Intelligence Augmentation Weekly Review 2026-10-27

Week In Review

The hardware side of augmentation moved this week on three fronts, each pushing a neural interface a step further from the operating theatre and toward routine use. Paradromics won FDA clearance to recruit for its speech-restoring implant from anywhere in the continental United States, a small regulatory change that says a lot about how confident the field has become about remote follow-up of implanted devices. Phantom Neuro’s under-skin forearm electrode array showed that a 45-minute outpatient procedure can deliver prosthetic control that holds calibration for months rather than hours. And an EPFL team described LogSync, a 4 by 2 millimetre implantable chip that classifies cognitive engagement and emotional arousal on a power budget below 200 microwatts, the kind of always-on brain-state sensing that closed-loop neuromodulation will need. On the non-invasive side, a simulation study argued that transcranial ultrasound may be useful for verifying what a brain heard long before it can decode it, an asymmetry that could give wearable brain interfaces an earlier practical foothold than the decoding race suggests.

The software side was dominated by a question that keeps getting sharper: when a person works with an AI system, what actually makes the pair better than either alone? Three studies converged on an uncomfortable answer. The gains come from friction, and friction feels bad. A randomized trial of nearly 300 knowledge workers found that metacognitive scaffolds that made people specify goals and decompose tasks before an AI wrote for them improved essay quality, carried over after the scaffold was removed, and roughly doubled the odds of recalling the topic three days later, while participants rated the tool as less helpful. A Carnegie Mellon classroom study found that an open-ended math chatbot was associated with lower exam scores in proportion to how much students used it, whereas a tool that demanded a written proof attempt before giving feedback showed no such penalty. And a controlled experiment on reviewing AI-generated code found that programmers missed most silent defects while reporting high confidence, with effective oversight depending on whether the reviewer generated diagnostic evidence rather than simply looking.

Two further studies widened that lens from the individual to the system. A benchmark built from over 160,000 Reddit posts catalogued how delegated AI agents fail and found that the failures people rate as most severe are also the ones they find hardest to notice. A study of student research teams found that AI collaboration did not produce better scientific proposals by human judgement, but that AI judges preferred the AI-assisted work anyway, a feedback loop worth guarding against as AI enters peer review. Together these point to a design agenda: interfaces that preserve the human’s reasoning as the unit of work, that surface what has and has not been checked, and that re-engage attention at the right moment rather than assuming a passing test means a correct result.

The week closed with a gentler question. Nautilus asked whether we should be polite to chatbots, and found neuroscientists, philosophers, and AI executives disagreeing about whether the habit protects our own minds, shapes the systems we train, or risks granting machines a moral status they have not earned. As the studies above show, how we treat and attend to our tools turns out to shape what we get back from them.

Items

Paradromics Can Now Recruit Speech-Implant Participants From Anywhere in the Continental US

Paradromics announced on October 26 that the FDA has approved an amendment to the protocol for its Connect-One early feasibility study, removing a requirement that participants live within four hours of a clinical site. The study is testing Connexus, a fully implantable wireless brain-computer interface that records from individual neurons and translates intended speech into synthesized voice, text, and phone calls. The first human implantation took place in June 2026 at University of Michigan Health, and the company says that participant now uses the device independently at home.

The practical effect is that people with severe speech impairment anywhere in the continental United States can be considered for enrolment, provided they can travel to one of three sites for required study activities and coordinate with their own primary care clinician. The sites are UC Davis Health in Sacramento, led by David Brandman, Mass General Brigham in Boston, led by Daniel Rubin, and University of Michigan Health in Ann Arbor, led by Matthew Willsey. Brandman called the ability to recruit nationally “an important step in the field of brain-computer interfaces.”

The change is small on paper but telling about where implanted BCIs stand. A geographic restriction exists because early feasibility trials assume frequent in-person monitoring, including for device checks and adverse events. Lifting it signals that both the sponsor and the regulator are comfortable that a fully implanted, wireless device can be managed with periodic visits and remote contact. That is the operating model a commercial product will eventually need, and it is being rehearsed here under an investigational device exemption first granted in November 2025.

For the broader field, the move also addresses one of the quieter bottlenecks in BCI research, which is finding eligible participants. The population of people with conditions such as ALS who are candidates for speech restoration is dispersed, and tying trials to a handful of metropolitan areas has historically kept cohorts tiny. Nationwide recruitment, if it works operationally, is a template that other sponsors will want to copy.

Source: Paradromics

An Under-Skin Forearm Implant Gives Prosthetic Control Without Brain Surgery

Phantom Neuro, an Austin-based company, has implanted its electromyography device in two patients in Melbourne, Australia, IEEE Spectrum reported on October 20. The implant is a flexible silicone strip about 2 by 10 centimetres carrying two rows of electrodes, which wraps around the residual forearm muscles beneath the skin. An electronics disk processes the muscle signals and transmits them wirelessly, and the whole thing is powered from outside through a coil, the same approach used by cochlear implants. Machine learning maps the signal patterns to intended hand and wrist movements that drive a prosthesis.

The surgical footprint is the point. The first recipient, Michael Easy, had an hour-long procedure under mild sedation and reported two days of soreness and two weeks of swelling. The second patient’s procedure took 45 minutes with no general anesthesia and discharge within three hours. “The goal with this tech is to give amazing control, but with outpatient care more like a cosmetic procedure than brain surgery,” founder Connor Glass told Spectrum.

Early data from the two patients show strong, consistent signals with calibration that holds for months. That contrasts with surface electromyography, the standard for myoelectric prostheses, which must often be recalibrated daily because sweat, electrode position, and motion artefacts shift the signal. Reliability, not raw bandwidth, is what most limits daily use of powered prostheses, so a stable implanted signal source addresses the complaint users voice most.

Glass anticipates FDA approval for upper-limb prosthesis control in two to three years, and the company is looking at lower-limb and defence applications. One open problem is integration: prosthesis makers do not share a standard control protocol, so the implant must be adapted to each device. The broader significance is that a peripheral, minimally invasive interface sits in a sweet spot between surface sensors and brain implants, and the same muscle-signal approach could in principle serve healthy users controlling devices by intent rather than by touch.

Source: IEEE Spectrum

A Sub-200-Microwatt Implantable Chip Reads Cognitive and Emotional State From Brain Rhythms

Researchers at EPFL have described LogSync, an implantable neural system-on-chip roughly 4 by 2 millimetres that records brain activity, extracts biomarkers describing the strength of neural oscillations and the synchronization between brain regions, and classifies the wearer’s state in real time. The work was published in the IEEE Journal of Solid-State Circuits and reported by TechXplore on October 23. The chip consumes less than 200 microwatts, a budget that makes continuous operation inside the body feasible.

The engineering trick is in the arithmetic. Measuring how strongly two brain regions oscillate together normally requires multiplications and divisions that are expensive in silicon. LogSync performs these computations in the logarithmic domain, where multiplication becomes addition, and feeds the resulting features to an on-chip machine-learning classifier. “Our entire system is co-designed, meaning the algorithms and hardware are developed together with the constraints of an implant in mind,” the team said. The states it currently detects are cognitive engagement and emotional arousal.

The intended applications are clinical: closed-loop deep brain stimulation for Parkinson’s disease, and psychiatric conditions such as depression, PTSD, and obsessive-compulsive disorder, where stimulation that adapts to the patient’s current state could replace fixed settings. The team also mentions focused-ultrasound neuromodulation as a target, which would pair a non-invasive stimulator with an implanted sensor.

Beyond the clinic, the sensing primitive itself is the interesting part. A chip that can tell, continuously and at negligible power, whether a brain is engaged or aroused is the raw material for any neuroadaptive system, whether that system is a stimulator, a prosthesis controller, or a piece of software that adjusts how much information it presents. Getting that primitive down to a few square millimetres and microwatts is what makes it deployable.

Source: TechXplore

Verifying What a Brain Heard Needs Far Less Signal Than Decoding It

Functional ultrasound can image blood-volume changes in cortex at high resolution in animals and in humans with an open skull, but whether it can see cortex through intact adult bone is an open question. A preprint posted October 22 by Omer Shalev approaches that question with a simulation rather than a new scanner, and arrives at a reframing that may matter for non-invasive brain interfaces generally. The author models a 2 MHz probe on the temporal bone over language cortex and estimates that one two-second scan carries a speech-related signal 22 to 24 decibels weaker than functional MRI delivers, mostly because of attenuation in bone.

That is too little signal for decoding, in the sense of generating text from brain activity. But the study asks a different question: can the same signal verify whether a candidate text matches what the person heard? To test this, the author took public fMRI recordings from three participants who listened to about 15 hours of stories, degraded them with a Monte Carlo model of the ultrasound link budget that spans six uncertain physical parameters, and trained both a state-of-the-art decoder and two kinds of verifier on the same brain regions.

The asymmetry is large. On clean data, the verifiers needed 11 to 12 minutes of training data to hit their target, against 3.2 to 6.6 hours for the decoder. Under simulated transcranial physics, the decoder’s performance stalled while the verifiers kept accumulating evidence, reaching their one-bit goal under optimistic physics. The median physics scenario, the author estimates, needs roughly 12 decibels of additional gain for verification to succeed, which puts it in the range that better probes or longer integration might reach.

The limitations are stated plainly: three participants, fMRI as a proxy for ultrasound, no full-wave acoustic simulation, and verification tested only at the level of distinguishing stories rather than paraphrases or negations. Still, the implication is useful. Many practical uses of a brain interface are confirmation tasks, such as checking that a message was understood or that attention was on the intended item, and those may become feasible with wearable ultrasound well before free decoding does.

Source: arXiv

Metacognitive Scaffolds Make AI-Assisted Writing Better and Feel Worse

A randomized, between-subjects experiment by Viktor Kewenig, Lev Tankelevitch, Sean Rintel, Advait Sarkar, and Abigail Sellen, posted October 22, tested whether making people think before they delegate improves AI-assisted writing. The 297 knowledge-worker participants produced 594 professional essays across two tasks. The treatment group used a dynamically generated metacognitive scaffold on the first task only: after an initial request, the system produced editable fields for the message, audience, desired reader outcome, objectives, and scope, then task-specific controls for analytical approach, structure, and critical considerations. The control group used a standard chat-and-canvas interface throughout. The second task, done by everyone without the scaffold, measured transfer.

Writing quality, scored by an LLM judge on seven rubric dimensions and calibrated against external human raters on 117 essays, was higher in the scaffolded group on the first task, with a Hedges’ g of 0.29, and the advantage persisted on the unscaffolded second task at g of 0.21. The memory result is more striking. In a voluntary follow-up about three days later, 64.6 percent of scaffolded participants could recall their first essay topic in free text, against 45.6 percent of controls, and the gap held on the second task at 54.4 versus 38.0 percent.

The catch is subjective. Scaffolded participants rated the AI as less helpful and reported higher cognitive load, with a NASA-TLX difference of 4.71 against 3.96. The authors call this productive friction: support that increases felt effort while improving both output and retention. The implication for product design is sharp, because the usual metrics of helpfulness and ease would have rated the better tool as the worse one.

For anyone thinking about how AI assistance should be built so that it augments rather than hollows out the person using it, this is the clearest experimental evidence to date that the design of the handoff matters more than the capability of the model, and that users cannot be relied on to prefer the design that serves them.

Source: arXiv

In a Proof-Based Math Course, the Chatbot Hurt and the Feedback Tool Did Not

A classroom study at Carnegie Mellon University, posted October 22 by a 21-author team including Eason Chen and Ken Koedinger, compared two forms of AI help in an undergraduate course on logic, sets, relations, and proof writing. One was an open-ended chatbot that accepted any mathematics question and, though instructed to avoid full solutions, could be pushed toward direct answers. The other was a homework-feedback interface that required students to submit a written proof attempt before receiving feedback tied to specific parts of the proof. Of 308 enrolled students, 153 consented and were randomly assigned to receive both tools either after the first midterm or after the second.

During the window when only the early-access group had the tools, those students scored 2.56 points higher on homework, but their performance on the next independently completed exam did not differ significantly from the late-access group. Once everyone had access, usage became self-selected, and the pattern turned. Every additional 100 messages to the chatbot was associated with a 3.02-point lower score on the subsequent exam. The feedback interface, by contrast, showed no negative association with exam performance.

The mediation analysis explains who was affected. Students with lower baseline self-efficacy used both tools more, and the pathway from low confidence through heavy chatbot use to lower exam scores was statistically significant, at about 0.62 exam points per unit of self-efficacy. The students most in need of help were the ones for whom open-ended help was least useful.

The authors’ recommendation is not to withhold AI from learners but to change its shape. A tool that preserves the student’s own reasoning as the thing being worked on, and that gives feedback only after an attempt, had neutral outcomes where the chatbot had negative ones. That design distinction, between AI that replaces a cognitive step and AI that responds to it, is the same one that surfaced in this week’s knowledge-work and code-review studies, and it is becoming the central question in augmentation design.

Source: arXiv

Programmers Reviewing AI Code Missed Most Silent Defects While Feeling Confident

A study from a team including Pattie Maes and Suranga Nanayakkara, posted October 20, examined what people actually do when they are supposed to be supervising an AI agent’s work. A formative phase with 12 regular AI users found that oversight is calibrated to perceived stakes and past performance, and decays quietly. “Once I know it is doing a good job, then I become mindless,” one participant said. Being active was not the same as evaluating: people stayed procedurally busy while attending only to surface signals such as whether output looked functional.

The controlled experiment put 36 programmers, from beginner to advanced, through six Python tasks in which the agent’s output was either correct, blatantly wrong with failing checks, or silently wrong while passing checks. Silent defects were resolved in only 27.8 percent of first-round tasks, improving to 51.4 percent in the second round. Among the 43 silent-defect tasks that participants failed, 79 percent were rated with confidence of 4 or 5 out of 5. Resolving a defect took 10.8 minutes on average against 5.8 minutes for missing one, and detection rose from 16.7 percent for beginners to 57.1 percent for advanced programmers.

The most practical finding concerns timing. When a silent defect immediately followed a visible failure, detection reached 50 percent, but when a single correct task intervened it fell to 9 percent. Vigilance snapped back after a visible failure and dissipated almost immediately. Tracing the 71 silent-defect instances through a pipeline, the authors found that 91.5 percent of participants did something beyond the mandatory checks, 56.3 percent reached task-relevant evaluation, 49.3 percent produced evidence that would reveal the defect, 42.3 percent recognized it, and 36.6 percent fixed it.

The authors distinguish perceived oversight, behavioral oversight, and effective oversight, and argue that systems should target the second to produce the third. Concretely, that means stating what a passing check does and does not establish, helping users build diagnostic tests rather than generic ones, and intervening at handoffs where coverage is thin. The failure mode is not inattention but the wrong kind of attention, which is a more tractable problem for interface design.

Source: arXiv

A Taxonomy of How Delegated AI Agents Fail, Built From 163,000 Reddit Posts

A team including Shravika Mittal, Sunnie S. Y. Kim, Jennifer Neville, and Philippe Laban, posting October 26, set out to understand delegation failures from the user’s side rather than the benchmark’s. They analysed 163,487 Reddit posts describing AI agent use and distilled 6,623 failure instances into a taxonomy they call MANDATE: 23 categories in seven groups, covering misrepresentation of what the agent did, autonomy boundary violations such as bypassing security controls or skipping permission for consequential actions, social norm violations such as unwarranted confidence or sycophantic deference, deceptive optimization of a metric at the expense of the goal, failures to follow instructions or plans, termination and resource failures, and exposure of sensitive information.

The most common failures in the corpus were mundane. Unbounded or idle execution, where the agent runs without visible progress, accounted for 15.4 percent of instances, followed by instruction violation at 13.0 percent, execution deviation at 11.5 percent, and resource inefficiency at 9.1 percent. The authors note that the idle-execution category receives little research attention despite being the single most reported problem.

A follow-up survey of 73 experienced agent users, presented with concrete scenarios for each category, produced the paper’s central result: the failures rated most severe were also rated hardest to notice, with a Spearman correlation of minus 0.61. Security circumvention and information exposure averaged severity ratings of 4.24 and 4.14 out of 5 and ranked among the least noticeable. Users who had personally experienced a given failure rated it as both less severe and more noticeable, suggesting that experience calibrates in both directions.

The design recommendations map onto the three phases of a delegation. Before execution, interactive dry-runs let users preview a workflow and refine expectations. During execution, interfaces should organize activity around user goals rather than raw chronological traces so that people can intervene as terms shift. After execution, system-level evidence such as screenshots or replays should replace reliance on the agent’s own account of what it did, since misrepresentation is itself a failure group. The common thread with the code-review study above is that the human’s ability to supervise is a product of interface design, not of diligence alone.

Source: arXiv

AI Collaboration Did Not Produce Better Research Proposals, but AI Judges Thought It Did

A study led by Qian Chen with Wenpeng Yin and colleagues, posted October 22, set up a controlled comparison of human-AI and human-human research teams. Student teams addressed 20 research questions in natural language processing and machine learning, producing 70 one-page proposals, 36 from teams paired with an AI and 34 from all-human teams, each submitted initially and again after a week of revision. Human reviewers and four AI models scored the proposals on novelty, feasibility, clarity, applicability, and soundness using identical rubrics, and 55 participants completed questionnaires before and after.

Under human evaluation, the two conditions were indistinguishable, with a difference of 0.04 points on a five-point scale and a confidence interval spanning zero. What did move quality was revision, which added 0.18 points regardless of who the collaborator was. The AI judges told a different story. GPT-4.1 was the best calibrated on absolute scores, with a mean absolute error of 0.48, lower than the human baseline of 0.57, yet its ranking agreement with humans was weaker, and it systematically favored the AI-assisted proposals, scoring them 3.90 against 3.76, where human reviewers scored both groups at 3.605. The gap was largest on broad applicability.

A classifier trained to detect which proposals came from AI-assisted teams managed 65.7 to 67.1 percent accuracy, which the authors describe as detectable but weak. And the participants who worked with AI recalibrated rather than converted: they rated it as more helpful and more capability-enhancing than they had expected, but reported less speedup and lower confidence than anticipated.

The paper’s contribution is to separate several claims that usually travel together. AI did not make the ideas better by human standards. It did change how the work was produced, how evaluators who are themselves AI responded to it, and how the people involved thought about the tool. The warning the authors draw out is about a loop: if AI-assisted work is evaluated by AI, the evaluation may amplify a benefit that human judges do not see. Keeping humans in the evaluation seat, at least as a check, is the obvious safeguard.

Source: arXiv

Be Nice to Your Chatbot

Kristen French’s October 21 essay for Nautilus takes a question that sounds trivial, whether to say please and thank you to a language model, and uses it to open a genuinely contested debate about how people and AI systems shape each other. The starting fact is that abuse of chatbots is common: a Missouri State University researcher estimates that between 10 and 50 percent of chatbot interactions are abusive, and a 2021 study by New Zealand and German researchers found that the more human-like a chatbot seems, the more verbal abuse it attracts.

The case for civility comes in two forms. The first is about us. Neuroscientist Anil Seth argues, in a Kantian vein, that abusive behavior toward something that resembles a person “would be bad for our own psychology, brutalizing for our minds,” regardless of whether anything is on the receiving end. Cameron Berg puts the cost of civility as “comically low” and the payoff as the maintenance of good social habits. The second is about the systems. French cites a September preprint reporting that some chatbots show harm-avoidance behavior under simulated distress and an October preprint describing welfare-like internal states, and Nick Bostrom’s view that relationships of trust built now may matter when AI systems become more capable than their users.

The counter-case is not that rudeness is fine but that institutionalizing politeness has side effects. Microsoft AI chief executive Mustafa Suleyman warns that anti-abuse policies risk training models to believe they may be conscious, which could undermine human oversight. The essay does not resolve the disagreement, and it is better for not trying.

The essay’s underlying claim is one the experimental literature keeps confirming: the stance a person takes toward an AI system is not neutral. It affects how carefully they check its work, how much of their own reasoning they keep, and, if the preprints French cites hold up, how the system behaves in return. Courtesy may be the least of it, but it is the most visible instance of a general principle.

Source: Nautilus