Working with AI and spotting AI-generated text

Generative artificial intelligence (AI) tools like ChatGPT can be used to draft text, summarise documents and find information. When used carefully, they can enhance readability and support analysis. However, they can introduce errors, bias or misinformation.

This article sets out practical principles for using AI effectively in work that involves research and information, such as work done by people in Parliament. It is part of the Commons Library’s good information toolkit, which reviews the strategies and criteria for spotting misinformation and working out what information can be relied upon.

AI should be treated as an assistant, not an authority. It can support research and analysis, but it cannot replace professional judgement, subject expertise or trusted information sources. This is why many experts recommend always keeping a human in the loop – AI users should check everything generated by an AI tool and use their own judgement. The Commons Library's editorial policy explains how our researchers may use AI in line with our commitments to impartiality and accuracy.

MPs and their staff can always contact the House of Commons Library to check generated material with an expert.

What AI tools can and cannot do

Generative AI tools work by identifying patterns in large volumes of existing material. They do not understand issues in the same way people do. And they do not return information in the way an internet search or a request to an Alexa-like device does.

AI tools can be useful for:

• brainstorming ideas, for example developing structures for reports or suggesting ways to respond to emails
• summarising long texts or documents
• generating questions or lines of enquiry
• suggesting alternative wording, for example making a sentence easier to understand or more formal
• helping to explain complex topics succinctly
• summarising meeting transcripts

For all these things, AI can be more helpful if the user already knows about the subject matter they are asking the tool to help with. This is because they can spot mistakes, identify areas the AI has missed or mis-represented and add nuance to the final version.

AI is less reliable for:

• providing definitive factual answers
• resolving contested or politically sensitive issues
• interpreting law, policy or parliamentary procedure
• producing impartial analysis without careful human oversight

Hallucinations

AI tools can produce information that sounds plausible but is incorrect, incomplete or misleading. This is sometimes described as “hallucination”. These can appear as incorrect facts, made-up citations or misleading arguments. They happen unexpectedly, sometimes in response to simple questions and sometimes when the prompts are more complex.

The best guard against hallucinations from AI is to check everything generated carefully, ideally with an expert. It is often useful to precisely identify each fact or claim in the generated material and then verify each one with an independent source.

Using AI as part of a workflow

When using AI for research and information, the following workflow prioritises accuracy, impartiality and accountability:

1. Define the task clearly and decide whether AI is appropriate
2. Use AI to do the task
3. Check all factual claims against reputable, authoritative sources
4. Review outputs with someone who has subject expertise
5. Edit, contextualise and take responsibility for the final content

Always ask AI to provide sources and always check that they are trustworthy by clicking on any links the AI provides. For more on this, see the good information toolkit article on evaluating sources.

Asking better questions of AI

The quality of AI output depends heavily on the quality of the instructions (or “prompt”) that the user puts into the tool. Vague prompts tend to produce vague results. All AI tools are already prompted by their programmers to respond in a certain way with “system prompts”. Users must then give additional instructions so to get back information or material that is useful to them in a format appropriate to their needs.

There are lots of methodologies for designing effective prompts. One example, described by Microsoft, is the “GCSE” framework:

• G – Goal: Clearly define what you want the AI to do (for example, "Summarise this document in 600 words").
• C – Context: Provide background information to guide the response, such as who the audience is, why you need it, and what the subject is.
• S – Source: Tell the AI where to look for information (for example, by specifying websites or organisations).
• E – Expectation: Define the format, tone, and style (for example, no bullet points, all in British English table, use impartial language).

Checking AI‑generated information

Because AI tools are known to hallucinate, make errors, and misrepresent sources, it is important to fact-check everything it generates. Users should treat every factual statement produced by AI as unverified until they have thoroughly checked it. For more on fact-checking, see the good information toolkit article How to check facts.

A useful workflow would be:

1. Identify a “fact” or claim made by the AI, such dates, figures and quotations
2. Confirm facts with primary or authoritative secondary sources. This can involve talking to an expert (such as the Commons Library) or doing some independent research
3. Test whether claims are current or out of date
4. Seek independent confirmation from more than one source

Reputable sources might include official statistics from the Office for National Statistics, primary legislation, government departments, recognised regulators, peer‑reviewed research, and briefings published by the House of Commons Library.

Spotting AI‑generated text

AI-generated text requires little effort to produce; someone can generate and share pages of text without ever reading it. AI-generated text produced this way is more likely to contain errors and is less likely to reflect anyone’s judgment or opinion about an issue than text that was carefully written by a real person. It is therefore useful to be able to spot AI-generated text, to assess how likely it is the text is inaccurate and whether it is likely to reflect genuine thought.

But as the tools become more effective at generating useful text and as AI-generated material becomes increasingly widespread, spotting AI-generated material is increasingly difficult.

Indicators of possibly AI-generated text include:

• Use of horizontal lines between clauses or sections
• Lots of bulleted lists, and often nested lists (with sub-bullets)
• American spelling (“color”, “emphasize”) in documents by people or organisations that generally use British English
• Lots of similar ideas or concepts presented in slightly different ways
• Language that is fluent but vague, with few concrete details
• Assertions without references or evidence
• Lots of references or citations to obscure sources, or citations that are loosely connected to the point
• Inconsistent levels of detail or sudden shifts in tone
• Errors in dates, names or institutional roles

These indicators are not definitive. Human-written content can share some of these features, and AI‑generated content can be edited to disguise them. AI detection tools do exist, but they are unreliable and should not be treated as conclusive.

The most robust approach is to assess the content itself. Ask whether claims are supported by evidence and whether sources are reputable and verifiable.

AI generated material in MPs’ casework

MPs’ caseworkers are increasingly receiving enquiries from constituents that appear to have been generated or partially drafted by AI tools. Often these appear as long lists of similar or technical questions and requests framed in legal or procedural language. While these messages often originate from genuine constituent concerns, they can be harder to interpret and more time‑consuming to respond to.

These casework requests can include inaccurate but convincing explanations of law or policy, misused legal terminology, or unrealistic “asks” that would be hard to respond to fully. They can also mask the underlying issue a constituent is facing, making it more difficult to identify what support an MP’s office can usefully provide.

There are several approaches that caseworkers can use to deal with enquires that might have been AI‑generated. One is called “de-prompting”, which is an approach to moving from long, generic or overly complex message to a clear understanding of what it would be most helpful to provide to the constituent.

Caseworkers can use an AI tool yourself to carry out the steps suggested below or can work through them manually.

De-prompting an AI generated query

1. Identify what kind of query this is. This is about trying to understand the core issue. the crux of the question, and reverse-engineering the prompt that might have generated the query. Try asking an AI tool: “What question or instruction might have produced this AI-generated query?”.
2. Deconstruct the query. This helps narrow down topics, see which claims are treated as given and find out what’s missing. Try asking: “Break this query down into: distinct sub-questions; assumptions it relies on; issues it does not mention but may imply.”
3. Identify the underlying concerns. This is not a diagnosis but instead helps understand what the constituent thinks and ask questions. Try asking: “What underlying concerns, motivations, or anxieties might lead someone to ask this question? List several potential interpretations.”
4. Generate follow-ups. This step is to help formulate follow-up questions to ask the constituent to clarify their requests or work out how best to help them. Try asking: “What clarification questions would a constituency office need to ask before responding properly? Prioritise questions that: establish personal impact; clarify what outcome they want; check key assumptions.”

In practice, de‑prompting allows caseworkers to prioritise what matters, reduce duplication and return to the constituent with a clearer, more focused set of questions, improving both response quality and efficiency.

As with all casework, asking the constituent about their personal circumstances and how they led to the enquiry is a useful way to reframe enquires so they are more direct.

MPs and their staff can always contact the House of Commons Library to help with questions, verify claims or to check the accuracy of references to legislation.

Understanding bias in AI

AI systems reflect the data they are trained on and the choices made by their developers, which can result in generated material that is biased towards a particular point of view or group of people.

The prompts that a user writes to generate material can also have a big influence on how biased the material they generate can seem. AI-generated material can sometimes omit perspectives or privilege commonly repeated views.

This is particularly important when impartiality is essential, such as in the work of Parliament.

To mitigate against bias, AI users should actively consider whose perspectives are included in an AI-generated response and whose are missing. They can also add text to their prompts to instruct the AI tool to avoid bias. This could be simple, such as telling it to avoid bias or include a diverse range of perspectives; or more subtle, such as telling them to adopt the persona of someone who answers questions slowly and rationally.

Data protection and security

AI tools should not be used to process sensitive, personal or confidential information. It is not always clear how AI tools store, use or share the data that is inputted into them. There is a risk that inputted data might be used to train future models, so information included in a prompt could be used to generate material for other users in the future.