Intelligence Augmentation Weekly Review 2026-11-03
Week In Review
The week’s most consequential result is also its most sobering, and it comes from a machine reading the literature on itself. A Harvard and MIT team released AutoAnalysis, a self-updating meta-analysis that pools 1,955 effect sizes from 420 papers on human-AI collaboration. Its two headline numbers point in opposite directions. People working with AI outperform people working alone by a medium-sized margin, but the pair underperforms whichever of the two would have been better on its own. The gap between “augmentation” and “synergy” is the frame for everything else this week, because the other studies are, in effect, attempts to find the conditions under which that second number turns positive.
Several of them succeed, and the pattern is consistent: the human contributes most when placed early and given structure. A Google Research randomized trial of 802 adults found that a structured visit-preparation agent built on the same model as an unstructured chatbot produced measurably better-prepared patients by both clinician and patient standards. A Cornell-led writing experiment with 333 participants found that all-AI story chains were more creative but less diverse than human ones, and that a human doing only the planning stage kept the AI’s creativity while restoring the diversity. On the design side, a Sydney Business School group formalized when a fleet of agents should interrupt one supervisor with clarifying questions, and showed that treating the human’s attention as a shared resource cuts coordination cost substantially.
The caveats cluster around trust. A 180-participant study found that when a model asks a clarifying question before answering, people feel better understood, review the answer for a third less time, and accept more wrong responses. A five-week classroom study found that students’ judgements of an AI teammate shifted week by week in ways no single-session evaluation would catch. And Dutch clinicians who use prediction tools in spine surgery and gynaecological oncology told researchers they consult the model only when in doubt, discuss its output with colleagues, and find it “a little bit weird” to rely on it alone. That selective, socially mediated reliance looks less like a failure of adoption than like the behaviour the meta-analysis would recommend.
Two items sit a level below the interface. Dartmouth researchers showed that the part of a language model’s representation that predicts human brain activity during listening occupies only one to three percent of the model’s dimensions and transfers across people, a compact bridge between artificial and biological language processing. And Shanghai’s NeuroXess closed a RMB 700 million Series B to take a flexible-electrode implant through multicentre trials and into a factory sized for 10,000 systems a year, a reminder that the hardware side of augmentation is now being built at industrial scale on more than one continent. A Google engineer’s essay in IEEE Spectrum ties the threads together with the claim that an engineer’s value in the AI era lies in knowing how to frame the problem and when another expert must lead, which is a fair one-line summary of what the data say.
Items
A Self-Updating Meta-Analysis Finds Humans Gain From AI but Pairs Lose to the Best Partner Alone

Michelle Vaccaro of Harvard Business School and Abdullah Almaatouq and Thomas Malone of MIT Sloan posted AutoAnalysis on October 30, a pipeline that continuously re-runs a meta-analysis of human-AI collaboration as new studies appear. The system identifies candidate papers by citation snowballing, screens them, extracts effect sizes, and regenerates the pooled estimates on a schedule. Its most recent run integrated 1,955 effect sizes from 420 peer-reviewed papers across human-computer interaction, artificial intelligence, and neighbouring fields, and the results are explorable at auto-analysis.dev.
The central finding separates two questions that are often conflated. Does a person with AI do better than the same person without it? Yes: the pooled effect for human-AI combinations against humans alone is a Hedges’ g of 0.53, a medium-sized improvement. Does the combination beat whichever of the human or the AI would have done better alone? No: against that stronger baseline the pooled effect is negative, g of minus 0.35. In the authors’ terms, human augmentation is reliably achieved, but human-AI synergy, on average, is not.
The moderator analyses say something about where synergy might be found. Decision tasks showed a larger AI advantage over humans than creation tasks. An explicit division of labour between the person and the system improved both augmentation and synergy. Providing AI explanations improved augmentation. Crowdworkers benefited less than other participant groups, which suggests that expertise and stakes change what people do with a tool.
The methodological point is as important as the numbers. Many studies report only the comparison with humans alone, and that comparison will almost always look good. The authors argue that collaboration must be evaluated against both baselines, and they have built the apparatus to keep doing so as the field, which they chart as growing steeply since 2000, produces more evidence. A living estimate that can be checked next month rather than next decade is a useful thing for a fast-moving field to have.
Source: arXiv
A Randomized Trial of 802 Patients Shows Structure, Not the Model, Makes an AI Visit-Prep Assistant Work

A Google Research and Google DeepMind team led by Xincheng Tan reported on October 30 a prospective randomized study of an AI agent that helps patients get ready for a clinical appointment. The agent gathers the patient’s medical context, clarifies what the visit is for, delivers relevant health information, assembles a ranked list of questions to ask, and produces a summary. It was built on Gemini 2.5 Flash through system prompting alone, with no modified weights. The comparison arm was a minimally guided chatbot using the same underlying model, which makes the trial a test of structure rather than of raw model capability.
The study enrolled 802 US adults through the Google Health app, generating 866 health conversations. Three board-certified primary care clinicians with at least seven years of post-residency experience conducted blinded reviews of 272 completed sessions on medical safety, clinical appropriateness, comprehensiveness, and alignment with the patient’s situation. The structured agent achieved comprehensive preparation in 67.2 percent of sessions against 50.6 percent for the control, and alignment with the patient’s health context in 98.9 percent against 90.4 percent. Safety and appropriateness were maintained across all arms.
Patients agreed. Among 460 survey respondents, 73.8 percent of those using the structured agent reported feeling prepared for their visit against 57.0 percent of controls, 81.4 percent against 67.1 percent felt confident about what questions to ask, and 74.0 percent against 62.0 percent were satisfied overall. Clinician ratings and patient ratings did not correlate with each other, which the authors read as evidence that the two groups are measuring genuinely different aspects of whether an AI tool is helping.
The limitations are stated candidly. Only 57.7 percent of conversations completed the full flow, raising the possibility of survivorship bias. Only 95 people answered the 14-day follow-up, too few to assess clinical outcomes. Participants were Android users and research volunteers, likely more technologically engaged and health literate than the general population, and demographic data on income, race, education, and insurance were not collected. Within those bounds, the trial is a clean demonstration that a well-designed conversation scaffold turns the same model into a better partner for a person preparing to make decisions about their own care.
Source: arXiv
In Creative Writing, a Human Who Plans and an AI That Writes Beats Either Alone
Chenyi Li, Nori Jacoby, and colleagues from Cornell, Princeton, MIT, Cornell Tech, and Purdue posted a study on October 27, revised October 30, that asks where in a creative process a human adds the most. The design borrows from cognitive models of writing. Stories are produced through a four-stage chain of plan, write, review, and revise, with each stage receiving the previous stage’s output, and the researchers vary which stages are performed by people and which by GPT-4 Turbo. Across the experiments, 333 participants contributed to 427 stories.
Two measures were tracked. Creativity was rated by a separate pool of non-expert judges using the Consensual Assessment Technique, supplemented with dimensions from the Torrance Tests. Diversity was defined as one minus the average semantic similarity between a story and the other stories produced from the same prompt under the same condition, computed with text embeddings. The two measures turned out to pull apart. All-AI chains produced more creative but less diverse stories than all-human chains. The AI wrote well, but it wrote the same way every time.
Inserting a single human into an otherwise AI chain restored diversity, and the position mattered. Human planning was the most effective placement: it preserved AI-level creativity, increased diversity that persisted through the downstream AI writing and revision stages, and maintained participants’ enjoyment. In the final revision round, the human-plan condition scored a mean creativity of 4.90 and a mean diversity of 0.358, though diversity still declined across rounds.
The practical reading is that the human’s comparative advantage is in direction rather than polish. A person who decides what the story is about, and then hands off, contributes a kind of variety that the model does not generate on its own and that survives several rounds of machine elaboration. For anyone designing creative tools, the result argues for interfaces that put the person at the front of the pipeline rather than at the end checking the output.
Source: arXiv
When an AI Asks a Clarifying Question, People Review Its Answer Less and Accept More Errors

Chengbo Zheng of the University of Queensland, with co-authors at Tongji University, Renmin University, and the Chinese Academy of Sciences, posted a controlled experiment on October 31 on what happens when a language model asks the user a clarifying question before responding. Clarifying questions are a popular design pattern because they elicit preferences and make answers feel personalised. The study tests whether that feeling has a cost.
The researchers ran 180 participants through six everyday decision scenarios. They manipulated whether the model asked a clarifying question and whether its response violated an explicit constraint the user had stated. When a violation was present, it was placed either adjacent to the topic the clarifying question had covered or somewhere else in the response. Participants’ review time, use of sources, acceptance of responses, and post-study ratings of trust and satisfaction were recorded.
Clarifying questions raised perceived understanding and satisfaction, as expected. They also cut the mean time participants spent reviewing a response from 80.7 seconds to 55.0 seconds, and raised the acceptance rate of erroneous responses from 39 percent to 52 percent. Violations located near the topic of the clarifying question were accepted more often than violations elsewhere, which suggests that the question created a halo of assumed correctness around whatever it had touched.
The authors list the limits: behavioural measures miss cognitive detail, the scenarios carried no personal consequences, only constraint violations rather than fabrications were tested, and the question format was fixed. Even so, the result adds a specific mechanism to the growing list of ways that interface features which improve the experience of working with AI can reduce the scrutiny that makes the collaboration safe. The design implication is not to stop asking questions, but to avoid letting a question stand in for verification.
Source: arXiv
A Framework for Deciding When Parallel AI Agents Should Interrupt Their Human Supervisor
Hanzhao Wang of the University of Sydney Business School, Shengjie Liu of Amazon, and Zhengchao Wang of the University of Sydney posted a paper on October 30 on a problem that is becoming routine as people run several AI agents at once. Each agent may hit a gap in its instructions, such as a missing specification or an unstated preference. It can search existing records for the answer, ask the human, or carry on with parts of the task that do not depend on the missing information. The human has limited attention, so every question one agent asks delays the others.
The paper models this as a dynamic programming problem and introduces a Pairwise Coordination Policy. The policy solves exact one-task and two-task versions of the problem to estimate how much any pair of tasks delays each other when they share the same human, and uses those estimates to decide, for each agent, whether to ask now, search, or work on something else. The running example is a procurement manager overseeing several purchase agents, where a missing maximum lead time might be found in earlier transaction records or might need the manager’s input. The paper’s opening figure uses two coding agents that both need a decision from one developer.
On a synthetic benchmark of 864 cases, the policy came within 0.07 percent of the optimal cost on average and reduced costs by 18 to 31 percent relative to three baseline heuristics. In a case study with language-model agents, every task completed correctly and the policy achieved the lowest mean cost in all configurations. The authors prove the policy is exactly optimal when at most two requirements remain unresolved and within a factor of two of optimal when every unresolved item needs the human.
The underlying insight is simple and easily missed: an agent cannot decide on its own whether to ask, because the right answer depends on what the other agents are doing and how busy the person is. Human attention is the scarce resource in a multi-agent workflow, and systems that treat it as such will get more out of both the agents and the human.
Source: arXiv
Five Weeks With an AI Teammate: Students’ Judgements Kept Changing After the First Session
Nia Nixon, Jaeyoon Choi, Pedro Martins de Bastos, and Mohammad Amin Samadi posted a longitudinal study on October 29 of what happens when the same AI agent sits in a team for more than one meeting. Seventeen undergraduate teams, 48 students in all, in a learning analytics course worked with a persistent language-model agent in weekly sessions over five weeks. The agent used a trigger-based architecture to decide when to speak. Two versions were tested: a Cognitive Agent that challenged quick consensus, and a Supportive Agent that invited quieter members to contribute.
The researchers combined chat discourse analysis, post-session surveys, and spoken reflections, and organised the results into three patterns. Accommodation came first and fast. Students matched their language to the agent’s role from the first session, using substantially more social language with the Supportive Agent. Evaluation came next. Students’ ratings of how helpful the agent was were recalibrated session by session against what it actually did, rather than staying anchored to their prior attitudes toward AI.
Emergence was the slow pattern and the interesting one. Measures of collaboration quality and positive-affect language in Supportive-Agent teams started below those of Cognitive-Agent teams but overtook them by the end of the five weeks, even though those same students consistently rated the Supportive Agent as less helpful. What the team felt about working together and what they said about the tool diverged, and only the longer study could see it.
The authors’ conclusion is methodological: single-session evaluations of agentic AI may miss changes that emerge only through repeated interaction. For teams adopting AI members, the finding cuts both ways. First impressions are unreliable guides to how the collaboration will settle, and the agent that feels most useful on day one is not necessarily the one that makes the group work best by week five.
Source: arXiv
Experienced Clinicians Use Prediction Tools Selectively, Socially, and With a Healthy Discomfort

A Radboud University team led by Simon Fischer, with Pim Haselager and colleagues from the Donders Institute and Radboud University Medical Center, posted a mixed-methods study on October 29 of how experienced clinicians actually use AI decision support in practice. The study combined a survey of 12 clinicians with semi-structured interviews of 10, drawn from two cohorts. In the first, spine specialists use a logistic-regression tool that predicts the success of surgery, a combined physical-psychological programme, or physiotherapy for chronic low back pain from patient questionnaires. In the second, gynaecological oncologists use a Bayesian-network tool that estimates lymph node metastasis risk in endometrial cancer from biomarkers and shows the contribution of each feature.
The headline finding is that reliance is conditional. “When I’m in doubt, I consider the prediction. If it’s a really clear case, I don’t even look at the DSS,” one participant said. The model’s output gained weight in complex cases and faded in routine ones. Clinicians also frequently discussed recommendations with colleagues before deciding, so the tool entered a conversation rather than replacing one. How the system was integrated into the workflow substantially shaped how much it was used at all.
The paper’s title comes from a participant describing cases where the clinical examination contradicted the prediction: “It feels a little bit weird to only trust on the AI.” That discomfort coexists with appreciation. “The models are helpful because I think every doctor already calculates some kind of risk, but not in real numbers,” another said. Quantifying an intuition the clinician already holds was the benefit most readily named.
The authors propose a socio-technical framework for understanding reliance that includes human factors, the system’s features, the clinical context, and the organisation, and they argue that simplified models of one person facing one AI miss most of what determines whether the tool helps. The picture that emerges is of expert judgement using a model the way it uses a second opinion, which is close to the arrangement that the week’s meta-analytic evidence suggests actually produces gains.
Source: arXiv
The Part of a Language Model That Predicts the Listening Brain Is Only One to Three Percent of It

Zach Paris and Emily Finn of Dartmouth’s Department of Psychological and Brain Sciences posted a study on October 29 that sharpens a question behind much recent neuroscience: when a language model’s internal representations predict human brain activity, which parts of those representations are doing the work? Using ridge reduced-rank regression, they searched for the brain-predictive subspace within model embeddings across three naturalistic listening datasets: an intracranial recording of nine participants with 1,330 electrodes hearing a 30-minute podcast, stereo-EEG recordings from the Brain Treebank during a film, and an fMRI dataset of eight participants listening to spoken stories over five sessions.
They compared six model conditions, including autoregressive models from GPT-2 to Llama 3 8B, the bidirectional RoBERTa Large, a randomly initialised GPT-2, and GloVe word vectors. The finding is that full linear encoding performance can be recovered using only one to three percent of the original embedding dimensionality. Neural prediction saturated with tens of dimensions. Principal component analysis, by contrast, needed far more dimensions to reach the same predictive performance, which means the directions most useful for predicting brain activity are not the directions that carry the most variance in the model.
The subspaces generalised across held-out participants and partially across datasets. What they encode differed by recording type: the intracranial datasets emphasised features like position within a sentence and prediction entropy, while the fMRI data emphasised word concreteness. Transfer loss between datasets tracked how much the stimulus-dependent structure of the embeddings shifted. Code is available in the thefinnlab/low-rank-alignment repository.
The result matters for any effort to connect language models to the brain in either direction. A compact, shared, interpretable subspace is a far better target for decoding, for cross-subject models that do not need lengthy per-person calibration, and for understanding what the two systems have in common than a 4,096-dimensional embedding whose brain-relevant content is scattered among its noise.
Source: arXiv
NeuroXess Raises RMB 700 Million to Take a Flexible-Electrode Implant to Scale

TechNode reported on October 28 that Shanghai brain-computer interface company NeuroXess has closed a Series B round of RMB 700 million, roughly 100 million dollars, led by SDIC Juli, Shenzhen Capital Group, and Qinzhi Capital. The company develops implantable systems built on flexible electrodes with wireless power and data transmission, designed to decode movement and speech and to control external devices.
The clinical track record is the notable part. NeuroXess has completed 65 clinical cases, and in July 2026 it started a registration trial involving nearly 20 hospitals. The new funds are earmarked for multicentre clinical trials, regulatory submissions, expanded manufacturing capacity, and next-generation product development.
The manufacturing plan is what signals the company’s intentions. NeuroXess is building a 14,300-square-metre facility with a planned annual capacity of 10,000 systems. No implanted BCI anywhere has yet shipped at that volume, and a factory sized for it, alongside a registration trial across many hospitals, indicates that the company is planning for a product rather than a study.
For the field as a whole, parallel programmes on multiple continents, each with its own regulatory route, clinical cohorts, and supply chain, make it more likely that implanted interfaces reach routine use this decade, and raise the stakes for interoperability and standards along the way.
Source: TechNode
An Engineer’s Value in the AI Era Is Framing the Problem and Knowing When to Defer

Sufian Aldogom, an IEEE member, research affiliate at MIT, and senior leader at Google, argued in IEEE Spectrum on October 28 that AI is changing what engineering expertise is for. As machines take over research synthesis, design generation, and simulation testing, the engineer’s work shifts from execution toward coordination: deciding what to ask, how to test the answers, and which parts of a problem belong to a machine, to oneself, or to a specialist in another discipline. “An engineer’s value is not in knowing everything,” he writes. “It is in knowing how to frame the problem, how to ask the right questions, how to test the answers, and when another expert must lead.”
His examples are concrete. A software engineer can now use AI to review biomedical literature and build preliminary models for a medical device without becoming a physician, which accelerates exploration of an unfamiliar domain but does not remove the need for the specialist’s judgement. An AI system can generate several candidate designs for a medical sensor and predict how each will perform, and an engineer can test the predictions and reject designs that violate technical requirements. In each case the machine widens what one person can attempt, and the person’s expertise determines what survives.
Aldogom is careful about where the boundaries still lie. Deep expertise, he argues, is what gives engineers the foundation to judge what is sound, and that matters more rather than less when AI can produce convincing answers from weak assumptions. Machines do not assess context, values, or responsibility. The discipline he recommends is systems thinking: understanding the interconnections, constraints, and trade-offs across technical domains well enough to allocate work sensibly among humans and tools.
He closes with an institutional ask, that professional bodies such as IEEE establish standards for human-machine collaboration and build problem-centred communities that cross engineering fields. The essay is an opinion piece, but its description of the engineer as coordinator of a mixed human-machine system matches what this week’s empirical studies found about where human input pays off.
Source: IEEE Spectrum