Using AI well on an open problem

1. Orient with AI, then verify without it

At the start, when you know very little, an AI model is good at orientation: it can help orient you to the terminology, the main actors, the debates, what to read in your area of interest. That is relatively low risk, because you are asking it to help you find things and building a list of things for you to check rather than collecting facts.

Once you need numbers and citations, you might use AI like a different tool. It can tell you where evidence probably lives. Finding it and checking it is still something you need to do. Without even realizing it, you might run into problems on your project if you stop after the first stage.

Ask yourself: am I using this to orient, or as a source? Have I done the second step?

2. Make it argue with you

If you ask AI tools what to work on, you won’t get very good answers. If you ask them general things, you’ll likely get general answers. Ask about problems generally and you might get topics or a list of symptoms. To get at the specifics of a system or a problem, you have to push and iterate between using tools and traditional research methods. Three prompts that may help:

• Who benefits from this staying the way it is?
• What are the resources that support this system, where do they come from, and how are they allocated?
• What would have to be true for the obvious fix to have failed?
• Which part of my thinking has the least evidence behind it?
• Ask the model after it gives an answer what is it unsure about? What is weak or thin data and these give you clues for things you may want to check first.

3. Ask where the information came from

Every type of data source is strong at some things and weak at other things.

• NGO reports: good on severity and ground detail; they advocate, and they often emphasise severity.
• Government statistics: good on scale and official baselines; slow, and often politically shaped.
• Journalism: vivid and recent; often anecdotal or small sample sizes.
• Academic papers: good on mechanism and causal evidence; narrow, and often dated can be small sample sizes.
• An AI answer with no named source: reads well; but there may be nothing behind it.

To check a number: ask the tool for the source, go find it yourself, and confirm the number is there and means what the model said.

No source at all is a signal that the tool may be pattern-matching rather than reporting.

Ask yourself: what kind of source is this, and what is it weak at?

4. Triangulate across types, not just across sources

Choose source types so their blind spots do not overlap: an NGO report for severity, a government dataset for scale, a journalist for texture, an academic paper for mechanism.

If they all point the same way, you can have more confidence in the claim. If they disagree, the disagreement is worth exploring.

Two AI models agreeing is a thin form of triangulation. It is the same source type. You can ask AI to help you spot which type you are missing, rather than stand in for one.

Ask yourself: which type am I missing, and where do my sources disagree?

5. Thin data tells you where to look closer

Data often is unavailable in a few predictable ways. It was never collected. It was tried and failed and never written up. Or it is known in the field but not published, because it is sensitive, would put someone at risk, or is a commercial secret.

An AI model will fill all three gaps with something plausible. If the answer on a thin topic reads smoothly, be suspicious and look closer.

All of this so far is intended to help you prepare for time in the field with users or experts doing interviews. This research will give you clues for what to ask a practitioner: what has been tried, what do people in the field know that is not written down, and why is it not?

The second kind of gap is the most useful one. People who tried and failed know more than most reports you will find from your desk.

6. Do not spend an expert’s hour on things you could have looked up

Every question you ask that your desk research could have answered is a better question you won’t have time or get to ask. Do the triangulation first. Walk in knowing the numbers and the debates. Then you can make the most of the in-person learning opportunity with people with deep lived experience and contextual knowledge of the problem area, what’s been tried and failed, what counterintuitive (to novices) things would it actually take to make progress, and so on.

Ask yourself: where is my knowledge weak and which of my questions could only this person answer?

7. Some parts of the system you can move this term and many you cannot

In a stakeholder map, every arrow should be a flow you can name and source. Many teams will map systems reasonably well, but systemic problems can be too big for class projects to tackle. What teams skip is asking which parts of the map they can actually touch. Mental models and big policy shifts take decades. Practices, information flows and specific relationships can change faster. If your effort is a term, that narrows your actionable system map. It is fine to say: this is the real leverage point, but I cannot reach it from here, so here is the smaller one I chose to work on in this timeframe.

Ask yourself: what can I move from where I stand, in the time I have?

Four things you might do as you kick off your project

• Know the difference between an interest area and a problem. Write your problem in one sentence a practitioner would recognise. If they would say “well, sort of,” you still have a topic (not yet a defined problem).
• Find someone who already tried. Most of these problems have a pilot that quietly ran out of money or interest, and that person knows more than most reports. Try reaching them on LinkedIn.
• Every project needs resources. Knowing what they are and how they are allocated shows you where a system has been vulnerable to scarcity, and that can suggest a better approach.
• Start a list now of what surprises you as your learning evolves and keep it up as you work. Do not wait until the end. Your understanding will change and you will forget the turning points.

If your course asks for an AI-use statement

If an argument in your paper would not exist without the model’s wording, it is generated. If it exists because the model sent you somewhere, it is assisted.

How the briefs here fit

Each write-up on this site is a head start on triangulation, not a substitute for it. This system gives you figures with named sources, what has been tried and why it stalled, and where a team might start. Use it to orient, know its limitations, and go and check to verify as you prepare to make the most of your time in the field learning alongside real users and experts.