0727

AI Insights & Research

Journal paper | Randomized field experiment in knowledge workOrganization Science·

Similar work can fall on opposite sides of AI’s capability frontier

A randomized experiment with 758 consultants found improved speed and quality on tasks within GPT-4’s capabilities but reduced accuracy on a deliberately selected task outside that frontier. The 2026 journal version follows a 2023 working paper. Its model and task setting cannot directly measure the capabilities of all current systems.

Source
Bank conference summary | Healthcare practices and expectationsMorgan Stanley Research·

Post-discharge follow-up is an emerging healthcare AI use case

Morgan Stanley’s healthcare-conference summary describes agents checking follow-up appointments, refill needs and continuing symptoms, alongside expectations for AI in drug development and operations. These are participant accounts and forecasts, not clinical-trial results. Continued patient engagement does not establish fewer hospital readmissions.

Source
Vendor partnership case | Early wildfire detectionMicrosoft Unlocked·

AI spots smoke; firefighters decide the response

Microsoft describes ALERTCalifornia using camera feeds and AI to identify early smoke and filter noise before dispatchers and firefighters assess the scene. It is a partner case account and interviews, not an independent estimate of avoided losses. Hypothetical fire-spread comparisons are not observed counterfactual outcomes.

Source
Vendor perspective | Commissioned healthcare-leader surveyMicrosoft Cloud·

Healthcare AI depends on connected, governed data

A Microsoft-commissioned survey of 500 healthcare decision-makers across seven countries highlights gaps between AI confidence and execution, including fragmented legacy systems. The article advocates governed data and workflow integration. Self-reported perceptions do not establish that a particular platform improves clinical outcomes or delivers scalable returns.

Source
Vendor transparency-report introduction | Self-reported governance practicesMicrosoft·

AI governance moves toward continuous operational checks

Microsoft’s transparency-report introduction describes governance of agent identities, permissions, runtime monitoring and interactions across systems, with red-team findings converted into repeatable tests. This is a company’s disclosure of its own practices rather than an independent audit; implementing controls does not demonstrate elimination of all risks.

Source
Journal paper | Human–AI dermatological diagnosis experimentsNature Medicine·

The same AI explanation can affect experts and novices differently

Experiments with 623 lay people and 153 primary care physicians found overall diagnostic gains from AI, but stronger deference to incorrect language explanations among non-experts. Experienced clinicians were more resilient, and presentation order mattered. These dermatology-task findings do not generalize to every clinical setting or validate self-diagnosis.

Source
University research account | Statistical generation of extreme-event scenariosMIT News·

AI-generated extremes support scenarios, not exact disaster forecasts

MIT describes η-learning combining spatial data with extreme-value statistics to generate plausible precipitation scenarios beyond training examples. Such scenarios may support infrastructure stress tests, but still depend on relevant data and statistical assumptions. A return-period scenario is not a prediction of an event’s precise time or location.

Source
University research account | Crystal generation and computational validationMIT News·

Constrain chemistry before generating new materials

MIT’s CrysVCD incorporates valence-related chemical constraints before generation to improve computational stability while targeting desired properties and reducing downstream screening. It is best suited to ordered crystalline structures. Computational results do not establish successful synthesis, manufacturing scale-up or performance in commercial products.

Source
University project account | Interdisciplinary educator-training pilotMIT Schwarzman College of Computing·

AI education needs teachers who connect tools with disciplines

MIT’s pilot brought 19 higher-education instructors together to adapt machine-learning materials to their disciplines and develop students’ critical judgment. The approach builds teaching capacity through collaboration. This workshop account and participant feedback do not constitute controlled evidence of improved student learning.

Source
Journal paper | Vehicle deployment and human-factors experimentsNature·

Driving explanations help humans anticipate system behavior

CW-Net grounds driving-planner decisions in interpretable concepts. Vehicle deployment and user studies find improvements in people’s predictions of behavior, especially in surprising situations. This demonstrates useful interpretability in the tested system, not population-level accident reduction or a guarantee that humans can always intervene in time.

Source
Stay curious. See you in the next read.

Join the conversation after approval.

Comments, replies and messages are reserved for approved, signed-in collaborators.

  1. Submit your collaborator profile
  2. Wait for profile review
  3. Sign in after approval to participate
Already a member? Sign inBecome a collaboratorThis is a flow preview. Account sign-in and approval checks are not connected yet.