A company is built to align people and ship. A university lab can convene people who would rarely fit one reporting line — poets, fashion scholars, designers, behavioral scientists, clinicians, statisticians, and software engineers — and give their disagreements a scientific process. The point is not eclectic brainstorming. It is independent evidence for the customer and business decision you need to make next.
Your team knows the product, market, and constraints. The lab is useful because it can test the attractive idea without needing it to win. A null or negative result can save more value than a flattering story.
We state the null hypothesis, primary outcome, meaningful effect, sampling plan, exclusions, and decision rule before the result. We estimate uncertainty, correct multiple tests, try to falsify the claim, and replicate before scale.
The unusual team is deliberately temporary and problem-specific. Your product organization keeps moving; the lab brings in the exact intellectual friction a consequential decision needs, then turns it into evidence your organization can use.
We do not sell predetermined conclusions or promise growth. We decide in advance what result would support the intervention, what would count as too small to matter, and what would tell us to stop. Positive, null, and negative results can all strengthen a real business decision.
That is a feature. A poet may care about what an answer does to dignity; a clinician about safe action; a designer about who is excluded; an engineer about reliability and cost. Revenue, adoption, and retention still matter because your business must decide — but they are measured alongside the human consequence, not instead of it.
These people may disagree about what the problem even is. We do not force an artificial consensus. We convert competing readings into rival hypotheses, and the research contract — not seniority or presentation skill — decides what survives.
Sees voice, sequence, metaphor, omission, and the implied reader — including what the answer quietly asks a person to believe or do.
Sees status, belonging, taste, cultural signals, and who feels addressed or excluded before a conventional usability metric notices.
Sees hierarchy, affordance, cognitive load, modality, and the moment an intelligent answer becomes difficult to use.
Sees confounds, stress, motivation, appropriate action, harm, and the gap between what people say and what they actually do.
Asks whether the effect exceeds noise, whether the analysis was chosen after seeing the answer, and whether one lucky comparison is being mistaken for a finding.
Makes the intervention reproducible, isolates model and interface changes, and keeps latency, reliability, tokens, and serving cost inside the decision.
Close reading, narrative theory, cultural analysis, fashion studies, and design research reveal mechanisms a benchmark may miss. We operationalize the relevant insight as a controlled intervention and measure what real people understand, choose, pay for, use, remember, or safely act on. Not every human experience should be reduced to a number; any product claim made from it should survive an empirical test.
The method is familiar science, applied to human–LLM interaction and business decisions. A p-value is one diagnostic — never a truth machine. Effect size, uncertainty, design quality, practical importance, and replication decide whether the result deserves action.
Interviews, close reading, field observation, cultural analysis, and product data locate the human mechanism worth studying.
We define the null hypothesis, rival explanation, primary outcome, and minimum effect that would be large enough to change the business decision.
Population, sample, exclusions, stopping rule, intervention, comparator, outcomes, and confirmatory analysis are fixed before results can tempt us to move the goalposts.
We control the model, prompt, interface, and exposure where possible; randomize or match conditions; document the deployment context; and include ethics and risk in the design.
We report effect sizes, intervals, p-values where appropriate, practical thresholds, and uncertainty. A small p-value does not prove that an effect is true or important.
When several confirmatory hypotheses are tested, procedures such as Holm correction control the family-wise error rate instead of rewarding the luckiest result.
A promising effect is replicated across a new sample, language, model, workflow, or setting before it becomes a general claim or a large deployment.
Trying to recreate this mix permanently inside one reporting structure either neutralizes the dissent or burdens the product organization with people it was never designed to manage. A finite university partnership preserves both strengths: your team keeps its speed and judgment; the lab supplies independent friction, method, and evidence.
Universities have long built centres around problems that no single discipline can own. These are precedents, not affiliations: each demonstrates a part of the method we adapt for human-centric LLM research with measurable customer and business outcomes.
An interdisciplinary AI institute spanning Stanford’s seven schools, guided by human impact and the aim of augmenting people.
Visit the centreAn interdisciplinary lab built around unconventional mixing of seemingly disparate areas — including AI, art, design, health, robotics, music, and human–machine interaction.
Visit the centreBrings writers, artists, and other creative thinkers together with scientists, engineers, and technologists, connecting human narratives to scientific questions.
Visit the centreStudies people, technology, and society through the scientific traditions of computer science and human sciences, at the intersection of engineering, behavioral science, and design.
Visit the centreA research collective that applies computational criticism to literature — a direct precedent for humanistic questions investigated with reproducible analytical methods.
Visit the centreTreats computation, data science, and the humanities as reciprocal partners, combining critical interpretation with data-intensive research and software engineering.
Visit the centreBrings philosophy, humanities, and STEM researchers together with leaders in technology, business, government, and civil society to provide independent perspectives on AI.
Visit the centreA long-running interdisciplinary HCI lab connecting information studies, computer science, psychology, education, English, engineering, journalism, and other fields.
Visit the centreA knowledge-design lab rooted in arts and humanities, working across code and narrative, data and ethics, networks and archives.
Visit the centreConnects fashion expertise, brands, and technology partners through proofs of concept and prototypes, then returns that learning to the university.
Visit the centreA people-centred inclusive-design research centre working with business, government, academia, and the third sector across design, psychology, neuroscience, health, and systems thinking.
Visit the centreThis is the compact research contract behind every sponsored study. It protects your team from an impressive result that answers the wrong question — and protects the lab from quietly redefining success after the data arrive.
Start with one consequential customer or business decision. We will assemble the smallest genuinely heterogeneous team that can challenge it, then design a study whose result remains useful even when the original idea does not survive.