Doctoral proposal · 2026 / Human–AI collaboration for sustainability decision-making
Answers arrive faster than judgement.
Half of what a model hands you may be wrong, or right somewhere else, for someone else, in another watershed. This is a proposal to study how young people build that judgement — not in a classroom, but in the ordinary moments where they and the world decide something together.
01 — THE GAP, AND THE GAP UNDER IT
The literature is crowded. The depth is not.
On 21 July 2025, Nature Sustainability published a systematic analysis of 792 papers applying artificial intelligence to Sustainable Development Goal research.[1] What it surfaces is not scarcity — AI is already everywhere in sustainability work. What it surfaces is shallowness.
We identify a critical gap: only a few studies combine advanced AI applications with deep sustainability expertise.Artificial intelligence in sustainable development research — Nature Sustainability, 2025
The authors push further: sustainability must "strike a balance between contextualization and generalizability to provide tangible knowledge that will lead to responsible change."[1] Context is where knowledge lives. Generalization is how it travels. Most AI pipelines are engineered to optimise the second and quietly discard the first.
That is a gap between two expert communities. This proposal names the gap one floor below them: between the people who build and publish the models, and the enormous number of young people already making small, real decisions inside those models' outputs — what to cite, what to trust, what to build, what to plant, where to work, what to stop doing. The 2030 Agenda is carried by 17 interlinked goals.[3] It is also carried by millions of unremarkable daily choices that almost nobody studies.
Papers systematically analysed [1]
Interlinked SDGs, adopted 2015 [3]
Of SDG targets reported on track, UN 2024 [2]
Studies bridging advanced AI & deep sustainability expertise [1]
02 — THE WHEEL OF QUESTIONS
Some questions have to be put on the table.
Otherwise they stay buried under the sheer volume of everything else — the feed, the deadline, the next fluent paragraph. Turn the wheel. Each face is a question this project refuses to answer too quickly.
Six faces. You will not see the others unless you turn it — which is, more or less, the point.
03 — WHAT WOULD ACTUALLY BE STUDIED
Three strands, one longitudinal cohort.
Not a survey of attitudes. A study of traces: what a young researcher actually did, in a real sustainability task, at the moment an AI output appeared.
Trace
Capture the moment of acceptance. Micro-logs and think-aloud records of every point where a learner accepts, edits, rejects or silently absorbs a machine output inside a live SDG-linked problem.
Output — an open corpus of decision traces
Calibrate
Measure the distance between confidence and correctness, over months. Accuracy is a snapshot; calibration is a skill. The research claim is that calibration, not recall, is the transferable competence.
Output — a calibration curve per learner, per domain
Narrate
Follow the story a learner tells about their link to a place, a system, a community — and test whether that narrative survives contact with a generalising model, or is quietly overwritten by it.
Output — longitudinal narrative analysis
- H1AI-assisted learners converge faster on fluent answers and slower on locally valid ones.
- H2Calibration under uncertainty predicts decision quality better than subject-matter test scores.
- H3Literacy formed inside real practice transfers across domains; literacy taught only as instruction does not.
"AI literacy" may be a moving target that dissolves as interfaces improve — study it and you risk measuring an artefact of 2025 tooling. The answer has to be structural: measure decisions and calibration, not tools. Interfaces churn every eighteen months. The act of deciding under an unreliable, confident source is at least as old as the library.
04 — WHY NOW, AND WHY THIS COHORT
The bottleneck moved. Nobody re-trained for it.
Generation is now nearly free. Verification is not. Whoever learns to verify cheaply and locally is the one who effectively decides — regardless of job title. That is the shift, and it is happening to students before it happens to institutions.
The urgency is not that AI will replace judgement. It is that the volume of plausible material now exceeds any individual's capacity to check it, at exactly the moment when sustainability decisions have the least room for error: a reported 17% of SDG targets on track as of 2024[2], with the 2030 deadline five years out. Fluency is scaling faster than the feedback loops that correct it.
And this cohort is observable only once. Students entering research between 2025 and 2030 are the last group who will remember both sides of the shift. A longitudinal baseline not taken now cannot be taken later — which makes this a narrow, dated, and therefore fundable window.
05 — THE FUNDING RUNWAY
Programs that already fund exactly this shape of work.
Human-centred AI, responsible AI and graduate fellowship programs, each with an open application page. Links as supplied in the source brief.