I think we are moving into a more difficult stage of misinformation. The old version of the problem was already serious: false claims, misleading headlines, poor sourcing and people sharing things too quickly. The newer version is more uncomfortable because it looks more convincing and moves through more systems. A fake image can look like evidence from a war zone. An AI answer can sound like a reliable summary. Both can be wrong, and both can travel a long way before anyone slows down enough to check them.
This blog focuses on two types of misinformation because they feel especially important right now. The first is fake or misleading conflict imagery, including AI-generated war videos, altered satellite-style images, and old footage shared as if it shows a current event. The second is AI hallucination inside organisations, where an AI system produces a wrong answer that is then copied into a report, stored in a knowledge base, or used in a decision. They look like different problems, but they share the same basic risk: AI can make unreliable information faster to create, easier to trust, and harder to trace.
Fake conflict images are the most visible threat
Conflict is a perfect environment for misinformation because people want answers before reliable evidence is available. During a fast-moving crisis, a short video or dramatic image can feel like proof. If someone sees smoke over a skyline, a missile launch, or a satellite-style image of damage, they may not pause to ask whether the image is real, whether it is from the right location, or whether it has been shared before. That is why visual misinformation is so powerful. It does not only tell people what to think. It gives them something that appears to show it.
This is already visible in recent conflict coverage. Full Fact has tracked AI-generated images, old videos and false viral claims linked to the Middle East conflict, including AI-generated images supposedly showing US soldiers captured by Iran, digitally created footage of a US bomber in the Persian Gulf, and old footage wrongly shared as a recent attack on Dubai. The Financial Times has also reported on AI fakes being used to create satellite-style war misinformation, which is especially concerning because satellite imagery often carries an automatic sense of authority.
What makes this more difficult is that fake conflict content is not always created for one clear political purpose. Sometimes it is created for attention, followers, engagement, or revenue. X has said it will suspend creators from revenue sharing if they post unlabelled AI-generated war videos, after fake videos linked to the Iran conflict circulated widely on the platform. I find that detail important because it shows how the misinformation economy has changed. The incentive is not always persuasion in the old-fashioned sense. Sometimes the incentive is simply to create something dramatic enough to be shared.
How AI can help detect fake conflict content
AI can help with this problem, but only if we are realistic about what it is doing. It is not looking at an image and discovering ‘the truth’ in one step. It is looking for signals. One system might compare a video frame with older online footage to see whether it has appeared before. Another might look for visual signs of AI generation or editing. Another might check whether the claimed location matches buildings, roads, weather, shadows or landmarks. Another might look at how the post is spreading and whether the same claim is being pushed by several accounts at once.
This is why the idea of a detection pipeline is more useful than the idea of one detection model. The first step is to work out whether the content contains a checkable claim. ‘This is frightening’ is a reaction. ‘This video shows a missile strike in Dubai today’ is a claim. Once the claim is clear, the system can compare the image, caption, account behaviour, timing, location and available evidence. Wired has reported that fake AI content about the Iran conflict has spread widely on X, while also noting that detection remains difficult because tools are limited and the volume of content is high.
This is where ensemble voting can be useful, although the concept is simpler than the phrase sounds. Instead of trusting one model, several checks are allowed to contribute. If the image check is uncertain, but the footage has appeared before, the location does not match and the account network looks suspicious, those signals together make the post a stronger candidate for review. I would not present this as a perfect technical solution. It is better understood as a way of making judgement more structured. A single signal can be wrong, but several independent signals pointing in the same direction give analysts something more defensible to work with.
AI hallucinations are the quieter organisational risk
The second type of misinformation is less visible, but it may affect far more organisations day to day. AI hallucination happens when a system produces information that sounds plausible but is false, unsupported or misleading. NIST describes this as ‘confabulation’, where generative AI systems confidently present erroneous or false content in response to prompts. Most people now know that chatbots can make mistakes, but I think the bigger risk is what happens after the mistake leaves the chat window.
A simple example makes the issue clearer. Imagine an AI assistant reviews a customer record and wrongly concludes that a person should be rejected for a loan. If a human checks the decision and catches the error, the damage may be limited. But if the decision is stored, summarised in a report, reused in a dashboard, or included in a future training dataset, the wrong output starts to behave like a fact. A later employee may not see an AI mistake. They may see a recorded decision. Another model may not see uncertainty. It may see labelled data. That is how misinformation becomes operational.
This is the part I think deserves more attention. Many discussions about AI hallucination focus on the moment of generation, as though the main problem is the model saying something wrong. In a business setting, the more serious problem is often the journey after that. Where does the output go? Does it update a record? Does it trigger an action? Does it feed another system? As AI systems move from answering questions to acting inside workflows, mistakes can become harder to spot and harder to undo.
The bridge between the two risks
Fake conflict images and AI hallucinations may seem like separate subjects, but they follow a similar pattern. In both cases, AI produces or supports information that looks credible. In both cases, the information can be shared, reused and trusted before it has been properly checked. The difference is the route it travels. Fake conflict images move through public platforms, where they are amplified by attention, emotion, algorithms and sometimes money. AI hallucinations move through organisations, where they are amplified by workflows, databases, reports and automated decisions.
That bridge is important for data professionals. Misinformation is not only something that happens ‘out there’ on social media. It can also happen inside the systems we build and use. A fake war image can distort public understanding because people treat it as evidence. A hallucinated AI output can distort organisational decision-making because a system treats it as data. In both cases, the professional task is not simply to be sceptical. It is to create processes that make verification possible before unreliable information becomes trusted.
This is also where I think analysts have a more important role than they may realise. Data professionals already understand messy inputs, weak labels, model limitations, audit trails and the danger of treating outputs as facts without checking how they were produced. Those skills are directly relevant to misinformation. The issue is not only whether AI can detect something fake. The issue is whether an organisation can explain why it trusted a piece of information in the first place.
What data teams should do differently
The first practical change is to separate output from evidence. An AI-generated summary, image label, risk score or recommendation should not be treated as proof by itself. The system should make the supporting evidence visible. For conflict imagery, that might mean showing previous appearances of the image, source history, geolocation checks and verified reporting. For internal AI systems, it might mean showing the documents, records or data fields that support the answer. A confident output without traceable evidence is not a reliable basis for a serious decision.
The second change is to build review into the workflow at the points where harm could happen. This does not mean every output needs to be manually checked. That would be unrealistic and, in many settings, impossible. It means high-impact outputs should not move automatically from suggestion to action. If an AI system is making or influencing decisions about money, employment, health, education, public communication or customer access, there should be clear points where uncertainty is escalated, evidence is checked and a person can intervene.
The third change is to keep a record of what happened. This sounds simple, but it is often where systems fail. If a false claim, fake image or hallucinated answer enters a process, the team should be able to trace where it came from, where it was used, and what needs correcting. Without that trail, organisations end up fixing the visible error while leaving the same bad assumption hidden somewhere else. In my view, traceability is one of the most practical safeguards we have, because it turns misinformation from a vague problem into something that can be investigated.
Why this matters for data analysts
For data analysts, this topic is a useful reminder that modern analytics is not only about producing insight. It is also about protecting the reliability of information as it moves through systems. That may sound like governance language, but in practice it is very concrete. It means checking sources, understanding how data was created, knowing when a model output is uncertain, and making sure decisions can be explained later.
AI will help detect misinformation because the volume of content is too large for humans to manage alone. It can compare images, cluster repeated claims, identify unusual sharing patterns and flag outputs that need review. But AI can also create convincing errors. That is the tension. The same technology can help us verify information and help misinformation spread faster.
I do not think the answer is to become anti-AI or to treat every output as suspicious by default. The more useful position is to be professionally cautious. AI can be valuable, but only when its outputs are supported by evidence, checked in the right places and traceable through the systems that use them. In a world where fake conflict visuals can shape public understanding and AI hallucinations can become part of organisational knowledge, that caution is not a weakness. It is becoming part of responsible analytics practice.
Want to learn more? IoA membership not only recognises your professional status it also provides access to 100s of extensive, on-demand training courses. We recommend our Governance training pathway, The Data Governance Practitioner and The Co-Pilot Expert courses, all included for IoA Members. Check out the right IoA membership for you here:Institute of Analytics (IoA) | Data Careers & Membership
