Healthy Person's AI - ForkLog

# AI of a Healthy Person

How Artificial Intelligence is Changing Medicine

Despite the hype, speculation, and frightening predictions, few experts doubt that artificial intelligence will indeed change the world. But who will benefit from the changes and what price will have to be paid for them — these remain open questions.

History shows that technological breakthroughs, along with opportunities, almost always bring crises, forcing society to seek a new balance. But there is one field where the benefits of technological progress have seemed almost indisputable for decades: medicine.

ForkLog has investigated how the use of artificial intelligence today is accelerating the creation of new drugs, optimizing laboratory processes, improving diagnostic accuracy, and changing approaches to treating diseases.

Drug Development

Most drugs work by interacting with receptor proteins — molecular structures that regulate cell function and are involved in almost all processes of the body.

Artificial intelligence systems can analyze the structure of receptor proteins and predict which compounds can interact with them most effectively and with minimal side effects. As a result, tasks that previously required many years of laboratory research are increasingly being solved in months.

According to estimates by experts from the World Health Organization (WHO), in the coming years, most new pharmaceutical drugs will somehow be developed using AI.

AlphaFold and Isomorphic Labs

In 2024, the Nobel Prize in Chemistry was awarded to David Baker, Demis Hassabis, and John Jumper. The latter two work at Google DeepMind, and they were honored for developing methods for predicting protein structures, including AlphaFold, which is based on machine learning.

In 2018, AlphaFold took first place in the molecular prediction competition Critical Assessment of Structure Prediction (CASP), demonstrating effectiveness in the most challenging categories. Two years later, at the next CASP, a new version — AlphaFold 2 — won.

In 2021, Google DeepMind released the AlphaFold2 code and a database of predicted protein structures to the public. Around the same time, Hassabis founded Isomorphic Labs — a subsidiary of Alphabet that develops AI for drug discovery.

In 2024, Isomorphic Labs entered into partnerships with Eli Lilly and Novartis. The deals provided funding for the company's AI research of up to $1.7 billion and up to $1.2 billion, respectively. In 2026, Isomorphic Labs also announced a partnership with Johnson & Johnson.

In February 2026, Isomorphic Labs introduced a universal drug design environment, the Drug Design Engine (IsoDDE), built on AlphaFold technologies.

Currently, Isomorphic Labs is working on solutions in oncology and immunology. Despite accelerating development with AI, projects remain at the preclinical research stage. The company expects to begin first human trials in the coming years.

Exscientia and Recursion Pharmaceuticals

Founded in 2012, Exscientia became one of the first companies to systematically apply machine learning to drug development.

In 2020, the drug DSP-1181 for OCD therapy became the first product created using AI to enter clinical trials. The development was carried out jointly with the Japanese pharmaceutical company Sumitomo Dainippon Pharma, which handled synthesis and laboratory tests based on Exscientia's theoretical results.

By 2023, the company had 8 candidate molecules ready, developed 'significantly faster' than the industry average.

In 2024, Recursion Pharmaceuticals acquired Exscientia in a deal worth $688 million. Some research programs were closed.

By that time, several drugs had reached the second stage of clinical trials — testing efficacy and side effects on a group of 100–300 patients.

The merger with Recursion Pharmaceuticals allowed the use of Exscientia's AI systems in combination with an automated laboratory complex for testing. Additionally, Recursion built its own AI supercomputer, BioHive-2, based on NVIDIA H100, for training specialized models.

The company also participated in the development of the open generative model Boltz-2, designed for predicting three-dimensional protein structures.

By 2025, Recursion Pharmaceuticals had concentrated efforts on four programs in oncology and two related to rare diseases. Several drugs are already at the transition stage between Phase 1 and Phase 2 trials:

• REC-4881 for the treatment of familial adenomatous polyposis — a disease that increases the risk of colorectal cancer;
• REC-617 — for the treatment of malignant ovarian tumors;
• REC-1245 for combating lymphoma and other forms of malignant tumors.

The drug REC-3565, intended for the treatment of chronic lymphocytic leukemia, is undergoing Phase 1 clinical trials.

Insilico Medicine

Founded in 2014, Insilico Medicine is another significant player in AI drug development.

In 2017, Insilico Medicine was listed among the top 5 projects for social impact according to Nvidia.

The company uses artificial intelligence at all stages of the development cycle:

• PandaOmics is responsible for finding biological 'targets' — molecules that need to be 'turned off' or regulated as part of therapy;
• Chemistry42 provides generative design of suitable compounds;
• InClinico optimizes clinical trial forecasting.

One of Insilico Medicine's early AI achievements is the drug Rentosertib (ISM001-055), related to the treatment of fibrosis. Development took 18 months from target discovery by the AI system to obtaining a candidate molecule. As of 2025, Rentosertib is undergoing Phase 2 clinical trials.

Additionally, in 2024, the AI-developed immunomodulatory drug ISM3312 for COVID-19 and other viral infections completed Phase 1 trials. ISM3091, related to cancer therapy, was approved for patient testing.

Diagnostics and Research

According to experts, about 90% of all medical information is represented by images such as X-rays and tomograms. This data is critical for diagnosis, but its analysis is a labor-intensive and nontrivial task.

Machine learning methods, especially convolutional neural networks, are well-suited for recognizing complex visual patterns. Similar to human vision, such systems can distinguish contrast edges, shapes, and textures in an image. This allows for high-confidence detection of tumors, hemorrhages, and other abnormalities.

For training AI models, inherently high-quality data is available — arrays of documented images with expert annotations.

In 2024, researchers from Harvard Medical School presented an AI model called Chief, capable of detecting several forms of cancer. According to the developers, the solution correctly identified signs of disease in digital images in 94% of cases.

In 2025, the U.S. Food and Drug Administration (FDA) granted 'breakthrough device' designation to the Damo Panda model from Damo Academy, the research division of Alibaba.

According to the developers, the system can detect signs of pancreatic cancer on tomograms even before symptoms appear, which is especially important for this form of the disease.

In 2026, a significant breakthrough in AI diagnostics was the REDMOD system, developed by the American nonprofit organization Mayo Clinic.

The model, also intended for detecting pancreatic cancer, outperformed specialists in diagnosing the disease at early stages. According to researchers, the system found pathological changes on tomograms on average 475 days before diagnosis.

Google Initiatives

Google is one of the key providers of AI for medical diagnostics and research.

The company offers a line of open models for analyzing medical texts, images, and audio — MedGemma, based on Gemma 3.

Through Health AI Developer Foundations, developers have access to open sets of weights and AI tools.

Google collaborates with a number of clinics and research organizations, focusing on the development of fundamental technologies.

In 2019, the company presented a model for detecting and predicting lung cancer. The model performed on par with or better than a group of six certified radiologists.

In 2020, as part of a collaboration with Northwestern Medicine, researchers demonstrated a system for analyzing mammograms that could detect cancer at the level of a specialist.

In 2024, Google Cloud and the German pharmaceutical company Bayer announced the launch of a platform for screening X-rays. The system analyzes the history of images and medical history data, forming hypotheses about possible pathologies.

NVIDIA and GE HealthCare's X-ray Robot Radiologists

Tech giant Nvidia and American medical technology company GE HealthCare, which produces fluoroscopy equipment, are developing their own AI system for autonomous image acquisition.

Unlike models that analyze already taken images, this solution is intended to reduce the routine burden on specialists and make diagnostics more standardized.

In the first phase, the system will work with X-rays and ultrasound images.

GE HealthCare also plans to use NVIDIA Isaac for Healthcare, a platform for developing autonomous medical systems, including surgical robots.

PathAI Diagnostic Platform

Founded in 2016, PathAI developed the 'digital pathology platform' AISight Dx, intended for primary diagnosis in clinical settings.

The system offers an environment for working with medical images with the ability to integrate third-party algorithms for data analysis.

Support is claimed for a set of CE-IVD-certified AI-based solutions, in particular — 'plugins' for oncological diagnostics:

• DeepDx Prostate allows automatic highlighting of tissues in the image and marking areas potentially important for diagnosis;
• Histotype Px Colorectal builds disease progression predictions based on images, assesses the appropriateness of chemotherapy, and offers therapeutic recommendations;
• Visiopharm identifies and quantifies biomarkers for various forms of cancer.

The platform has its own functions for automatic image analysis, assistance in forming diagnoses, and report writing, but they are currently intended 'for research purposes only' and are not allowed for clinical use.

AISight Dx also offers built-in auxiliary AI tools:

• ArtifactDetect — for finding scan artifacts and other errors in images;
• Case Priority — for prioritizing clinical cases based on tissue analysis;
• AIM-Tumor Cellularity — for assessing tumor cellularity.

In 2022, the solution received U.S. FDA 510(k) clearance and the European CE mark, indicating the product's safety for consumers and the environment.

In 2025, PathAI announced a partnership with the Moffitt Cancer Center in Florida, USA, to integrate AISight Dx into diagnostic processes. In 2026, the company entered into a similar agreement with the University Hospital Zurich.

In May 2026, the Swiss pharmaceutical company Roche announced the acquisition of PathAI in a deal worth over $750 million.

Problems and Limitations

As in other industries, the use of AI in medicine exacerbates systemic problems and creates new ones.

AI assistants, especially those based on LLMs, are not immune to hallucinations.

In Google's research paper on the Med-Gemini model, an error was found: the model 'invented' a nonexistent brain region called the basal nuclei.

The hallucination was formed from two real anatomical names: basal ganglia and basilar artery. The developers cited a typo, but several specialists called the incident a worrying example of the risks of deploying AI assistants in medicine.

Researchers from Stanford University discovered that AI models have the ability to convincingly diagnose diseases from medical images without access to the images themselves.

One of the analyzed systems 'blindly' showed high results in a radiology test. Models GPT-5, Gemini 3 Pro, and Claude Opus 4.5 'confidently described visual details' on nonexistent images.

According to a study published in June of the same year, in a medical context, 7.1% of GPT-4's responses to patient questions were incorrect and could have led to significant harm. In one in 156 cases, the error posed a life-threatening risk.

According to 2025 data, tools for automatic documentation based on patient dialogues introduced errors in 70% of clinical notes. Models added false facts to the conversation transcript, omitted points, and confused concepts.

In addition to LLMs fabricating organs, they are characterized by opaque logic, which makes it difficult for humans to analyze how certain conclusions were reached.

Lack of representativeness in datasets can form biases and attachment to false patterns in models trained on them.

Furthermore, typical problems of AI assistants, such as user cognitive dependence and data privacy, are only exacerbated in the healthcare context.

WHO experts classify the use of artificial intelligence in medicine as a high-risk area.

Under the European AI Act, from August 2026, AI systems in this category will be required to comply with a number of special requirements related to risk management, reporting, and human oversight.

Despite the challenges and potential risks of implementation, the WHO views the prospects of artificial intelligence in medicine positively, provided there are proper regulations and oversight from government agencies.

The U.S. FDA is also optimistic about the prospects of medical AI, although it acknowledges that current regulations are outdated. Formally, in the U.S., such systems are classified as software under the Software as a Medical Device category.

In 2025, the FDA published a set of recommendations related to the AI product lifecycle, risk management, and marketing.