What is artificial intelligence?¶
AI systems can write a summary, spot a pattern or recommend what to watch next. They take input and infer how to produce an output: put simply, they work out a result using what they receive and how they have been built or trained. For you, “AI was used” is a useful starting point. The next questions are which system was involved, what went into it, what came out, and what happened next?
A simple way to think about AI¶
The term AI covers many different systems. They do not all work in the same way, but this is a useful starting point:
The system's objectives are what it is intended to achieve. Inferring an output does not mean it has proved the answer or understands the situation as a person would. A convincing answer can still be wrong.
This follows the broad idea in the OECD definition of an AI system: a machine-based system that takes input and infers how to generate outputs such as predictions, content, recommendations or decisions. NIST's AI Risk Management Framework uses the same broad idea. The formal definitions help keep the terminology consistent; the diagram above gives you a practical way to think about the system.
A familiar example¶
Suppose Jane opens an AI chat application, such as ChatGPT, and asks it to summarise a public consultation document.
Jane now has a summary she can compare with the document. The system has processed the supplied material using patterns and relationships represented in the model and the context available to the application — the information it can draw on when producing its response.
That same input → system → output relationship appears in other AI uses:
A system produces text, an image, audio, video or code from supplied instructions or material.
A system evaluates input and returns a predicted category, score or other classification.
A system uses available information to rank or suggest options.
An AI output may feed into another process that makes a decision or performs an action, with varying levels of human involvement.
The phrase artificial intelligence tells you the broad type of technology involved. You still need to identify the model, application, provider — the organisation supplying the service — and any data sources or process used to reach a decision.
Why that matters to an investigation¶
Suppose you are told:
“The suspect used AI to write the messages.”
That is a useful line of enquiry. Now you have something to test.
A screenshot, device record or material from a provider may help identify an AI application worth examining. A visible prompt — the request or instructions supplied to the system — and its response may also help explain how particular content was produced. The next questions are more specific:
- Which application or service was used?
- What input or material was supplied?
- What output did the system produce?
- Is the model or version known?
- What other information or source material could the system draw on?
- Which account, session, device or person was associated with the relevant activity? A session is a period of interaction with the service.
- Did the output remain a suggestion, or did another system or person act on it?
A more precise account of what technology was involved, what input and output should be looked for, and which application, device, account or provider may hold relevant records.
It does not by itself prove which model produced the material, that the output was accurate, who operated the system, why they used it or who is responsible for the resulting conduct.
Those questions give the enquiry somewhere to go. Once you identify the system, input, output and what happened next, you have concrete things to preserve, compare and check against other evidence.
AI is not one particular kind of product¶
AI may be part of a chat service, an image or audio tool, a search or recommendation feature, business software, security software or a larger automated process. Some systems generate content. Others sort information into categories, rank options, predict outcomes or help make decisions. Some need a person to direct each step; others can use tools or trigger actions with less immediate human involvement.
The label AI is therefore similar to labels such as “cloud service” or “online account”: useful for getting your bearings, but you need the detail to investigate what happened.
