A dentist reviewing an X-ray, a scientist testing a protein, and an engineer measuring an algorithm have something in common: they can check whether an AI system helped. The answer does not depend on whether its output sounded convincing.
That is the useful starting point for a discussion about AI. Ask what a system produced, how it was tested, and what the test actually establishes. The examples below cover 21 applications across healthcare, biology, robotics, mathematics, software, materials, and weather.
Each example names the kind of technology involved. Claude and GPT-4 are large language models. Some systems combine language models with specialist tools or symbolic reasoning. Others use different forms of machine learning. Those distinctions matter when assessing what has been achieved.
A living record of AI progress. This collection is reviewed weekly for significant advances, stronger evidence, and corrections. Each entry explains what happened, why it is useful, and what remains unproven. Last updated: 19 October 2026.
Medicine and diagnosis
1. Helping dentists make fewer false positive diagnoses
Technology: specialised clinical AI. Evidence: randomised crossover study.
Thirty dentists assessed 50 panoramic radiographs with and without AI assistance. Their accuracy in identifying periapical radiolucencies, dark areas around tooth roots that can indicate disease, increased from 91.6% to 93.3%. The false positive rate fell from 4.3% to 2.0%. Cone-beam CT scans provided the reference standard.
This was a modest improvement with a useful practical implication: fewer healthy findings were incorrectly labelled as disease. Less experienced dentists benefited most. Sensitivity did not improve, and the study did not establish that AI-assisted decisions led to better long-term patient outcomes.
Source: Journal of Dentistry, randomised trial of AI assistance for periapical radiolucencies.
2. Improving physicians' management reasoning
Technology: GPT-4. Evidence: randomised clinical case study.
In a trial involving 92 physicians, participants answered questions about five complex clinical cases using either conventional resources alone or those resources plus GPT-4. Access to GPT-4 improved management-reasoning scores by an adjusted 6.5 percentage points, with a 95% confidence interval of 2.7 to 10.2 points.
The assisted group also took roughly two minutes longer per case. These were case exercises scored against expert-developed rubrics, so the result establishes a benefit for that task rather than a demonstrated improvement in patient health.
Source: Goh and colleagues, management-reasoning trial, original preprint.
3. Producing strong diagnostic reasoning, with a revealing limitation
Technology: GPT-4. Evidence: randomised study with blinded expert grading.
A separate study recruited 50 physicians to work through diagnostic case vignettes. In an exploratory comparison, the model alone scored approximately 16 percentage points higher than physicians using conventional resources. However, giving physicians access to the model did not significantly improve their diagnostic-reasoning scores.
Both findings belong in the same account. A capable model does not automatically make an effective clinical workflow. How people use its output can determine whether its capabilities translate into a benefit.
Source: JAMA Network Open, Large Language Model Influence on Diagnostic Reasoning.
4. Testing language-model assistance in radiology
Technology: multimodal GPT-4. Evidence: randomised reader study.
A 2026 study asked 101 US radiologists to diagnose 20 cases containing clinical descriptions and CT or MRI images. Researchers compared no model support with three ways of presenting GPT-4's suggestions.
Explanations that set out the reasoning produced the strongest diagnostic performance. This makes the study useful beyond the question of whether a model can read an image: the presentation of its advice affected the human result. The findings came from controlled case assessments, not a prospective trial of routine patient care.
5. Reaching regulated medical use
Technology: several forms of medical AI. Evidence: regulatory authorisation.
The FDA maintains a list of AI-enabled medical devices authorised for marketing in the United States. It includes applications across medical imaging, with radiology particularly prominent.
This is evidence that AI has progressed into regulated products. It is not evidence that every listed device improves patient outcomes, nor that every device uses a language model. Each authorisation and intended use needs to be considered separately.
Source: FDA, List of Artificial Intelligence-Enabled Medical Devices.
6. Supporting breast-cancer screening at population scale
Technology: specialised imaging AI. Evidence: population randomised trial.
The Swedish MASAI trial randomised 105,934 women to AI-supported mammography screening or standard double reading. Its 2025 screening analysis reported 338 detected cancers in the AI-supported group and 262 in the control group, corresponding to a cancer-detection rate ratio of 1.29.
The study also reported a substantial reduction in screen-reading workload. Unlike the clinical vignette studies, this involved an actual screening programme. Detection, false positives, interval cancers, and longer-term outcomes remain distinct measures; an improvement in one should not be presented as proof of all the others.
Source: The Lancet Digital Health, MASAI screening-performance analysis.
7. Helping doctors prepare for visits with real patients
Technology: AMIE, a conversational language-model system. Evidence: peer-reviewed, prospective feasibility study with physician supervision.
A study published in The Lancet on 8 October 2026 evaluated Google's AMIE before urgent primary-care appointments at Beth Israel Deaconess Medical Center. Of 114 enrolled adults, 98 completed both the AI conversation and their appointment. Doctors monitored every interaction; none required a stop under the predefined safety criteria, although supervisors recorded one hallucination and supplied clinical information in five interactions.
Doctors completed follow-up surveys for 60 cases. In the 44 where they had reviewed the AI transcript before the appointment, they found it useful for preparation in 33, or 75%. The practical aim is better-prepared consultations.
The Alphabet-funded study involved English-speaking adults at one centre and had no control group. It demonstrates feasibility and perceived usefulness under supervision; it does not establish improved patient outcomes, reduced workload, or safe autonomous diagnosis.
Biology and laboratory results
8. Identifying an uncharacterised enzyme system
Technology: Claude agents. Evidence: computational discovery and initial laboratory investigation.
In September 2026, Anthropic reported that roughly 950 Claude agents spent 21 hours searching genomic data and identified an unusual system involving a reverse transcriptase, repeated DNA sequences, and an accessory protein. The researchers named it array-associated reverse transcriptases, or ART.
The underlying enzyme had appeared in earlier research. The reported discovery concerned the associated biological system and its defining relationships. Human scientists performed the laboratory work, and the system's function remained unknown at the time of the announcement. It is an early research finding, with further characterisation still needed.
Source: Anthropic, Claude discovers a novel enzyme system with CRISPR-like repeats.
9. Designing proteins that bind in physical experiments
Technology: Claude operating specialist protein-design tools. Evidence: external laboratory testing.
Anthropic reported that Claude-led design campaigns produced confirmed protein binders against 14 of 15 evaluated targets. Adaptyv Bio and Twist Bioscience synthesised and tested the designs. Reported hit rates ranged from 22.6% to 35.1%, depending on the model and campaign setup.
Claude coordinated established protein-design and structure-prediction models. The outcome was therefore a result from an integrated research system, with physical binding assays providing the check. A working binder is a research result, not an approved medicine or proof of safety and effectiveness in patients.
Source: Anthropic, Claude accelerates protein design and analytical chemistry.
10. Finding an antibiotic candidate
Technology: deep learning. Evidence: laboratory assays and mouse experiments.
A neural network helped researchers identify antibacterial activity in halicin, a compound structurally different from conventional antibiotics. Subsequent experiments demonstrated activity against a range of bacteria, including treatment of C. difficile and highly resistant A. baumannii infections in mouse models.
The AI prediction was the beginning of the investigation. Biological experiments supplied the evidence. The study demonstrated a promising antibiotic candidate, not an established treatment for people.
Source: Cell, A Deep Learning Approach to Antibiotic Discovery.
11. Making protein-structure prediction far more useful
Technology: AlphaFold, specialised deep learning. Evidence: comparison with experimental structures.
AlphaFold2 made highly accurate protein-structure prediction practical at a scale that changed how researchers investigate proteins. Predictions can be compared with structures determined experimentally, providing a concrete measure of performance.
Demis Hassabis and John Jumper shared half of the 2024 Nobel Prize in Chemistry for protein-structure prediction. A predicted structure still has uncertainty and does not establish every aspect of a protein's function, but its scientific usefulness goes well beyond generating plausible prose.
Source: Nobel Prize in Chemistry 2024, official announcement.
Robotics and autonomous driving
12. Turning visual and language knowledge into robot actions
Technology: RT-2, a vision-language-action model. Evidence: robot experiments and simulation.
DeepMind's RT-2 represents robot actions as tokens, connecting web-trained vision-language models with robotic control. Its evaluations demonstrated improved generalisation to unfamiliar objects, instructions, and situations.
On the Language Table task suite, it achieved a 90% success rate in simulation, compared with 72% to 77% for the cited baselines. That figure describes the simulation test, not a 90% success rate for arbitrary real-world tasks. The broader work is evidence that language-model-derived systems can guide physical actions.
Source: Google DeepMind, RT-2: New model translates vision and language into action.
13. Evaluating driverless vehicles over millions of miles
Technology: autonomous-driving AI. Evidence: observed crashes and human-driver benchmarks.
A peer-reviewed Waymo study compared crash rates across 56.7 million rider-only miles with human-driving benchmarks. This gives researchers a way to evaluate a deployed autonomous system using events on actual roads.
These comparisons require care about geography, operating conditions, reporting thresholds, and the construction of the human benchmark. They do not establish that autonomous driving is equally safe everywhere. They do establish that its performance can be studied through large-scale operating data.
Mathematics and working software
14. Opening hundreds of research mathematics results to scrutiny
Technology: an unreleased OpenAI frontier model. Evidence: public research manuscripts and partial Lean formalisation; review is ongoing.
On Tuesday, 6 October 2026, OpenAI released a large collection of mathematical research produced by an internal model. By 9 October, the repository listed 719 manuscripts organised into 372 related families. OpenAI reports that the model was given approximately 4,000 problems, with each retained result using roughly three hours of ChatGPT Pro-equivalent thinking compute on average. These are research outputs that people can inspect, rather than a score on a school-maths benchmark.
One striking example is a claimed matrix-multiplication exponent of at most 9/4 over the complex numbers. In plain language, this concerns how the arithmetic needed to multiply large matrices grows with their size. If established, such a result advances the theoretical foundations of a calculation used throughout science and computing; it does not by itself demonstrate faster software on today's hardware.
The public proofs are the important part. OpenAI's 7 October update reports Lean formalisation for 300 of 719 main results, about 42%. Lean can mechanically check a formal proof, while checking that its definitions and assumptions match the intended mathematical claim remains essential. The collection has mixed verification status: that same update withdrew three manuscripts after a sign error and revised 14 others. A manuscript count is therefore not a count of independently accepted breakthroughs.
This is a substantial demonstration of AI generating research candidates and proof artifacts at scale. Its lasting value will be determined result by result, through mathematical scrutiny, formal checking, and what other researchers can build on. We will track the repository's revisions as that evidence develops.
Sources: OpenAI's 6 October announcement; the manuscript repository and methodology; the mathematical overview, family 107; corrections and formalisation history.
15. Discovering new mathematical constructions
Technology: FunSearch, an LLM with an automated evaluator. Evidence: executable programs and checkable constructions.
FunSearch generated programs that produced new constructions for the cap-set problem in combinatorics, improving previously known results. The same approach produced better heuristics for online bin packing, a problem involving the efficient allocation of items to containers.
The evaluator tests candidate programs and rejects failures. The achievement is a productive combination of language-model generation, a carefully specified problem, and external checking. It did not solve the entire cap-set problem.
Source: Nature, Mathematical discoveries from program search with large language models.
16. Producing Olympiad-level geometry proofs
Technology: AlphaGeometry, a neural language model with symbolic deduction. Evidence: independently checkable proofs.
AlphaGeometry solved 25 of 30 Olympiad geometry problems in its published evaluation, compared with 10 for the previous best method and an average of 25.9 for human gold medallists on the benchmark.
The language model proposes useful constructions while the deduction system checks the resulting reasoning. The result demonstrates substantial ability in a defined area of mathematics, not unrestricted mathematical competence.
Source: Nature, Solving olympiad geometry without human demonstrations.
17. Discovering faster sorting routines that shipped
Technology: AlphaDev, reinforcement learning. Evidence: code review, correctness checks, and deployment.
AlphaDev discovered sorting routines that were translated into C++ and incorporated into LLVM's libc++ library. The paper reported improvements of up to 70% for certain short sequences and approximately 1.7% for sequences longer than 250,000 elements on the tested configurations.
Those are specific benchmark results rather than a universal speedup for software. The important achievement is that automatically discovered algorithms survived review and entered a widely used software library.
Source: Nature, Faster sorting algorithms discovered using deep reinforcement learning.
18. Finding better matrix-multiplication algorithms
Technology: AlphaTensor, reinforcement learning. Evidence: explicit mathematical constructions.
AlphaTensor found matrix-multiplication algorithms that improved on known results for several matrix sizes. Its outputs take the form of tensor decompositions whose correctness can be checked independently.
The applicable number system, matrix size, and optimisation target matter. A reduction in arithmetic operations does not automatically imply faster execution on every computer. What makes this a strong example is that the mathematical output can be tested directly.
Source: Nature, Discovering faster matrix multiplication algorithms with reinforcement learning.
19. Helping researchers find gaps before submitting papers
Technology: Gemini-based agents in Google's Paper Assistant Tool, now Deep Review. Evidence: conference pilots and author surveys; technical methods described in a preprint.
On 14 October 2026, Google opened Deep Review to the public after its Paper Assistant Tool had provided private feedback on more than 16,000 manuscripts across four conference programmes. The system divides a document into sections, examines its claims, and combines the findings into feedback for the author. The practical value is helping researchers catch problems while there is still time to fix them.
The earlier ICML 2026 pilot provides a concrete measure of usefulness. It processed approximately 4,500 papers and received 869 survey responses. Among responding authors whose papers contained theoretical results, 35.4% reported gaps that took more than an hour to repair. Among those with experimental results, 31% said the feedback prompted additional experiments. These figures come from a retrospective coauthored by the ICML programme chairs and Google researchers.
These are self-reported outcomes from voluntary participants and survey respondents, not an independent audit of every flagged error or a controlled estimate of improved research quality. The pilots used earlier system versions, so their figures should not be treated as a performance guarantee for the new release. Feedback was separate from official peer review; authors still had to check it. The demonstrated contribution is useful criticism within a human research process, not automatic certification that a paper is correct.
Sources: Google's 14 October Deep Review announcement; ICML's pilot retrospective and survey results; the Paper Assistant Tool preprint.
Materials and weather
20. Expanding the map of potentially stable materials
Technology: GNoME, graph neural networks. Evidence: computational physics and experimental structure matches.
GNoME identified more than 2.2 million crystal structures stable relative to earlier datasets, with approximately 381,000 new entries on its updated stability convex hull. The paper also reported 736 structures matching independently realised experimental structures.
The distinction between prediction and synthesis is essential. The study did not physically manufacture millions of new materials. Its contribution was a much larger set of computationally supported candidates, with experimental matches providing an additional check. Those matches should not be described as hundreds of experiments commissioned by the AI.
Source: Nature, Scaling deep learning for materials discovery.
21. Generating accurate medium-range weather forecasts
Technology: GraphCast, graph neural networks. Evidence: forecast evaluation against reference weather data.
GraphCast generated global ten-day forecasts in under a minute on the hardware used by the researchers. The published work reported better performance than ECMWF's high-resolution deterministic system on around 90% of 1,380 verification targets.
This describes a particular published comparison, not a permanent lead over every forecasting system. Forecast variables, lead times, extreme events, and uncertainty all deserve separate evaluation. Even with those boundaries, the work demonstrates that learned models can make useful predictions about a complex physical system.
Source: GraphCast, Learning skillful medium-range global weather forecasting, published in Science.
What these results establish
These examples carry different kinds of evidence. A randomised screening programme is different from a case exercise. A protein binding in a laboratory is different from a medicine helping a patient. A mathematically correct algorithm is different from a robot operating reliably in an unfamiliar home.
Keeping those boundaries visible makes the achievements easier to assess. AI systems have contributed to new mathematical constructions, experimentally tested molecules, improved performance on some clinical tasks, deployed algorithms, and useful predictions about the physical world.
For language models specifically, the strongest examples often involve tools, evaluators, and people with domain expertise. Their usefulness comes from the whole process: generate a candidate, test it, inspect the result, and keep what survives.
The fact that AI generated an output tells us where it came from. Evidence tells us what to make of it.
Originally adapted from AI Evidence Report 2026; maintained as a living collection. Initial source review: 9 October 2026. Latest addition, 19 October 2026: Deep Review and the conference-pilot evidence behind it. This weekly review also checked the OpenAI maths repository and Anthropic enzyme announcement; neither required an entry revision.