OpenEvidence, the medical search AI app “most used by doctors in the US,” announced a new round of funding in which it tripled its valuation, which now exceeds US$3.5 billion. The round, categorized as Series B, was US$210 million and was led by Google Ventures and Kleiner Perkins funds. Sequoia Capital, the fund that had led the Series A in January, also participated.

Why it matters

Founded in 2021 by Daniel Nadler and Zachary Ziegler, and driven by Mayo Clinic’s startup accelerator program, OpenEvidence is part of a new breed of apps that use AI to ease the burden on doctors when answering clinical queries in consultation.

Faced with any medical question, these apps present in a few seconds a summary of the most relevant scientific articles, ordered according to aspects such as the reputation of the scientific journal, the number of citations or the date of publication, far surpassing conventional AI chatbots, such as ChatGPT or Grok, when faced with clinical or scientific queries.

At least in the case of OpenEvidence, the business model is based on advertising: it is free to use for physicians, while advertisers are offered direct access to a huge number of doctors. In its press release, the company claims that over 40% of U.S. physicians use the service.

How these services revolutionize medicine

In a recent article on how AI is changing medical practice, written by a gastroenterologist and professor of medicine at the University of North Carolina, he posits that these new chatbots take a very different approach to traditional digital tools, such as UpToDate, a platform known for its clinical summaries on a wide range of specialties.

While in that system doctors must read abstracts until they find what they are looking for, in these new apps like OpenEvidence or Consensus (its competitor) they can directly ask a question. They are fast, convenient and specific, in a context where doctors don’t have enough time, he says. In addition, he cites a cardiologist who says he uses OpenEvidence daily and calls it revolutionary.

There is also criticism. Although they guard against “hallucinating” like traditional chatbots, since each answer is backed up by a scientific paper, on internet forums users report occasional errors in reasoning or answers with exaggerated conclusions.

Another criticism comes from MEDCalc founder Dr. Graham Walker, who is concerned about what he calls automation bias: that these tools diminish the specialist’s ability to critically analyze his patients.

New feature: DeepConsult

OpenEvidence used the announcement to launch a new feature, DeepConsult, similar to the “deep” function that is trending in chatbots in general. A feature that allows for more thoughtful, multi-step approach answers to be crafted (Consensus had launched something similar a few weeks earlier).

According to the company, DeepConsult addresses a different need than its main search engine, intended for quick answers during the query. DeepConsult would instead be for when physicians have more time to delve into a specific clinical topic.

These were some of the points they highlighted about their new feature:

– Provides physicians with a set of “PhD-level” AI agents capable of performing advanced research. 


– Makes greater referrals and connections

– Uses 100 times more computational capacity

– Maintains the app’s signature freebie feature.

To dig deeper:

This AI startup founder became a billionaire by developing a ChatGPT for doctors (Forbes)

How AI is changing the way doctors access medical knowledge (Forbes)

Could be one of the most important companies of the next decade (Sequoia)

Consensus app, competition

Microsoft announced a pioneering diagnostic system, the Microsoft AI Diagnostic Orchestrator (MAI-DxO), which outperforms human physicians in complex medical diagnoses. According to their blog post, The Path to Medical Superintelligence, MAI-DxO achieved an 85% accuracy rate in diagnosing challenging cases from the New England Journal of Medicine (NEJM), compared to a 20% accuracy rate for a group of experienced doctors.

Developed under Mustafa Suleyman’s leadership, MAI-DxO emulates a panel of expert physicians tackling intricate diagnostic challenges. Paired with OpenAI’s o3 model, it was tested on 304 NEJM case studies, correctly solving over eight out of ten cases. In contrast, physicians, working without access to colleagues, textbooks, or AI tools, accurately diagnosed only two out of ten cases, ensuring a fair comparison of raw human performance against the system.

MAI-DxO follows a clinician’s diagnostic process, analyzing symptoms, posing targeted questions, and ordering tests like bloodwork or imaging to reach a diagnosis. Unlike traditional AI benchmarks, such as the U.S. Medical Licensing Examination, which rely on memorization, MAI-DxO’s sequential approach mirrors real-world clinical reasoning, as detailed in Microsoft’s blog.

Cost efficiency is a key advantage. MAI-DxO orders fewer unnecessary tests, potentially reducing healthcare costs, a critical issue as U.S. spending nears 20% of GDP. Microsoft emphasizes that MAI-DxO is designed to complement physicians, not replace them. “Their clinical roles are much broader than simply making a diagnosis,” the blog states, highlighting doctors’ ability to navigate ambiguity and build patient trust—skills beyond AI’s scope.

The system is not yet approved for clinical use and requires extensive safety testing and regulatory review. The research, pending peer review, focused on complex cases, and physicians in the study lacked typical resources, which may have impacted their performance. Microsoft is partnering with health organizations to validate MAI-DxO in real-world settings, aiming to enhance diagnostic accuracy and accessibility.

This breakthrough could reshape healthcare by supporting clinicians and reducing costs, particularly in underserved areas. Microsoft’s blog underscores the potential for MAI-DxO to empower patients and clinicians, paving the way for more efficient, accurate diagnoses in challenging medical scenarios.

Google announced plans to release new open-source AI models tailored for drug discovery, aiming to accelerate pharmaceutical research. The initiative, revealed on March 18, 2025, seeks to democratize access to advanced AI tools, enabling researchers worldwide to develop new treatments more efficiently.

The models, built on Google’s AI expertise, are designed to predict molecular interactions, optimize drug candidates, and streamline preclinical testing. Unlike proprietary systems, these open models will be freely accessible, fostering collaboration across academia, startups, and established pharma companies.

Google’s move aligns with its broader push into healthcare innovation. The company has previously developed AI for medical imaging and diagnostics, but this marks its first significant step into open-source drug discovery tools. By sharing these models, Google hopes to address the high costs and lengthy timelines of traditional drug development, which often takes over a decade and billions of dollars.

Industry experts see this as a game-changer. “Open AI models could lower barriers for smaller players, driving innovation in a field dominated by big pharma,” said Dr. Emily Chen, a biotech analyst. However, some caution that ensuring model accuracy and addressing ethical concerns, like unintended biases in AI predictions, will be critical.

Google plans to roll out the models later in 2025, with documentation and support for researchers. The company also intends to collaborate with academic institutions to refine the models based on real-world applications.

This initiative comes amid growing competition in AI-driven drug discovery, with companies like DeepMind and Insilico Medicine advancing similar technologies. Google’s open-source approach, however, sets it apart, potentially reshaping the future of pharmaceutical research.

Craif, a Japanese biotech startup founded in 2018 as a spin-off from Nagoya University, has raised $22 million in a Series C funding round to expand its non-invasive early cancer detection technology. The funding, led by existing investor X&KSK and joined by Unreasonable Group, TAUNS Laboratories, Daiwa House Industry, and Aozora Bank Group, brings Craif’s total funding to $57 million, with a valuation nearing $100 million.

The company’s flagship product, miSignal, uses urinary microRNA (miRNA) analysis combined with AI to detect seven types of cancer—pancreatic, colorectal, lung, stomach, esophagus, breast, and ovarian—at early stages, such as Stage 1. Unlike traditional blood-based tests that rely on cell-free DNA (cfDNA), Craif’s approach leverages miRNA, which is actively secreted by early cancer cells and offers clearer biomarker signals due to urine’s lower impurity levels. This non-invasive method allows patients to collect samples at home, improving accessibility and reducing screening barriers.

Co-founder and CEO Ryuichi Onose, motivated by his grandparents’ cancer diagnoses, partnered with Nagoya University associate professor Takao Yasui to launch Craif just one month after meeting. Yasui’s innovative urinary biomarker method laid the foundation for the company’s technology. “Our goal is to make early cancer detection simple and routine, especially for those deterred by invasive tests or limited access to medical facilities,” Onose said.

In Japan, miSignal is already generating revenue through partnerships with over 1,000 medical institutions and 600 pharmacies, serving 20,000 users. Craif reported $5 million in revenue in 2024 and aims to reach $15 million by the end of 2025. The company offers single tests and subscription models, with many users opting for regular screenings. Craif plans to expand miSignal to detect 10 cancer types by year-end and explore applications for non-cancerous conditions like dementia.

The new funding will fuel Craif’s U.S. expansion, including clinical trials targeting FDA approval by 2029. The company operates an R&D lab in Irvine, California, and plans to open a business office in San Diego. Craif has begun collecting pancreatic cancer samples with 30 medical institutions across 15 U.S. states. “Our technology has the potential to transform cancer screening into a routine part of life,” said Onose.

Craif faces competition from companies like Grail and Freenome, which focus on cfDNA-based tests, but its urine-based miRNA approach sets it apart. With growing recognition of miRNA’s role in cancer biology—highlighted by its link to the 2024 Nobel Prize—Craif is poised to make early detection more accessible and effective, potentially saving lives through timely intervention.

As healthcare systems face strain, many turn to AI chatbots like ChatGPT for medical guidance. A survey indicates one in six U.S. adults uses chatbots for health advice monthly. However, an Oxford Internet Institute study highlights that these tools often deliver unreliable advice due to communication challenges.

The study, involving 1,300 U.K. participants, used doctor-designed medical scenarios. Participants identified health conditions and decided on actions, such as visiting a doctor, using chatbots or traditional methods like online searches. Tested AI models included GPT-4o (ChatGPT’s default), Cohere’s Command R+, and Meta’s Llama 3.

Adam Mahdi, study co-author, told TechCrunch, “The study revealed a two-way communication breakdown.” Users often left out key details in queries, resulting in chatbot responses that mixed accurate and misleading advice. This complicated informed health decisions. The study found chatbot users were less accurate in identifying conditions and more likely to underestimate severity compared to those using other methods.

The findings reveal a gap between AI’s healthcare potential and its current limitations. Companies like Apple and Microsoft are developing AI for health applications, but skepticism persists among doctors and patients about its reliability in critical scenarios. The American Medical Association advises against using chatbots like ChatGPT for clinical decisions, and OpenAI cautions against relying on its tools for diagnoses.

Mahdi emphasized the need for improved testing, stating, “Current evaluation methods for [chatbots] do not reflect the complexity of interacting with human users.” Unlike doctors, chatbots cannot ask follow-up questions or interpret non-verbal cues, which are crucial in medical contexts.

The study recommends relying on trusted sources for healthcare decisions, Mahdi noted. The findings suggest that until chatbots better handle human interactions, professional medical guidance remains more reliable than AI advice, according to the researchers.