Intelligence Augmentation Weekly Review 2026-11-17
Week In Review
The week’s brain-computer interface news came from two directions at once: the laboratory and the people who volunteer for it. At the Society for Neuroscience meeting, three groups reported communication implants that are becoming expressive, durable, and portable, including a synthetic voice that carries pitch well enough for a participant to sing and a brain-to-text decoder running on a phone (Brain-computer interfaces: A new voice for people with paralysis). In the same week, the trial participants who make such results possible published their own set of demands, asking for co-authorship, access to their neural data, and guarantees about what happens to their implants when a study ends (More autonomy, fewer NDAs: Volunteers have new requests for brain implant trials). A review of non-invasive interfaces in dementia care rounds out the picture, with modest but real gains in attention and memory from neurofeedback training and a frank accounting of how small the trials remain (Brain-computer interfaces in dementia). Together these suggest a field maturing on the engineering side while its social contract is still being negotiated.
A second cluster asked a sharper question than usual about AI advice: not whether people use it, but what makes it genuinely useful. Three preprints from human-computer interaction groups each isolated a different variable. One found that structured, multi-round deliberation between language models helps human fact-checkers more than a single AI answer does, while a “devil’s advocate” persona hurts (Understanding the Effects of Multi-Agent Deliberation for Human-AI Decision-Making). Another put a language model into three-person hiring discussions as a facilitator with no knowledge of the right answer, and nearly doubled the share of groups that found it (LLM Facilitation of Group Decision-Making). A third separated whether an advisor reaches your preferred conclusion from whether it reasons the way you would, and found that many people prefer an advisor that disagrees with them for reasons they respect (Right Advice, Wrong Reasons). The cautionary counterpart came from a four-experiment study showing that without outcome feedback, people keep following an assistant that is wrong every time (Do People Trust AI Assistants Even When They Are Consistently Wrong?). Read together, the lesson is that the design of the interaction, not the raw capability of the model, decides whether human judgment is amplified or merely displaced.
The remaining items apply the same test to learning and to physical skill. A preregistered experiment with 400 adults found that an LLM tutor that walks a learner through correcting their own error beats one that explains the error, however personalized the explanation (Turning Errors into Successes). A design-consultancy veteran laid out a practical framework for when AI belongs in a product at all, with physical controls kept for anything done “instantly, by feel, or under stress” (When Does It Actually Make Sense to Put AI Into Products?). And a San Francisco startup is building a robotic catheter system that lets a stroke surgeon operate across a city, with computer vision absorbing the spatial disorientation that makes the procedure hard to learn (Can Robots Bring Stroke Surgery to Remote Hospitals?). In each case the technology works best when it takes over the part of the task that was never the point and leaves the human in command of the part that is.
Items
Brain-Computer Interfaces Gain Expression, Endurance, and Portability at Neuroscience 2026

Three research groups presented communication-restoring brain implants at the Society for Neuroscience’s annual meeting, and each addressed a different limitation that has kept such systems in the laboratory. According to the Medical Xpress report on the presentations, the studies covered expressive speech synthesis, long-term independent use by a completely locked-in person, and a decoder compact enough to run on a smartphone.
The University of California, Davis group, represented by Maitreyee Wairagkar, showed a system that produces a synthetic voice capable of carrying changes in pitch. Earlier speech decoders produced flat, monotone output that conveyed words but not emphasis or emotion. The report says the added prosody improved how well listeners understood the output, and that one participant was able to sing with the device. Pitch is a large part of how meaning is carried in spoken language, so recovering it matters for more than musicality.
Daniel Rubin of Massachusetts General Hospital reported that a participant in a completely locked-in state, unable to move or speak at all, has used an implant that decodes intended movement from brain activity to communicate independently for more than a year. Duration is the quiet test for these devices. Signals drift, electrodes degrade, and decoders need recalibration, so a year of unassisted use is evidence that the engineering problems of stability are becoming tractable.
Chethan Pandarinath’s team at Georgia Tech and Emory University presented a portable brain-to-text system that achieves accurate, efficient decoding while running on an iPhone. Moving the computational load from a rack of equipment to a pocket device is what turns a laboratory demonstration into something a person could carry through their day. The report notes industry projections that the brain-computer interface market will grow from about $1.7 billion in 2022 to $6.2 billion by 2030, a figure worth treating cautiously but consistent with the pace of the clinical results.
Source: Medical Xpress
Brain Implant Volunteers Publish Their Own Terms for Participation

The people who receive experimental brain implants have begun to organize, and this week they published their first formal recommendations for how trials should treat them. Jessica Hamzelou reports in MIT Technology Review that the BCI Pioneers Coalition, cofounded in 2022 and led by Ian Burkhart, released a living document covering 13 categories of guidance, including compensation, privacy, ownership of brain data, and support after a study ends.
The requests are concrete. Participants want to be recognized as research collaborators, potentially with co-authorship on the papers their brains make possible. Burkhart describes taking part in his own trial two or three days a week, four hours at a time, a commitment the coalition argues deserves fair pay. They want unrestricted access to their own neural data and an end to non-disclosure agreements that prevent them from sharing that data with their personal physicians.
The sharpest demands concern what happens when a trial winds down. The coalition asks for extended device maintenance after studies conclude, adequate funding for removal surgery, and advance notice before an implant is taken out. The article recounts the case of Rita Leggett, an Australian epilepsy patient whose implant was removed against her wishes after the manufacturer went out of business, a situation ethicists have described as a potential breach of human rights.
The document is explicitly “version one,” and Burkhart plans to present it at an upcoming neuroscience conference. The timing is significant. As implant trials multiply and companies compete for participants, the volunteers are the scarce resource, and a shared set of expectations gives them leverage they have not previously had. It also gives the field a chance to set norms before the first large commercial rollouts make them harder to change.
Source: MIT Technology Review
A Robotic Catheter Could Let Stroke Surgeons Operate From Across the City

Thrombectomy, the removal of a clot from a brain artery, is one of the most effective treatments in medicine for acute ischemic stroke, and most eligible patients never receive it. Max G. Levy reports in IEEE Spectrum that only about one in ten candidates get the procedure, that one in five Americans lives more than an hour from a center that performs it, and that roughly 5 million people a year worldwide suffer an ischemic stroke. Remedy Robotics, a San Francisco company cofounded by surgeon David Bell and engineer Jake Sganga, wants to let a specialist perform the procedure remotely.
Its N1 system is a steerable catheter driven by a robotic arm with seven motorized modules. The interesting part is not the motors but the software. Catheter navigation is normally guided by two-dimensional X-ray images, and the catheter rotates as it advances, so a surgeon’s command to turn left may send the tip somewhere else entirely. The N1 tracks patterned markings on the catheter with computer vision, overlays a CT scan onto the live X-ray feed, and translates the operator’s directional commands into the correct physical motion automatically. The operator uses a joystick-style controller. According to the article, the company’s internal comparisons found the robot collided with vessel walls less often and finished faster than human-guided catheters.
In September 2025 the company ran its first human tests in Toronto. A neurosurgeon operated remotely on ten consenting patients undergoing non-emergency angiography procedures: six in an adjoining room, two elsewhere in the hospital, and two at a separate hospital across town. The article reports these procedures completed in roughly half the usual time. The system’s target latency is under 150 milliseconds, below human reaction time, and operators have shown they can compensate for worse links by slowing down.
The obstacles are substantial and the founders are candid about them. Rural hospitals often lack the network quality, CT scanners, catheterization teams, and intensive care units the procedure requires, so mid-sized regional hospitals are the realistic first targets. No company has full U.S. regulatory approval for remote surgery. Bell himself raises the questions of skill erosion and platform dependency among trainees, and emergency physician Jasdeep Bahra warns that telesurgery could concentrate expertise in wealthy centers unless the systems also train local clinicians. Remedy has raised $62 million in new funding, including from the University of California system, with on-site elective trials beginning within a year and remote procedures planned for late 2027.
Source: IEEE Spectrum
An AI Facilitator With No Answer Key Nearly Doubled Group Decision Accuracy
Groups routinely fail at a task they should be good at: pooling information that individual members hold privately. The classic “hidden profile” experiment gives each person a different slice of the facts about a set of options and watches the group converge on the wrong one because people discuss what they already share. A team including José Córdova, Chun-Wei Chiang, Miriam Schirmer, Ming Yin, and Emőke-Ágnes Horvát tested whether a large language model could fix this, not by knowing the answer but by running the meeting well.
The study put 414 participants into 138 three-person groups and gave them a hiring decision in an online chat, with some candidate information shared and some private. Groups were assigned to no facilitation, individual-level LLM assistance, or group-level LLM assistance. The model’s interventions were drawn from decades of group-process research: it summarized the discussion, nudged quiet members to contribute, and played devil’s advocate. Critically, it did not know which candidate was correct, so any improvement had to come from better process rather than leaked answers.
The effect was large. According to the preprint, groups without facilitation chose the correct candidate 32.6 percent of the time, while LLM-facilitated groups reached 56.5 percent. Facilitated groups exchanged more consensus messages, and the group-level condition in particular drew out significantly more disclosures of the private information that pointed to the right candidate. Individual-level facilitation reduced redundant repetition of already-shared facts. Participants also rated the group-level version as more usable.
The result reframes what an AI assistant in a meeting might be for. The temptation is to have the model produce the answer, but here it produced a better conversation, and the humans produced the answer. That design is robust to the model being wrong about the substance, since it never asserts any, and it leaves the judgment and accountability with the people in the room.
Source: arXiv
When Multiple AIs Debate, Humans Decide Better, but Not for the Reasons You Might Expect
Multi-agent deliberation, in which several language models argue a question across several rounds before producing an answer, has become a popular way to boost AI benchmark scores. Harry Yizhou Tian and Ming Yin asked a different question: does showing that deliberation to a person help the person decide better? Their two user studies, on verifying LLM responses and on fact-checking claims, suggest the answer is yes, but with a twist about what is doing the work.
The researchers deliberately pulled apart the two defining features of the approach, the presence of multiple AIs and the use of multiple rounds, and tested them separately against ordinary single-shot AI assistance. Deliberation outperformed single-round assistance on both tasks. But the gain came mainly from the multi-round structure, the iterative refinement of an answer, rather than from having a second model in the conversation. A second AI added value chiefly when the first one had room to improve.
Some popular design choices backfired. Giving one agent a devil’s advocate persona reduced both accuracy and user experience. Asking the human to participate actively in the deliberation improved accuracy in one task but raised frustration, a reminder that engagement has a cost and that a design that is better on paper may not be one people will tolerate.
The study is useful precisely because it resists the assumption that whatever improves a model’s score will improve a human’s decision. Human-AI decision support has its own dynamics, including attention, trust, and fatigue, and the authors lay out when, why, and how deliberation can be deployed so that those dynamics work in the user’s favor.
Source: arXiv
People Prefer an AI Advisor That Disagrees With Them for the Right Reasons
Alignment is usually discussed as a single property: does the AI do what you want? A preprint from Eunkyu Park, Vasudha Varadarajan, Wesley Hanwen Deng, Anisha Reddy, Gunhee Kim, Motahhare Eslami, Jana Schaich Borg, and Maarten Sap splits it in two. Decision alignment is whether the advisor recommends the option you would have picked. Reasoning alignment is whether the justification reflects the values you hold. The two can come apart, and the study shows that people care about the difference.
In a within-subjects experiment, 121 participants evaluated three AI advisors with different alignment profiles across eight everyday moral dilemmas, then chose which advisor they would want for various tasks. When the goal was to get advice, the single most popular profile, preferred by 40 percent, was the advisor that reached a different conclusion than the participant would have but reasoned in a way they endorsed. People, in other words, were willing to be told they were wrong by something that thought the way they did.
That preference was not stable across contexts. When participants were asked to pick an advisor for new tasks, preferences shifted toward decision-aligned advisors, the ones that would simply agree. And the willingness to hand over control fell sharply as the AI’s autonomy rose. Opt-out rates were 12 percent when the AI was merely consulted and 38 percent when it was allowed to act on its own.
The finding complicates the common assumption that aligning an AI’s values with a user’s will make the user comfortable delegating to it. Shared values earn a hearing, not a mandate. For anyone designing an assistant meant to improve decisions rather than ratify them, the result is encouraging: there is an audience for well-reasoned disagreement, as long as the person stays in charge of the final call.
Source: arXiv
Without Feedback, People Keep Following an AI That Is Always Wrong
The usefulness of an assistant depends on users trusting it roughly in proportion to its reliability. Cameron R. Jones, Julia Wong, Aaron S. Cohen, James Woolley, and Benjamin K. Bergen tested whether that calibration happens in practice by running four experiments in which conversational AI assistants were made consistently accurate, consistently wrong, or right half the time, and then watching how people’s reliance evolved across repeated interactions.
When participants received feedback on whether each piece of advice had been correct, calibration worked. According to the preprint, they followed the accurate model 67 percent of the time and the inaccurate one 21 percent. Without feedback, the pattern inverted: people followed the always-wrong assistant 48 percent of the time and the always-right one 38 percent. Initial reliance on AI was high and dropped only when evidence of poor performance accumulated, and arguments grounded in truth offered little advantage in persuading people.
The comparison with human advisors is striking. Participants reported trusting AI advisors more than human ones, at 83 percent versus 65 percent. That baseline deference is reasonable when a system is usually right and dangerous when the user has no way to check.
The practical implication is not that people are gullible but that feedback is a design requirement. Many of the settings where assistants are now deployed, from medical triage to legal research to strategic planning, do not deliver quick ground-truth verdicts on each answer. The study makes a case that systems in such domains should manufacture feedback where they can, through spot checks, confidence signals, or verifiable intermediate steps, so that users have something to calibrate against before the stakes are high.
Source: arXiv
An LLM Tutor Works Best When It Makes the Student Fix the Mistake
The obvious use of a language model in education is explanation: a student gets something wrong, and the model produces a clear, personalized account of why. Michael W. Asher, Meng Cao, and Paulo F. Carvalho ran a preregistered online experiment to test whether that is actually the best use, and the answer favors a slower, older pedagogy.
The study enrolled 400 adults practicing the setup of constraint equations, a skill that is procedural, teachable, and easy to get wrong. After an error, participants received one of four responses: the correct answer only, a generic explanation, a personalized explanation generated by a language model, or Socratic tutoring in which the model led them through a scaffolded dialogue toward correcting the mistake themselves.
Socratic tutoring won on every measure that mattered. Participants in that condition resolved 64 percent of their errors through dialogue and scored 0.24 standard deviations higher than the explanation groups on posttest questions aligned with the practice material. They also found the sessions more interesting, less frustrating, and less costly, despite the sessions taking twice as long. Personalized explanations showed no advantage over generic ones, and neither kind of explanation reliably beat simply showing the correct answer.
The authors connect the result to mastery-experience principles: learning sticks when the learner does the correcting. For AI in education this is a clarifying finding. The capability that language models are best at, fluent explanation, is not the one that produces learning, while the capability that requires patience and restraint, asking the next question and waiting, is. Tutoring systems built around the second will look less impressive in a demo and work better in a classroom.
Source: arXiv
A Designer’s Test for When AI Belongs in a Product

As companies rush to add AI to everything from toothbrushes to refrigerators, IEEE Spectrum’s Stephen Cousins interviewed Savannah Kunovsky, managing director at the design consultancy Ideo and head of its Emerging Tech Lab, about how to tell a genuine improvement from a gimmick. Her framework is simple and ordered: understand who the product is for and what they are trying to do, then assess technical feasibility, then consider the business case. The governing question is whether the feature helps someone do what they were already trying to do.
The interview’s examples of failure are instructive. The Humane AI Pin asked users to abandon a familiar, fast smartphone interface for a slower and less reliable one. A smart cooktop prototype revealed in testing that users instinctively reached for physical knobs that were not there when something went wrong. Touchscreen climate controls in cars replaced tactile controls and forced drivers to look away from the road. Kunovsky’s rule for when physical controls should stay physical is memorable: anything you need to do instantly, by feel, or under stress.
Her examples of success share a pattern. Hearing aids that use AI to isolate a voice in a noisy room, the Be My Eyes service that describes a scene for blind users, and earbuds that translate speech in real time all extend a human capability that was limited, rather than substituting a novel interface for one that worked.
The design principles she lays out apply well beyond consumer gadgets. Offer multiple ways to accomplish a task, including voice, app, and a physical switch. Make manual override easy, because it builds trust. Be clear about when the system is exploring versus behaving predictably. Anticipate needs without becoming intrusive. And because generative AI produces nondeterministic output, she argues it requires far more design intention than conventional software, with explicit boundaries and evaluation frameworks in place of a fixed specification.
Source: IEEE Spectrum
What Non-Invasive Brain Interfaces Can and Cannot Yet Do for Cognition in Dementia
A review published in the journal Human Brain and released through EurekAlert surveys the state of brain-computer interfaces in dementia and mild cognitive impairment, covering the neurophysiological mechanisms, the clinical applications attempted so far, and the barriers to translation. It is a sober document, and its value lies as much in its accounting of limitations as in its catalog of promising results.
On the positive side, the review describes EEG-based neurofeedback that targets theta and alpha rhythms to address attention and memory, and beta rhythms for executive function. Gamma-frequency entrainment shows exploratory promise in Alzheimer’s disease. The P300 event-related potential, a well-characterized brain response, has potential as a marker for detecting early cognitive decline, and portable EEG platforms have proven feasible for rapid screening. BCI-based cognitive training, the authors write, has produced modest gains in attention, visuospatial ability, and memory, occasionally exceeding what music or physical exercise interventions achieve.
For later disease stages, the review proposes communication aids and environmental control systems, with conditioning-based paradigms suited to patients whose executive function has declined too far for more demanding interfaces. Closed-loop stimulation that combines interface readout with transcranial or deep brain stimulation remains investigational.
The limitations are stated plainly. Most clinical studies are small pilots, typically under 30 participants, with short follow-up, heterogeneous outcome measures, few active control groups, and rare blinding. Ethical concerns about informed consent, neural data privacy, and equitable access are unresolved. The authors call for standardized protocols, large randomized controlled trials, and hardware refinement. Taken as a whole, the review is a map of a field in which the mechanisms are plausible, the early signals are real, and the evidence base has not yet caught up to either.
Source: EurekAlert