Bast

Evidence library

The research behind how we build.

If we ask AI to show its work, we should show ours. This library collects published research on three questions: whether AI that shows its work keeps people thinking instead of handing their judgment over, whether easing the load on caregivers and care workers helps patients, and who pays for caregiver support.

Every entry links to its source and says how strong the evidence is and what it can’t tell you. We kept the studies that cut against us. The point of AI is to augment the people in the room, and that starts with being honest about what the evidence does and doesn’t show.

Last updated · Refreshed weekly · 43 studies

Five of the strongest findings

Each of these comes from a randomized study or a meta-analysis. Read the caveat before you quote the number.

  1. Explainability · Strong evidence · Supports

    LLMs alone vs people using LLMs

    Tested alone, LLMs named the relevant condition in 94.9% of cases. Members of the public using the same LLMs got there in fewer than 34.5% of cases, no better than a control group using their usual sources.

    Bean et al., Nature Medicine, 2026 · Randomized, preregistered study of 1,298 UK adults

    Caveat: Written scenarios, not real patients, and members of the public, not clinicians. It tests a person working with the AI, not a hospital rollout.

  2. Explainability · Strong evidence · Supports

    AI tutors with and without guardrails

    Students with plain GPT-4 scored 48% higher on practice problems but 17% lower on a later exam without AI than students who never had it. Students with the guardrailed tutor scored 127% higher in practice, and on the exam they were statistically indistinguishable from students who never had AI.

    Bastani et al., PNAS, 2025 · Randomized field experiment with nearly 1,000 high school math students (exact count not verified)

    Caveat: Education, not healthcare, in one country and one subject. The guardrailed tutor avoided measurable harm but didn’t raise exam scores, and its students thought they had learned more than they had.

  3. Explainability · Strong evidence · Supports

    The “jagged frontier”

    On a task just outside the AI’s abilities, BCG consultants using GPT-4 were 19 percentage points less likely to get it right (84.5% correct without AI vs 60% and 70.6% in the two AI groups). On tasks inside its abilities, they did 12.2% more tasks, 25.1% faster, with significantly improved quality, and bottom-half-skill consultants gained the most.

    Dell’Acqua et al., Organization Science 37(2):403–423, March 2026 (DOI 10.1287/orsc.2025.21838) · Preregistered randomized field experiment with 758 consultants

    Caveat: Consulting tasks, GPT-4 in 2023, one firm, and only one task outside the frontier. Quality was graded by people.

  4. Caregiving · Strong evidence · Supports

    NYU Caregiver Intervention

    Counseling and support for spouse caregivers of people with Alzheimer’s delayed nursing-home placement by a model-predicted median of 557 days and cut the placement rate by 28.3%.

    Mittelman et al., Neurology, 2006 · Randomized controlled trial with 406 spouse caregivers

    Caveat: Volunteer spouse caregivers in one New York program, and 557 days is a model-predicted median.

  5. Caregiving · Strong evidence · Supports

    Caregivers in discharge planning

    Bringing family caregivers into discharge planning cut readmissions by 25% at 90 days and 24% at 180 days.

    Rodakowski et al., Journal of the American Geriatrics Society, 2017 · Meta-analysis of 15 studies (11 with readmission data)

    Caveat: The interventions varied a lot between studies.

The full library

All 43 studies, including the ones that complicate our view, with filters by question, direction, strength, and study type.

Enter your email and we’ll send a confirmation link. No login and no password.

Optional. Leave this unticked and we only email the confirmation link.

We’ll email you a confirmation link. How sign-ups are handled.

How to read this