New clues for doctors, genetic researchers and scientists—and careful tests of where AI can help.
This week, a medical chatbot met real patients under close supervision, a genetics model found more clues in research data, and two studies examined where AI may help specialists and scientists. The encouraging part is the evidence being gathered; the limits matter just as much. Our archive story goes back to a 2020 search for new antibiotics.
Worth your attention4 current stories
Primary carePilot
A medical chatbot joins 98 primary-care visits under supervision
In a single Boston clinic, 98 adults completed a text conversation with Google's AMIE before an urgent primary-care appointment. A physician monitored every exchange. None met the study's preset criteria for stopping a conversation, and doctors who read the transcript before the visit said it helped them prepare in 33 of 44 surveyed cases.
AI’s role: AMIE gathered a patient's account and drafted a summary and possible diagnoses for a doctor to review before the ordinary appointment.
Why it matters: A well-supervised intake tool might give clinicians useful context before a visit and give patients another way to explain their concerns.
Keep in mind: This was a one-clinic feasibility study without a comparison group. Supervisors recorded one hallucination and added information in five conversations. It did not test whether care, workload or outcomes improved, and it does not support unsupervised diagnosis or triage.
Sources & timing
The patient encounters took place April–November 2025; the peer-reviewed Lancet study was published October 8, 2026.
A new genetics model looks for disease clues missed by simpler scans
Illumina released SpliceAI2, a model that predicts how DNA changes may alter RNA splicing—the editing step that helps cells build proteins. In the team's tests, it performed better than earlier models on several splicing benchmarks and highlighted rare-disease candidate variants in deep intronic regions often missed by exome sequencing.
AI’s role: The model learns patterns from DNA and RNA data to estimate which variants may change the splice sites, junctions and transcripts a cell produces.
Why it matters: Better prioritization could help researchers investigate otherwise overlooked genetic explanations for rare diseases.
Keep in mind: The evidence is a company-authored manuscript and benchmark analysis, not a trial showing more diagnoses or better patient outcomes. Predicted variants still need laboratory and clinical interpretation.
Sources & timing
Illumina released SpliceAI2 and its research manuscript on October 8, 2026.
Illumina: SpliceAI2 release · October 8 release date and availability; performance claims are attributed to Illumina.
Cancer care accessResearch
A review finds supervised AI entering radiation-oncology workflows
A new systematic review found 18 studies of radiation-oncology AI in low- and middle-income countries. Six described clinician-supervised workflow integration, and one reported coordination across stages of care. The authors mapped where tools are beginning to enter real workflows and where the evidence stops.
AI’s role: The reviewed systems assisted bounded radiotherapy tasks and workflow steps; clinicians remained responsible for care.
Why it matters: Radiation therapy needs specialized staff. Carefully integrated tools could help teams use scarce expertise, and this review gives a concrete map for what needs validation before broader use.
Keep in mind: Most validation was retrospective; no study met the review's highest operational level, and continuous performance monitoring was not reported. The review does not establish improved access, treatment quality or survival.
Sources & timing
The systematic review was published October 9, 2026; it searched studies through December 2025.
Researchers used a physics-informed neural network to estimate hard-to-observe friction patterns along a fault from surface-motion data. Applied to a past slow-slip event in southwest Japan, the model reproduced the observed surface displacement pattern while remaining consistent with the team's fault-physics equations.
AI’s role: The neural network represented spatially varying fault friction; physical equations constrained the fit to measured ground movement.
Why it matters: Understanding slow fault motion could improve research models of how faults behave and help scientists test future forecasting methods.
Keep in mind: This reconstructs a historical event and explores model forecasts of slow slip. It is not an operational earthquake warning system and does not show that damaging earthquakes can be predicted.
Sources & timing
Peer-reviewed paper published October 12, 2026; it analyzes the 2010 Bungo Channel slow-slip event in Japan.
Authors' preprint · Earlier manuscript provides accessible methodological context; journal publication is the new development.
From the archive2020
From the archive · antibioticsResearch
An AI search found an unexpected antibiotic candidate
In 2020, MIT-led researchers trained a deep-learning model on molecules with known antibacterial activity and screened a drug library for different candidates. The model flagged a compound they named halicin. Lab tests showed activity against several bacteria, and it treated two infections in mouse models.
AI’s role: The model ranked molecules likely to have antibacterial activity despite structures unlike familiar antibiotics; researchers then tested the promising hits in the lab and in mice.
Why it matters: The work showed how AI could widen the search for antibiotics at a time when drug-resistant infections make new options valuable.
Keep in mind: Halicin was a preclinical candidate in this result, not a proven treatment for people. The 2020 study demonstrates a discovery method and animal results, not clinical safety or effectiveness.
Sources & timing
The Cell paper and MIT report were published February 20, 2020.