AI has the power to transform investigations. It can automate repetitive tasks, produce insights rapidly and free investigators to focus on analysis. But this potential comes with a critical responsibility: ensuring AI-generated insights are accurate, not just plausible.
AI systems generate outputs based on pattern recognition and probability. Without proper oversight, they can produce findings that sound credible but are factually wrong – known as ‘hallucinations’. In high-stakes investigative work, where decisions must be defensible in court or to regulators, plausible-but-wrong insights aren’t just unhelpful, they’re dangerous.
This risk is particularly acute in Open Source Intelligence (OSINT), where investigators work with vast, unstructured and publicly available datasets from across the internet. While AI might surface insights at remarkable speed, it can also invent information, miss critical context, or generate compelling narratives based on false data.
However, the solution isn’t to reject AI, but to ensure human investigators remain at the centre. AI should function as a collaborative partner, handling routine tasks while investigators apply the contextual judgement and ethical oversight that machines cannot replicate.
Credibly flawed insights
AI excels at collecting and processing data, presenting insights in a digestible format. What it’s not so good at is making judgement calls. The news is full of examples of AI models hallucinating, whether that’s reporting incorrect information in news headlines or reflecting biases present in training data.
One of the real dangers is that generative AI can present inaccurate information as fact with such confidence that people take it as truth. As they write like humans, people can assume they have human intelligence. This means insights appear highly credible, even when they’re partially or wholly fictitious.
Why do they do this? A crucial reason is that training data can be flawed. Online data is incredibly vast, and a lot of it is ‘nonsense’. Reddit, for example, is the most-cited source across all major AI platforms and any inaccurate content on forum chats can be taken as fact. That means there will be non-factual content in training data.
During training, models learn to make predictions by spotting patterns in the data, so if this data is inaccurate, incomplete or biased, these patterns will be incorrect. Most notably, in an attempt to fill in the blanks in any information or context they might not have, they still generate an answer using these patterns rather than stating that they don’t know.
Of course, this presents a real risk to investigations, where using genuine information is crucial to accurately following leads and correctly looking into any persons or entities of interest. If the AI hallucinates and investigators take its insights at face value, then they could end up conducting investigations into the wrong people.
But this doesn’t mean that AI should never be used. Instead, it’s a question of correct application. Gaps in context are exactly where human expertise and insights come into play.
Turning AI from a risk to an asset
Human oversight is always needed to maintain data accuracy and ensure output quality of AI solutions. Wrongful investigation outcomes can’t be blamed on AI – accountability lies with the humans overseeing AI and the consequence for breaching regulations will fall on an organisation, not on a machine.
So, what are some of the safeguards that make AI an investigator’s asset?
Explainability is a vital feature for enterprise and investigative AI use. Many large, commercially available answer engines are ‘black-box’ systems. The problem with using these systems for OSINT is that their internal workings are unknown, so it’s not possible to understand how or why they have produced their outputs. And if you can’t see what insights or information the model used, investigators can’t audit AI decision-making and verify answers.
Solutions that are purpose-built for investigators, on the other hand, put additional emphasis and focus on repeatability, auditability and sourcing. They can provide reasoning for how they arrive at conclusions, and therefore enable investigators and regulators to understand the steps they’ve taken in generating information. With these solutions in place, it’s then a case of establishing clear governance for what tasks AI can carry out and where human oversight is essential.
For example, AI can automate repetitive activities like data collection and then investigators can apply contextual analysis on these insights and check their accuracy. What’s more, using prompt frameworks like AUTOMAT can enable teams to set their own risk parameters and ensure outputs are reliable, explainable and compliant.
Harnessing agentic AI in OSINT
Agentic AI refers to autonomous, goal-oriented AI systems which can independently reason, plan and execute multi-step tasks with minimal supervision. An agentic AI system consists of AI agents which can autonomously collect data from multiple sources, assess their options and then take the necessary steps to meet a defined goal or objective. Essentially, this is how an investigator works. Crucially, they’re flexible and adaptable, with their capabilities allowing them to react to situations in real time.
When it comes to OSINT, this means AI agents could independently carry out tasks like entity resolution (where they find and link different online data and records to the same real-world identity). Investigators can then use this consolidated information to inform and carry out their investigations, with more time to detect patterns in data.
But while the autonomy and decision-making of AI agents is their major benefit, without robust governance and oversight of their use, these capabilities are also their major risk.
That’s why human-in-the-loop principles are so important, as they ensure investigators can see what the system is doing, why, and intervene at key stages within the agentic AI workflow. The key to agentic AI’s effective use is to utilise incredibly specific prompts and contain the data and tools it can access.
Matching other industries
Due to the sensitivity of information and high-stakes nature of investigations, the idea of effective machine-human collaboration for OSINT might seem far-fetched. But the aviation industry is a great example of this partnership working and evolving over time. The sector automates much of the flying experience to reduce errors and enable pilots to focus on high-level tasks and responsibilities.
We can see the use of AI and agentic AI for OSINT in a similar way. The technology can independently automate key repetitive tasks like data collection to empower investigators to focus on more critical activities. Yet human oversight is fundamental to mitigating risk. Investigators mustn’t take plausible AI-generated insights at face value or use the technology for tasks that require human expertise – and they definitely shouldn’t give AI full autonomy. You wouldn’t feel comfortable flying in a plane with no human pilot.
That’s why AI won’t replace investigators – but it can enhance their work significantly .
