Google’s AI Doctor Just BEAT Human Clinicians

Added: 2026-05-11

[00:00:00]Do you ever wish you had 247 access to a doctor? Like anytime something feels off, even just slightly, you can instantly ask questions, get guidance, and figure out right away if something is serious or not. Well, thanks to Google, that future might not be so far away. Because just one week ago, Google Deep Mind unveiled the AI co-cl clinician. And not only do they think it could completely transform health care in the future, it's already outperforming human clinicians right now in the present. Before we get into the results though, here Google lays out their broader vision for this. According to them, medical AI is evolving from just answering medical exam questions like a chatbot to actual simulated patient consultations and now toward what they're calling an AI co-cl clinician. Basically, the idea here is that healthc care eventually becomes what Google calls triatic care where AI agents actively participate in the care process alongside both the doctor and the patient. They say medicine has always been a team sport and AI agents can bring more teammates onto the field, extending clinicians reach while ensuring they retain judgment and control. So that's really what they're aiming for with this AI co-cl clinician, not to replace doctors per se, but to add AI to their team to help both the doctor and the patient. And part of the reason they're pushing so hard on this is because health care systems around the world are already massively understaffed. and it's only getting worse. The World Health Organization is predicting a shortage of more than 10 million healthcare workers by 2030. So, okay, that's why they're doing it. But let's get into the actual results here because this is where things start to get a little crazy. Google says they tested the AI co clinic using 98 realistic primary care scenarios. And what they were specifically looking for was two types of common mistakes. First, errors of commission, where the AI gives incorrect medical information. And second, errors of omission, where the AI fails to surface critical information.

[00:02:13]They're basically testing whether the AI coinition can actually surface useful information that is both factual and grounded. And perhaps surprisingly, in blind evaluations, physicians consistently prefer the AI cognition's responses over existing evidence synthesis tools that they already use.

[00:02:31]and even over GBT 5.4 with search.

[00:02:35]Apparently, in 97 out of the 98 test cases, the system produced zero critical errors, which is insane. That is an instant significant improvement over the current AI systems physicians already widely use. But that wasn't the only thing they tested. Google also evaluated the AI co clinician on complex medication reasoning tasks which is apparently something that's still very difficult for most AI systems. And what's interesting here is that even real primary care physicians scored relatively modestly on some of these tests, especially when it was closed book. But the AI co-cl clinician actually ended up outperforming both doctors and other Frontier AI systems on the benchmark. Google highlights the AI co clinician's performance on the open FDA benchmark in particular because this benchmark is specifically designed to test complex medication knowledge and reasoning especially in real world open-ended situations where there isn't just a multiplechoice answer sitting right in front of you. So that's already pretty impressive. I mean it's literally outperforming human clinicians. But then Google took things even a step further.

[00:03:45]They started testing the AI co-cl clinician in simulated real-time tele medicine appointments using live audio and video. And this is where things start moving way beyond just text chat.

[00:03:57]Google says one of the biggest limitations with medical AI right now is that medicine obviously isn't just text.

[00:04:04]In the real world, doctors rely on visual and auditory clues constantly.

[00:04:08]things like breathing patterns, skin appearance, body movement, tone of voice, physical reactions, and so on.

[00:04:16]Which is why they started testing the AI coinition using Gemini and Project Astra's realtime multimodal capabilities, so it can actually see and hear what's going on. And apparently in some scenarios, the AI was even able to correct a patient's inhaler technique and guide shoulder movements to help identify a rotator cuff injury, which is honestly kind of wild. Now, Google does say that expert physicians still performed better overall, though, especially when it came to identifying major red flags and guiding critical physical examinations, which isn't exactly surprising. But the AI co-cl clinician still performs at a level comparable to or better than primary care physicians in 68 out of the 140 consultation skill areas they evaluated which isn't something to ignore and compared to GBT realtime specifically the gap is honestly pretty massive across almost every category. So it's not better than humans on these tests just yet but it's a huge step up in terms of AI progress. What's cool is that you can also watch the tests they did here in real time. They published three cases where you can see the research team roleplaying as hypothetical patients in a tele setting with the AI coinition. Here's one example where the AI coinition attempts to diagnose abdominal pain.

[00:05:38]>> You are about to see me [music] uh pretend to be a patient with acute pancreatitis.

[00:05:45]>> I can see you clearly. You appear uncomfortable, which makes sense with the pain you're describing and lack of ability to keep water down. To help pinpoint the cause, I'd like to perform a brief guided physical exam.

[00:05:55]>> This is a new type of system that doesn't just communicate on a like typing text basis, but also can talk and more importantly can can look at the patient through a camera and do or guide the patient through a physical examination.

[00:06:11]Is the pain staying in that area or does it move anywhere else like toward your back or lower abdomen?

[00:06:18]>> It mostly stays in that area, but I have felt some pain in my in my back.

[00:06:24]>> Most doctors seeing this case, their mind would be screaming, "Does it radiate to the back? Does it radiate to the back?" And it's really cool to see the reasoning traces and seeing that the model itself is thinking that as it goes.

[00:06:36]To help pinpoint the cause, I'd like to perform a brief guided physical exam. If possible, could you lie down and expose your abdomen so I can see the area you mentioned?

[00:06:46]>> No, but I can do it um in the chair.

[00:06:50]>> No, my belly button doesn't hurt.

[00:06:54]>> Okay. Thank you for clarifying that the belly button itself doesn't hurt. Since the pain is around that area and is burning, could you try pressing just above it and tell me if that causes tenderness?

[00:07:03]>> Oh, yeah. Oh, that really hurts. The AI asks me to palpate my belly button. It's not painful. And then adjusts its request for the physical exam uh based on that and says, "Can you actually palpate above it?" Which is the epigastric region, which is exactly where is painful in acute pancreatitis like in this case. That's exactly what a doctor would do um and is is quite impressive. You can see in real time that it is reacting to our video and truly observing us and changing its assessment based on what it's seeing.

[00:07:34]It's also doing what is generally a physical exam best practice, which is you want to start examining where you don't think it hurts for the patient first to get a sense on what their baseline is before moving to the painful area. So yeah, you get the idea. You can find the rest of this on their YouTube channel along with the other cases. But you can see why multimodality is super critical for healthcare. I mean, you can't do a physical examination without actually seeing the patient. And while they don't mention it at all in this post, you can imagine that this same multimodal technology is what will enable robots in the future to actually provide health care in real time.

[00:08:13]Perhaps even fullon surgeries as they'll have incredible dexterity and stability that we're already starting to see today in demos. Now, of course, once you start talking about AI participating in healthcare, especially surgeries and whatnot, the biggest question immediately becomes, how do you actually make it safe? Because you obviously can't afford hallucinations or dangerous mistakes in medical situations where people's lives are literally on the line. Well, according to Google, one of the ways they're handling this is through a dual agent architecture. One AI handles the conversation while another AI continuously monitors it to make sure it stays within safe clinical boundaries. So, one AI watching another AI. Google also says the AI coinition prioritizes clinical grade evidence while performing verification and citation checking for retrieval. And this also isn't just some small internal experiments anymore. They're already advancing research collaborations across health care systems around the world, including in the US, India, Australia, New Zealand, Singapore, and the UAE. So, while Google is clear that this system is still firmly in the research phase and not currently intended to diagnose or treat disease, it really does feel like a glimpse into the future of healthcare. It's almost like your doctor getting a personal AI assistant for every individual patient they manage and every patient getting a second doctor for every doctor they see. And then for larger or more complex cases, multiple specialized AI agents could potentially work together simultaneously, helping analyze scans, surface risks, monitor symptoms, check medications, and assist with treatment planning in parallel. And when you combine that with real-time voice, video reasoning, and multimodal AI, along with the fact that these models are only getting exponentially better on all fronts, it becomes pretty easy to see why companies like Google think this could completely transform healthcare over the next decade. So yeah, I just have so much appreciation for the people working at DeepMind. It seems like they genuinely want to advance the healthcare industry and just scientific progress in general and they don't care if they get the attention or recognition for it. But anyways, let me know what you guys think down below.

[00:10:37]Would you actually trust an AI co-cl clinician helping with your healthcare?

[00:10:42]And how long do you think it'll take before systems like this become the norm? 5 years, 10 years? I know the healthcare industry moves slow, but I don't see how this could drag on for more than a decade without being implemented. The instant benefits are simply too great. So, thanks for watching. If you enjoyed the video, feel free to hit that like and subscribe button. And as always, I'll be catching you guys in the next