Intelligence Augmentation Weekly Review 2026-07-14

Week In Review

The past week produced a striking symmetry between two ends of the human-AI interface. On the invasive end, Neuralink disclosed a new transdural surgery that leaves the brain’s protective outer membrane intact, and CNBC reported that Chinese competitor BrainCo is betting on a very different theory of change — that the market for cognitive augmentation belongs to non-invasive wearables. Those two philosophies now define the field’s near-term axis: how deep should the electrode go before the benefit stops justifying the surgery?

On the software end, OpenAI moved GPT-5.6 into Microsoft 365 Copilot as the default model, updating the AI that hundreds of millions of knowledge workers now touch every workday. That change lands into an environment that Microsoft’s own 2026 Work Trend Index describes as a shift from assistants to agents — from tools that answer to tools that act, with human oversight moving from every step to periodic review.

Meanwhile, the research community held its annual reckoning at HHAI 2026 in Brussels, where hundreds of hybrid-intelligence researchers presented work on systems designed to enhance rather than replace human judgment. Clinical BCI progress continued in parallel: Paradromics has now implanted its first participant in the Connect-One study, and MIT Technology Review documented that BCI human trials have moved from single-digit patients per company to dozens across a growing cluster of firms.

Two research threads tied the week together. A Nature Machine Intelligence study reported that the brain processes spoken language in stages that closely mirror the layer-by-layer computations of large language models — a finding that a Frontiers review argues could accelerate the decoding of inner speech from non-invasive signals. And the long-running BrainGate consortium published data on a person with ALS using an implant at home, unsupervised, over thousands of hours — the clearest evidence yet that BCI has crossed from lab demo to daily-life tool.

Items

GPT-5.6 Becomes the Default Model Behind Microsoft 365 Copilot

On July 9, OpenAI announced that GPT-5.6 is now the preferred model powering Microsoft 365 Copilot inside Word, Excel, PowerPoint, Chat and Cowork — a rollout that affects the AI experience of every enterprise customer on the world’s most-used productivity suite.

OpenAI’s positioning of the model emphasizes efficiency rather than headline capability: “more useful work from every token,” stronger performance per dollar, and the ability to spend more compute on demand when a task warrants it. In Word, Microsoft says the model reduces the number of prompt-and-revise rounds needed to reach a clean draft. In Excel, it moves faster from raw data to interpretable insight while using fewer tokens per query. The most consequential integration is in Copilot Cowork, where GPT-5.6 handles multi-step agentic tasks — planning across files and tools to complete work end-to-end from a stated outcome.

The announcement lands in the middle of persistent reporting that the Microsoft–OpenAI relationship is straining. That business-relationship subtext aside, the practical significance is that a new default LLM has been installed into the daily workflow of an enormous knowledge-worker population without their action or consent. Individual users cannot opt back to the previous model; the upgrade simply happens under the branding of “Copilot.”

For anyone tracking human-AI collaboration, this is the shape of routine model turnover in a Copilot-first world: the interface stays the same while the intelligence behind it silently improves — or changes character — every few months.

Source: OpenAI

Neuralink Threads Electrodes Through an Intact Dura

Neuralink disclosed the first BCI implantation performed without cutting the dura mater — the tough protective membrane covering the brain. The procedure, done in May as part of the CAN-PRIME study at Toronto Western Hospital, uses slightly thicker insertion needles engineered to pass through the dura without bending or breaking, paired with new imaging that lets the surgical robot see blood vessels and the cortical surface through the intact membrane.

The clinical rationale is significant. Cutting the dura carries infection risk, complicates future revisions, and is one of the most delicate parts of any neurosurgery. Preserving it should make BCI implantation safer, more repeatable across surgeons, and easier to scale beyond specialist centers. Within an hour of surgery, the participant was using the device to control a cursor on a computer screen.

The company frames the technique as a step toward “high-volume production” — a phrase Musk has used to describe the ambition to move from bespoke research procedures to a standardized, largely automated surgery that a mid-tier hospital could offer. That vision is still years away from a commercial device, but this week’s disclosure is a real engineering milestone toward it.

For readers tracking the invasive-versus-non-invasive debate in neurotechnology, the transdural approach matters because it lowers the surgical cost of going invasive — which changes the tradeoff calculus against wearable alternatives.

Source: ALS News Today

CNBC: China’s BrainCo Bets on Wearable Brain Tech

While Neuralink and Synchron battle for the invasive-BCI market, CNBC’s July 11 profile of Chinese neurotechnology firm BrainCo makes the case for the opposite theory: that the mass market will belong to non-invasive wearables. BrainCo builds EEG headbands for education, focus training, and prosthetic control, and its products are already being sold to consumers rather than restricted to research participants.

The strategic argument is that surgical implants are a small addressable market — dominated by patients with severe paralysis, ALS, or blindness — while wearables serve the far larger population interested in attention monitoring, meditation feedback, sleep tracking, and eventually general-purpose neural input. BrainCo’s founder frames the choice as a bet on distribution: a device that requires neurosurgery cannot become mass-market on any reasonable timescale, but a headband can.

The report captures a real geographic pattern. American BCI development is concentrated in invasive systems (Neuralink, Synchron, Paradromics, Precision Neuroscience), while Chinese efforts skew heavily toward EEG wearables and the software layer above them. Both approaches are legitimate paths to intelligence augmentation, but they optimize for different users and different regulatory environments — and they will produce very different social effects if they scale.

Source: CNBC

HHAI 2026 Convenes in Brussels

The fifth Hybrid Human-Artificial Intelligence conference ran July 6–10 across the Vrije Universiteit Brussel and Université Libre de Bruxelles, gathering the research community explicitly focused on AI systems that work collaboratively with humans rather than replacing them. Of 104 main-track submissions, 39 papers were selected for the proceedings, published by IOS Press in the Frontiers of AI series.

HHAI is distinct from mainstream AI venues in its framing: the object of study is not the model in isolation but the joint human-AI system, and the success criterion is not raw benchmark performance but the quality of the human decision, learning outcome, or task result. This year’s program covered everything from AI-mediated scientific collaboration to interfaces that help clinicians interpret model outputs to design principles for AI tutors that avoid what one paper called “metacognitive laziness” in students.

The conference’s institutional home matters. Belgium’s FARI institute co-hosted, and the European research funding environment continues to weight hybrid-intelligence work more heavily than American funders do. As enterprise deployments increasingly hinge on where the human should sit in an agentic workflow, HHAI’s growing catalog of empirical results on that exact question is becoming a practical resource, not just an academic one.

Source: HHAI 2026

Paradromics Implants Its First Connexus BCI in a Human Patient

Paradromics announced the completion of the first surgical implantation of its Connexus BCI in the FDA-approved Connect-One Early Feasibility Study, performed at University of Michigan Health. The study will assess long-term safety and evaluate whether the device can restore communication — via synthesized text and speech — to people with severe motor impairment.

The Connexus system takes a different design path from Neuralink’s Utah-array-descended threads. It uses a high-density microelectrode array that records from the cortex and transmits signals to a discrete transceiver implanted in the chest, which then relays data wirelessly through the skin to an external receiver. That two-part architecture is intended to reduce failure modes at the skull site and simplify future device swaps.

Following FDA clearance of the Investigational Device Exemption in November 2025, Michigan, UC Davis, and Massachusetts General Hospital began screening eligible participants — the first Paradromics trial to run at multiple academic medical centers simultaneously. The first participant’s identity has not been disclosed.

The significance for the field is that a fourth serious clinical program (alongside BrainGate, Synchron, and Neuralink) has now put its device inside a human. This is exactly the kind of multi-program pipeline that MIT Technology Review flagged in its recent survey as the mark of a field finally scaling.

Source: Business Wire

BrainGate Reports Long-Term, At-Home Speech Restoration for a Person With ALS

The BrainGate consortium published results on Casey Harrell, a 45-year-old man with ALS who has now used an implanted BCI at home for thousands of hours across multiple years to communicate — with the system running independently, without researcher intervention.

The clinical picture matters. Harrell had lost the ability to speak clearly due to bulbar ALS. The BrainGate implant records from motor cortex regions associated with speech articulation and decodes attempted speech into text, which is then read aloud by the computer. The decoding accuracy is described by the authors as “the most accurate speech neuroprosthesis ever reported.” More important than the accuracy number is the setting: this is no longer a supervised lab demo requiring researchers to recalibrate the decoder every session. It is a tool the patient uses on his own, day after day, to talk with his family.

The paper marks a genuine transition point for BCIs — from proof-of-concept demonstrations that make headlines to sustained assistive technology that measurably improves a specific person’s daily life. That distinction is what the field has been chasing for two decades. It also sets a de facto benchmark that other clinical programs (Neuralink, Synchron, Paradromics) will now be measured against.

For patients with progressive motor neuron disease, this class of technology is the first credible answer to the eventual loss of all voluntary communication.

Source: Space Daily / BrainGate

The Brain Processes Language in Stages That Mirror LLM Layers

A study in Nature Machine Intelligence reports that the human brain processes spoken language in a sequence — from raw acoustics, to speech sounds, to individual words, to meaning — that closely mirrors the layer-by-layer computations of modern large language models. The finding emerged from a study of nine people with epilepsy who already had intracranial electrodes for clinical monitoring and who listened to a thirty-minute podcast while researchers recorded their neural activity.

The researchers compared the temporal profile of activity in different cortical regions against the activation profile of successive transformer layers processing the same audio. The correspondence was strong and directional: earlier brain regions matched earlier LLM layers on acoustic features, and later brain regions matched later LLM layers on semantic content.

The interpretation is subtle. The result does not claim that brains and LLMs are the “same” — they clearly use different implementations, learning regimes, and architectures. What it does suggest is that once a system is trained on a large corpus of natural language, a particular ordering of representations may be a near-optimal solution that both biological and artificial systems converge on. That would explain why LLMs work so well as models for probing brain activity, and it hints at practical decoders that borrow structure from language models to interpret neural signals.

Source: Nature Machine Intelligence coverage

MIT Technology Review: Brain-Computer Interface Trials Are Taking Off

MIT Technology Review published a survey of the clinical BCI landscape making a specific quantitative claim: participation in human trials is scaling from single-digit numbers per program to dozens, and the number of serious programs running at once has grown from two or three to at least six. The piece frames 2026 as the year in which the field crossed from proof-of-concept into a real regulatory pipeline.

The magazine’s synthesis maps closely onto the individual data points from this week. Neuralink, Synchron, Paradromics, and BrainGate are each running multi-site trials; the FDA has issued Breakthrough Device designations to several other programs; the surgical procedures are becoming standardized; and outcome measurements are converging on comparable benchmarks (words per minute, cursor control accuracy, error rates over sustained use).

The article also notes the growing tension between the medical-device path — slow, evidence-driven, focused on paralysis and ALS — and the transhumanist framing that a subset of investors and founders continue to push. That tension is likely to shape regulatory posture in 2027 and beyond, since a device approved for a specific medical indication will inevitably be asked to expand to healthier users. The FDA does not currently have a well-developed framework for elective cognitive augmentation.

For readers who want a single scene-setting piece on where the invasive-BCI industry actually stands, this is the one to read.

Source: MIT Technology Review

Foundation Models for Decoding Inner Speech From Non-Invasive Signals

A mini-review in Frontiers in Human Neuroscience takes stock of a fast-moving research area: using foundation models — the same kind of self-supervised pretraining used by LLMs — to decode inner speech from non-invasive brain signals such as EEG and MEG. Inner speech (imagined speech without articulation) is the holy grail for non-invasive BCI because it would let people communicate silently without moving.

The review’s central observation is that classical BCI decoders were trained end-to-end for narrow tasks and did not generalize across subjects or sessions. Foundation models pretrained on large heterogeneous neural datasets learn representations that transfer — the same architecture can be fine-tuned for different users, different tasks, and different recording setups with much less per-user calibration.

The state of the art is not yet clinically useful for inner-speech decoding, and the authors are careful to distinguish between promising benchmark results and real-world reliability. But the review documents a clear trajectory: performance is improving, calibration burden is shrinking, and open datasets are growing. If the trend continues, non-invasive systems could eventually offer a meaningful fraction of what invasive systems provide — at essentially zero surgical risk. That would be a very different market.

Source: Frontiers in Human Neuroscience

Microsoft’s 2026 Work Trend Index: Agents, Human Agency, and Oversight

Microsoft published the 2026 edition of its annual Work Trend Index, based on surveys of tens of thousands of knowledge workers and business leaders. The headline finding is stark: 97% of executives report that their organization deployed AI agents in the past year, and 52% of employees already use them.

The more interesting substance is the finding on oversight. 78% of companies plan to increase agent autonomy over the next year, and 34% already operate in what the report calls “let it rip” mode — agents act first, humans review afterward. That is a rapid retreat from the human-in-the-loop-on-every-step posture that dominated 2024 and early 2025.

The report also documents a widening perception gap: 65% of executives say their AI usage policies are “very clear,” while only 43% of knowledge workers agree. That gap is where governance failures happen. The organizations reporting the largest gains, per the survey, are the ones building explicit outcome accountability and workflow-level governance — not the ones deploying the most agents fastest.

For readers of this newsletter, the report is best read as an empirical baseline for how enterprise human-AI collaboration is actually being structured in the wild. The academic HHAI community’s designs for hybrid intelligence and the enterprise reality of “let it rip” agents are, at the moment, describing very different worlds.

Source: Microsoft WorkLab