Marc Andreessen did not mince words. In a recent conversation that quickly spread across social media, the venture capitalist declared Doctor ChatGPT superior to the vast majority of practicing physicians. “Doctor ChatGPT is a better doctor than 99% of doctors,” he said. Doctors hate when you say it, but it just is.
The remark, captured by the New York Post on June 29, 2026, landed at a moment when millions already turn to the chatbot for medical questions. OpenAI reports more than 40 million daily health-related queries. Yet fresh studies paint a far more complicated picture than Andreessen’s sweeping assertion. Some experiments show advanced AI models matching or exceeding human performance in narrow tasks. Others reveal consistent gaps in risk assessment and real emergencies.
“99% of the time, the answer that I’m getting from the AI is better than I would get from talking to basically almost any expert,” Andreessen told Joe Rogan earlier in June. The claim echoes his past statements on AI’s edge in coding. This time the stakes feel higher. Health decisions carry immediate consequences for patients.
Andreessen Horowitz, the firm he co-founded, has invested heavily in health-focused AI companies. That includes ventures like Hippocratic AI, Ambience Healthcare and OpenEvidence, tools many doctors already use. The connection does not invalidate his view. It does invite scrutiny. When the messenger stands to gain from widespread adoption, evidence matters more than ever.
The original spark for much of this discussion came from a Thenextweb.com article published today. It highlights the absence of specific data behind the 99 percent figure. It also notes recent research where ChatGPT Health missed over half of true emergency cases in tested scenarios. The pattern holds across multiple papers. AI often shines at suggesting likely diagnoses. It struggles more when asked to weigh treatment options or flag urgent threats.
Consider a February 2026 study led by Mount Sinai urologist Dr. Ashwin Ramaswamy. His team evaluated OpenAI’s specialized health model across 60 clinical situations. The system failed to direct 51.6 percent of genuine emergencies to the emergency room. Instead it recommended routine follow-ups. One simulated patient headed toward respiratory failure. Such misses raise obvious questions about standalone use.
But other findings complicate the narrative. A Harvard-led effort published in April 2026 tested an OpenAI reasoning model called o1-preview on emergency room tasks. The system outperformed two experienced physicians. It did so using only electronic health records available at the time of the original visits. Reviewers rated its diagnostic reasoning, test recommendations and case management at or above expert human levels. NPR covered the results on April 30, noting the model’s step-by-step approach mirrored strong clinical thinking.
Stanford Medicine researchers reached a similar conclusion in early 2025. They pitted a chatbot against physicians who had access to internet searches and standard references. The chatbot alone scored higher on nuanced clinical decisions. When doctors received their own AI support, they performed just as well as the standalone model. The lesson seems clear. Tools amplify human judgment. They do not yet replace the full chain of responsibility.
A 2025 meta-analysis examined 83 separate studies on generative AI for diagnosis. Published in Nature Digital Medicine, it found overall accuracy at 52.1 percent. That number sits near a coin flip. AI models showed no significant difference against non-expert physicians. They fell short of expert specialists by 15.8 percentage points. Several newer models, including variants of GPT-4o, Claude 3 and Gemini, performed closer to experts in some domains. The authors stressed high risk of bias in many of the underlying papers. Still the gap with top-tier human performance persisted.
Empathy offers another angle. Multiple reports show ChatGPT crafting responses patients rate as more compassionate than those from busy doctors. A widely cited study from 2023 found AI answers scored higher on both quality and bedside manner. Patients valued the thoroughness and lack of rush. One New York Times opinion piece from October 2024 carried the headline “I’m a Doctor. ChatGPT’s Bedside Manner Is Better Than Mine.” The author, a practicing physician, described the uncomfortable realization.
Yet empathy without accuracy creates its own risks. A 2024 Journal of Medical Internet Research experiment asked both doctors and ChatGPT-4 to answer 100 real patient questions. Lay readers preferred the AI versions for clarity and usefulness. Specialist reviewers flagged 15 answers as potentially harmful. Ordinary users could not reliably distinguish safe replies from dangerous ones. The gap between perceived quality and clinical safety remains wide.
And. Recent regulatory moves add urgency. Utah approved a pilot program in early 2026 allowing AI to handle certain medication refills without initial physician sign-off. The platform, Doctronic, claims 99.2 percent concordance with human decisions on low-risk cases. Robert Wachter, a prominent digital health expert, wrote on his Substack that the development marks an early step toward routine AI involvement in care. He expressed cautious acceptance for simple refills but warned against overinterpreting early results.
OpenAI itself launched ChatGPT Health around the same time. The feature aims to provide more structured medical conversation. Early feedback mixes praise for accessibility with concern over unverified advice. A Penn State Health study from June 2026 found large language models achieved nearly 76 percent accuracy on healthcare queries. That error rate still doubles the typical rate for human physicians in controlled settings. The researchers concluded current models work best as aids for trained clinicians rather than direct patient tools.
Physicians express mixed feelings. A Sermo poll from 2025 showed 58 percent believe AI will reshape their roles, either reducing scope or changing daily practice. Few expect full replacement soon. Most see augmentation. The volume of medical literature grows too fast for any person to track. AI can surface latest drug interactions or trial outcomes in seconds. Older patients with multiple conditions especially benefit from that breadth.
But. Clinical judgment involves more than data recall. It requires reading subtle cues during physical exams, understanding patient context and accepting uncertainty. AI models trained on text do not touch patients. They cannot smell infection or feel a spleen. Those sensory elements still matter.
Andreessen’s firm published its own podcast episode titled “Healthcare 2026: AI Doctors, GLP-1s, and Insurance Defection” in January. Guests discussed rising out-of-pocket spending on diagnostics and navigation services. Consumers appear willing to pay directly for faster, AI-supported insights when insurance falls short. That shift could accelerate adoption regardless of physician comfort.
Critics like AI researcher Gary Marcus have pointed out practical flaws. When Andreessen shared a detailed prompt instructing ChatGPT to “never hallucinate,” engineers noted the impossibility. Hallucinations stem from the model’s architecture. Instructions cannot eliminate them. Marcus called the episode “hilarious (and maybe a little bit scary).” The exchange underscored that even sophisticated investors sometimes overestimate control over these systems.
So what does this mean for the future of medical practice? Evidence suggests AI already augments certain tasks at expert levels. Diagnostic reasoning in controlled emergency scenarios. Literature synthesis. Patient communication drafts. Yet real-world deployment reveals persistent shortfalls in triage accuracy and treatment selection. The 99 percent claim lacks backing from peer-reviewed sources. Current data shows AI competitive with average performers in some areas while trailing specialists in others.
Doctors who integrate these tools effectively may outperform those who do not. A University of Virginia test from 2025 found physicians using ChatGPT Plus scored similarly to those using traditional resources. The AI alone scored above 92 percent on the same cases. The gap appeared tied to prompting skill. Humans still need to know when to trust the output and when to question it.
Policy conversations have begun. Regulators weigh how to supervise AI in direct patient care. Utah’s experiment offers one model. Others propose strict limits on standalone diagnosis. Insurance companies eye cost savings from automated triage. Patients simply want answers that work without unnecessary delays or errors.
The conversation Andreessen joined will not end soon. New models arrive monthly. Capabilities improve. Error patterns evolve. Today’s meta-analysis may look dated in six months. What remains constant is the need for rigorous testing in realistic conditions. Hype helps no one when symptoms are real.
Patients already blend sources. They ask ChatGPT for background, then check with their doctor. Or they skip the visit altogether when the bot sounds confident. That behavior carries risk. It also signals deep frustration with traditional access. Long wait times and brief appointments drive people toward whatever responds quickly.
Industry insiders watch closely. Venture funding flows toward companies promising AI physicians. Health systems pilot ambient scribes and decision support. Medical schools update curricula to teach prompt engineering alongside anatomy. The integration feels inevitable. The exact form and speed spark disagreement.
One thing feels clear from the accumulated research. AI will not replace doctors wholesale. It will change what doctors do. Routine information delivery may shift. Complex case management and human connection likely stay. Those who master the partnership stand to deliver better care than either party alone.
Andreessen’s provocation serves a purpose. It forces examination of assumptions. It highlights investment incentives. Most of all it underscores how quickly public expectations move ahead of validated performance. The studies keep coming. The technology keeps advancing. Patients and physicians navigate the tension together, one query at a time.
WebProNews is an iEntry Publication