Research Topic Artificial Intelligence
Artificial Intelligence
Artificial intelligence research examining how AI systems are built, evaluated and used, and how they reshape work, markets and decisions.
We examine how AI systems are built, evaluated and used, with attention to the evidence behind their performance and the consequences that follow deployment.
Beyond the model
Artificial intelligence is a system, not a single piece of software.
A model sits inside a wider environment of data, objectives, interfaces, people and operating rules. Performance depends on how those parts work together—not only on the model selected.
That wider view matters because a technically accurate output can still be poorly timed, misunderstood or used for a decision it was never designed to support.
Does the complete system help the intended user make a better decision under real conditions?
The task, user and intended outcome.
Sources, coverage, quality and permissions.
Users, reviewers and affected groups.
Workflow, incentives and operating constraints.
From idea to use
The work continues after a model is trained.
Useful AI requires a clear problem, suitable data, disciplined evaluation and ongoing review as users, inputs and operating conditions change.
Frame the problem
Define the task, the user and what a useful outcome would mean in practice.
Prepare the data
Examine relevance, quality, coverage, permissions and the gaps the dataset may contain.
Develop or adapt
Select, train, fine-tune or connect a model to the information and tools required for the task.
Evaluate
Test performance, limitations and failure modes against realistic cases—not only a headline metric.
Deploy with controls
Place the system inside a workflow with clear roles, review points and ways to challenge an output.
Monitor and revise
Watch for changing data, behaviour and effects, then update or stop the system when needed.
Capabilities and limits
AI can extend analysis. It does not remove the need for judgement.
The useful question is not whether a product is called AI, but whether it performs reliably in the setting where it will be used.
Classify and predict
Detect patterns in historical data and estimate categories, probabilities or future outcomes.
Generate and transform
Produce, summarise, translate or reorganise text, images, audio and code.
Detect and monitor
Identify unusual patterns, quality problems or changes that deserve closer attention.
Recommend and assist
Rank options, retrieve relevant information or support a person working through a complex task.
Evaluating an AI system
Start with the decision, then examine the model.
No single benchmark answers every practical question. Evaluation should reflect the task, the people affected and what happens when the system is wrong.
Purpose
Is the task clearly defined, and is AI suitable for it?
Look for: intended use, excluded uses, decision ownerData
Does the evidence represent the conditions the system will face?
Look for: coverage, quality, provenance, material gapsPerformance
Which errors occur, for whom, and under which conditions?
Look for: baselines, subgroup results, realistic test casesReliability
How does the system respond to unfamiliar, incomplete or changing inputs?
Look for: stress tests, drift, uncertainty, fallback behaviourHuman role
Can people understand, review and challenge the output in time?
Look for: authority, review points, escalation, trainingEffects
What changes after deployment, including effects beyond the target metric?
Look for: behaviour, access, workload, incentives, unintended outcomesWhere change appears
AI reaches the economy through tasks, products and decisions.
The effects are not uniform. The same capability may save time in one setting, create new review work in another or be unsuitable where errors carry greater consequences.
Industrial operations
Quality inspection, predictive maintenance, process monitoring and production planning.
Supply chains and markets
Demand signals, inventory decisions, route planning and market intelligence.
Knowledge work
Search, drafting, analysis, software development and document-heavy workflows.
Digital products
Recommendation, personalisation, moderation, assistance and automated interaction.
Science and materials
Literature analysis, candidate screening, experiment design and interpretation of complex data.
Institutional decisions
Decision support where transparency, contestability and accountability may be as important as speed.
Proportionate oversight
The greater the consequence, the stronger the control should be.
AI used to reorganise a draft is not the same as AI used to recommend an operational or consequential decision. Review, documentation and escalation should reflect that difference.
Routine assistance
Clear user review and an easy way to correct the result.
Operational support
Defined thresholds, monitoring, records and escalation paths.
Consequential decisions
Strong validation, accountable human authority and a meaningful route to challenge.
Questions we examine
Good AI research connects technical performance with real-world effects.
Open each question for a concise explanation.
01Where does AI genuinely improve a task?
AI is most useful when it saves time or improves quality without creating more checking, correction or hidden work elsewhere.
02Which tests show whether a model will remain useful outside a controlled benchmark?
The strongest tests use realistic tasks, unfamiliar inputs and failure cases that reflect how the system will actually be used.
03How do data, computing requirements and incentives shape access?
Access depends on more than the model. Suitable data, computing resources, cost and a clear reason for use all matter.
04What changes in productivity, skills and responsibility?
AI can speed up parts of a job while shifting skills and responsibility towards reviewing, judging and correcting its output.
05Which controls help people respond when a system fails?
People need to know who is responsible and how to review, challenge or correct an output when something goes wrong.
Continue exploring