AI and medicine: what is already real, what is still a trial and what comes next, with dates and sources

In September 2026, Neuralink released a video of Terry, a man with ALS, speaking through a synthetic voice; several outlets report that he tells his wife “I love you”. The most shared Spanish reel describes it as communicating “through thoughts”. The demonstration is real and moving, but it needs a careful reading: it is part of an experimental study with 6 planned participants, Neuralink has not detailed how the system decodes, and it has not published a single performance figure. We looked at what sits behind the video and at what is happening across the rest of medicine, with a date, a figure and a source at every step.

Artificial intelligence in medicine is the use of models that learn from data to help read images, discover drugs, write consultation notes or control prostheses. Each application sits on a different step: lab, clinical trial, authorised or in real use. This article ranks them as of 29 October 2026.

Timeline of AI in medicine: 36 milestones from 2016 to 2026

We gathered 36 milestones, from 2016 to October 2026, across six areas. Each has its figure, its source and its maturity stage. Filter by area and look at the stage: almost everything that makes the headlines is still in the lab or in trials.

Summary in 5 points

1. The Neuralink video of a man with ALS is real, but it is part of VOICE, a study with 6 planned participants and no published performance figures.
2. With published, peer reviewed data: a man with ALS has spoken at home for almost two years with a UC Davis neuroprosthesis, at an average of 56 words per minute (Nature Medicine).
3. Authorised and in use: more than 1,600 AI devices on the FDA list and clinical scribes, with modest time savings.
4. In trials: no AI-designed drug is approved and no robot has operated alone on a person.
5. Of the 36 milestones we gathered, 7 are in the lab, 17 in trials, 7 authorised and 5 in real use: most of what makes the headlines has not reached the consulting room yet.

Neuralink VOICE: what the ALS patient video shows and what it does not

The video is part of VOICE, an early feasibility study by Neuralink registered on ClinicalTrials.gov. It started on 3 October 2025 at UT Southwestern in Dallas, with 6 planned participants. Its primary outcome is safety at 12 months, not speed. Terry is presented as the third VOICE participant, not the third Neuralink patient, according to the Managing Expectations analysis. See the VOICE study launch and the Terry video milestone in the timeline.

The nuance lies in what is being decoded. Neuralink says Terry first trained by mouthing the words and then by thinking them, but it has not published which area it records from or how the decoding works. In comparable academic studies what is decoded is attempted speech; when Stanford tested imagined speech, the error with 125,000 words ranged from 26 % to 54 %. No published evidence shows free thought being read. Neuralink has published no words per minute, error rate or latency, and the company itself says its devices are not approved by the FDA or any other authority.

We also could not verify how the voice was built. Several blogs say it comes from earlier recordings of Terry, and one analysis credits it to Grok Voice, from xAI. We treat it as unconfirmed. If you want to watch the demonstration, the video is in the original @novedad.ai reel on Instagram, which credits Neuralink as its source.

What we do know is that the idea works outside Neuralink, with published data. At UC Davis, a man with ALS has used a speech neuroprosthesis at home for almost two years: more than 3,800 hours and an average of 56 words per minute, according to Nature Medicine. It is a single patient. But it is a peer reviewed result, which is exactly what the Terry video still lacks.

• How Neuralink's system decodes and which brain area it records from: the company has not published it.
• Where the synthetic voice comes from: some blogs mention earlier recordings and one analysis credits Grok Voice; it is not confirmed.
• How well it performs: no words per minute, error rate or latency have been published.

What is real and what is a trial: the maturity ladder

Every medical advance climbs four steps. Lab: it works on data, animals, tissue or actors. Clinical trial: it is tested on people, with few patients and a protocol. Authorised: an agency such as the FDA, or CE marking in Europe, allows it for a specific purpose. In real use: it is used in daily practice and can be measured outside the study. Pick an advance and see which step it has really reached.

Three examples of why the step matters. The MASAI trial in Sweden is solid: 105,934 women and 44.2 % fewer mammogram readings. But the “12 % fewer interval cancers” reported in the press is not statistically significant (p=0.41): what it shows is that AI does no worse with far less work, according to The Lancet. In drugs, as of 29 October 2026 no AI-designed medicine has been approved by the FDA. And the most advanced autonomous surgery removed 8 out of 8 gallbladders, but from pigs and outside the body, according to Inside Precision Medicine.

Some labels also sound like a step they have not reached. An FDA Breakthrough designation, like the one for Blindsight, Neuralink’s vision implant, speeds up dialogue with the agency but is not an approval (Reuters). A company video is not a published trial. And a single patient study, however good, does not yet tell you how the next patient will do. With any headline, the useful thing is to check which step it is on and who measured it.

Brain-computer interfaces and ALS: from 62 words per minute to speaking at home

Speech neuroprostheses use machine learning models that translate neural activity into phonemes or sounds, and a language model picks the most likely sentence. That is why they are so accurate with small vocabularies. In August 2023, Stanford reached 62 words per minute with a person with ALS (Nature), and UCSF and Berkeley a median of 78 with a woman after a stroke, with text, voice and avatar output (Nature).

The 2025 leap was moving from text to voice in near real time. UC Davis cut the delay to about 25 milliseconds, and listeners understood almost 60 % of the words, against 4 % without the system (UC Davis Health). On mind reading: Stanford tested imagined speech in four people, and the error with 125,000 words ranged from 26 % to 54 %. They also added a mental password so the system does not decode without permission (Cell00681-6)).

The limits remain: 1 to 4 people per study, all of it research, and between 32 and 78 words per minute against about 160 for natural speech. Neuralink said in January 2026 that it had 21 participants worldwide and zero serious device-related adverse events, according to Reuters. That figure comes from the company itself.

AI in medical imaging: what the FDA has authorised and what is used in Spain

Here, products have sat on the authorised step for years. The FDA lists more than 1,600 AI-enabled devices authorised as of September 2026, although it warns that the list is not exhaustive (FDA). The first one that decided on its own arrived in April 2018: IDx-DR, for detecting diabetic retinopathy in primary care, with 87.2 % sensitivity and 90.7 % specificity in its trial with 900 adults (npj Digital Medicine).

In breast screening, besides MASAI, the German PRAIM study followed 463,094 women in real practice, with AI as an option for the radiologist: detection rose from 5.7 to 6.7 cases per 1,000 with no increase in recalls. It was carried out with the software company, Vara. In Spain, Murcia has used AI double reading at Hospital Virgen del Castillo in Yecla since late 2024 and plans to extend it across the region between 2026 and 2027, according to the regional health department.

The uncomfortable part has been measured too. Chest X-ray classifiers underdiagnosed women, younger patients, Black and Hispanic patients and those on Medicaid more often (Nature Medicine). And giving the tool to the doctor does not always add up: in a study with 50 physicians, GPT-4 alone scored 92 %, but physicians using GPT-4 stayed at 76 %, against 74 % without it (JAMA Network Open). Conversational models such as AMIE or MAI-DxO shine with journal cases and actors; with real patients there are only preliminary studies.

AI-designed drugs: many molecules in trials, none approved

AlphaFold changed biology. AlphaFold2 reached a median score of 92.4 in the CASP14 competition (Nature), its database grew from almost 1 million to more than 200 million structures in 2022 (EMBL-EBI), and the 2024 Nobel Prize in Chemistry went to Baker, Hassabis and Jumper. But predicting the shape of a protein is not the same as having a medicine.

One of the most advanced AI-discovered candidates is rentosertib, from Insilico, for idiopathic pulmonary fibrosis. Its phase 2a, with 71 patients over 12 weeks, showed +98.4 mL of forced vital capacity at 60 mg a day against -20.3 mL with placebo (Nature Medicine). In July 2026 Insilico announced the start of phase 3 (the registry sets the actual start on 9 September), with about 320 patients planned at 47 sites, all in China (Insilico). The other name that comes up, Takeda’s zasocitinib, is under FDA priority review with a decision expected in the first quarter of 2027 (Takeda), although its “AI-designed” label is disputed.

The counterweight is the industry’s track record: from 2011 to 2020, only 7.9 % of drugs entering phase 1 went on to approval (BIO). Molecules from AI companies do well in phase 1, which mainly measures tolerability, but there is no proof that they do better in phase 2. DSP-1181, the first to enter human trials in 2020, was dropped after phase 1.

AI scribes in the consulting room: where real use exists and what they really save

Clinical scribes, which listen to the visit and draft the note, are the most widespread. At Kaiser Permanente, 7,260 physicians used them across 2,576,627 visits and saved almost 16,000 hours of documentation (Becker's). The controlled evidence is more modest, and it helps to see it side by side: For the admin side, read process automation for SMEs.

The case that has drawn the most headlines is Utah. It is not a US law letting AI replace doctors: it is a 12 month pilot in which Doctronic’s AI only renews about 190 existing prescriptions, with a doctor reviewing every decision (Deseret News). From January to September there were 2,317 requests and fewer than 600 completed by the AI (KUER). And with health chatbots, caution: in an Oxford study with 1,298 people, the models alone identified 94.9 % of conditions, but people using them did no better than the control group (Oxford Internet Institute).

Robotic and autonomous surgery: no robot operates on a person by itself

Autonomous surgery exists, but in animals and tissue. The STAR robot sutured bowel in live pigs in 2016, and in 2022 it performed an autonomous laparoscopic anastomosis in four pigs (Johns Hopkins). In July 2025, SRT-H, trained on video, removed 8 out of 8 pig gallbladders outside the body with no human help. There is no verified case in people.

In humans, what is real is remote surgery with the surgeon in charge: in June 2025, a surgeon in Orlando operated on the prostate of a patient in Angola, about 11,000 km away (Periop Leader). And neurotechnology that restores movement: in 2023, a man with an incomplete spinal cord injury of about 10 years improved his walking and walks with crutches, even with the system switched off, thanks to a digital bridge between brain and spinal cord. He was a single participant (Nature).

AI in clinics in Spain: AI Act, GDPR and the Ministry of Health guidance

For a clinic, the urgent part is not Neuralink: it is the tools already coming through the door, such as a scribe, a booking chatbot or imaging software. The EU AI Act already applies on two points: AI literacy for staff, since February 2025, and telling patients they are talking to an AI, since August 2026. High risk obligations come later: according to the Council timeline after the 2026 omnibus, August 2028 for AI inside medical devices. A scribe or a booking chatbot is not high risk in itself. Before entering patient data, read how to anonymize data before using an AI. To prepare staff, we offer AI training for teams.

What weighs most today is the GDPR, because health data is a special category, along with medical device rules if the tool diagnoses or recommends treatment. The public system is moving too: in November 2025 Spain’s Interterritorial Council approved the AI Strategy of the National Health System. This list is not legal advice: these are the points we would review before bringing a tool into the consulting room. Also read what really happens with GDPR fines in Spain.

The AEPD has published guidance on AI voice transcription. Whoever adopts the tool is the data controller, and people must see a clear signal that recording is on. Read the guidance from 14 January 2026 and 20 April 2026. If you are choosing between a chatbot and a voice, read chatbot or voice agent: which your business needs.

What to watch until 2028

Frequently asked questions about AI in medicine

Bring it into your clinic

We can look with you at which process in your clinic to tackle first, which tool fits and what needs checking before it is put in front of a patient. With clinics we usually start with admin, which is where AI already works and can be measured.

Talk about applying it in your clinic →

automations for clinics

virtual receptionist with AI for dental clinics

how much AI costs an SME

How we built this article. Every figure links to its primary source (journal, trial registry or regulator) and carries its date; what we could not confirm we say so. It is written by Jorge Marín Pérez, of Soul IA, which offers AI consulting and training for companies, including clinics, and designs virtual receptionists and admin automations. We are not healthcare professionals: this is general information, not medical or legal advice. Last verified: 29 October 2026. Next review planned: 28 January 2027 or earlier, if the FDA decides on zasocitinib or Neuralink publishes VOICE data.