving the quality of information we all rely on, we can build a more informed society where trust can grow. ▶️ Watch it here:

AI-generated false content now makes up nearly a quarter of everything we fact check, but how many of these pieces actually cause harm? In a new study funded through the AISI Challenge Fund, Full Fact and the University of Westminster analysed 112 pieces of online AI-generated or AI-altered misinformation that circulated in the UK between January 2025 and March 2026, collectively seen tens of millions of times. What we found: ▪️ Confirmed or suspected AI content accounted for 24% of our fact checks during the study period, up from 8% a year earlier. ▪️84% of the content created a substantively false or misleading understanding, contributing to a broad erosion of trust in information. ▪️However, only 41% met all the criteria for substantive potential to cause specific real-world harm. ▪️Virality isn't the same as harm. More than half the pieces viewed over a million times had limited harm potential, while a fake video of an MP defecting to Reform UK reached far fewer people yet ended up being discussed at Prime Minister's Questions. ▪️Politics featured in 63% of entries, often through imposter videos of public figures "announcing" policies that don't exist, from heating fines to flight limits. With fact checking and regulatory capacity finite, treating every fake as an emergency isn't sustainable. The question we need to ask first is: what's the harm? Explore the full findings and the model we used to assess harm: https://lnkd.in/eW8GURkY

100 days. That's how long it's been since a Conservative Party spokesman admitted that a claim in the party's Alternative King's Speech was wrong and said it would be corrected. We're still waiting. The claim was that welfare payments had overtaken income tax receipts "for the first time ever". That's not true. Welfare spending has actually been higher than income tax receipts for many years, and the position is expected to reverse this year. Even so, the claim was repeated in print, in the House of Commons, on social media and on the radio. You might know Full Fact for our fact checks, but that's only part of what we do. When public figures and institutions get things wrong, we follow up and push them to correct the record. Anyone who wants voters' trust should put things right when they make a mistake. Read more: https://lnkd.in/eSPJqCXS

Responding to the Prime Minister's announcement of a National Centre for Information Defence, defence secretary Wes Streeting said that means being "better coordinated within government, but also mobilising trusted voices across civil society to help the public to debate and discuss on the basis of fact." We agree. As our Chief Executive Chris Morris says: "We applaud the government's desire to confront information insecurity and hope that civil society organisations like ourselves can help to shape further thinking on how a national unit of this kind can be properly accountable, independent and robust." For a national unit like this to succeed, it needs to draw directly on civil society expertise, alongside government departments, intelligence agencies, law enforcement and technology companies. It should also be integrated with existing national resilience and crisis coordination structures.

At the UN General Assembly, Prime Minister Andy Burnham called for international agreement on global principles and standards for AI transparency and safety, and announced a new National Centre for Information Defence to detect and disrupt AI-enabled disinformation. This is an important step towards information resilience, and echoes a key recommendation in our 2026 Full Fact Report for a national Information Resilience Unit. But the National Centre's proposed focus on hostile states is too narrow. To succeed, it will need to draw directly on civil society expertise, be fully independent and be accountable to Parliament. As our Chief Executive Chris Morris says: "Concerted and coordinated global action to raise the bar on what is acceptable can only help to improve information integrity for us all in the longer term." Common standards for trust and transparency are at the heart of our public interest benchmarking work, where we ask five leading AI chatbots the same questions daily and assess their answers across five dimensions: factuality, transparency, timeliness, consistency and civic responsibility. Find out more about that here: https://lnkd.in/eZUDVBEh

Microsoft AI CEO Mustafa Suleyman told BBC Radio 4's Today programme that AI companies could help rebuild public confidence in their work through the process of independent auditing. We couldn't agree more. We've already embarked on exactly that with our LLM benchmarking work. 🔗 Read more here: https://lnkd.in/eZUDVBEh

Meta has started testing its community notes programme in 16 Spanish-speaking Latin American countries. We don't think that is a good idea. The community notes programme alone is simply not adequate when it comes to reducing the impact of harmful falsehoods on its platforms. Indeed, earlier this year, Meta's own Oversight Board advised it not to expand community notes to countries and contexts that face specific societal challenges. The International Fact-Checking Network, of which we are a member, believes professional fact checking programmes should continue both in the region and worldwide. Read the full statement here: https://lnkd.in/eMqk_5kv