Intelligence Augmentation Weekly Review 2026-11-10
Week In Review
The week’s strongest results came from studies that treat AI not as a replacement for a person but as a coach, a configurable delegate, or an instrument under expert control. The largest was a randomized trial on the Schoolhouse.world tutoring platform, Automated Feedback to Peer Tutors Increases Their Talk Moves and Learner Reasoning but Not the Balance of Talk, in which over a thousand volunteer tutors received model-generated feedback after each session and measurably changed how they taught. A survey of 289 American adults, What to Delegate, What to Preserve: Human Preferences for Agentic AI Across Everyday Life and Work, found that people want “bounded autonomy” from AI agents rather than an on-off switch, and that worry about losing personal fulfilment pushes against delegation even where the agent would save money. A Microsoft Research position paper, Goals as a Representational Layer for Human-AI Collaboration, argues that the missing ingredient in most of these systems is an explicit model of what the person is actually trying to achieve.
Two papers took up the harder question of what AI assistance does to human skill and how to measure whether a human-AI pair is actually better. Today’s Work, Tomorrow’s Workforce: Incentivizing Learning in AI-Assisted Work models workers choosing between solving a task and accepting an AI answer, and shows how novices can be trapped at low skill while their output quality looks fine, along with the incentive designs that break the trap. Complementarity without ground truth? tackles a methodological gap: most evidence of human-AI complementarity comes from prediction tasks with right answers, while much decision support, especially in medicine, has no single correct label. The authors propose constructing “adequacy labels” so that complementarity can be measured where it matters most.
Expert control was the theme of the two most concrete tools of the week. Lea: An Agent Backbone for Mathematician-Controlled Formalization in Lean from New York University lets a mathematician outline a proof in plain language and keep authority over its structure while agents handle the formal verification, and it produced a complete Lean proof from a natural-language sketch with one substantive correction. Fluidic Body: Swarm Robots as a Fluid Extension of the Hand turns a swarm of robots into something a person steers by opening and closing a hand, and its VR study found users completed combined tasks with far lower workload and a stronger sense of agency than with a standard controller.
Beyond software, Brain Stimulation Device Aims to Treat Women’s Health Conditions reports a consumer headband delivering transcranial direct current stimulation that just won European clearance for anxiety symptoms of premenstrual dysphoric disorder, a reminder that non-invasive neuromodulation is reaching the market with real, if modest, trial evidence behind it. Two Nautilus pieces supplied the long view. In Do We Think in Pictures, Words, or… Code? MIT cognitive scientist Evelina Fedorenko lays out the evidence that language is a communication tool rather than the medium of thought, and draws a parallel to large language models that separate language from reasoning. And We Attribute Personalities to Chatbots, Even When We Know Better reports a Mainz study showing that chatbots present themselves as more agreeable and conscientious than people, which the authors argue can breed excessive trust in systems that still err.
Items
Automated Feedback to Peer Tutors Increases Their Talk Moves and Learner Reasoning but Not the Balance of Talk

Volunteer tutors are abundant but rarely trained, and the tutoring field has long known that good tutors elicit a learner’s ideas and build on them rather than lecturing. A team led by Dorottya Demszky ran a randomized controlled trial across two four-week SAT mathematics bootcamps on that platform to test whether automated feedback could move tutors toward those practices. The study enrolled 1,030 tutors and 7,443 learners, and analysed 6,508 recorded sessions.
Tutors were assigned to one of four conditions: no feedback, automated discourse feedback, or discourse feedback combined with learner-facing participation feedback framed either around personal score gains or around benefit to peers. The feedback itself was generated by GPT-4 Turbo from session transcripts, paired with classifier models that detected specific “talk moves” such as eliciting learner ideas and taking up learner contributions. Within hours of a session, tutors saw a summary comparing their talk time and learner participation to earlier sessions and bootcamp averages, up to three highlighted transcript moments, and concrete suggestions for the next session.
The feedback worked on the targeted behaviours. Across both bootcamps, tutors who received it elicited learner ideas 8 percent more often, took up learner contributions 9 percent more, and their learners produced 15 percent more reasoning talk. What did not budge was the overall balance of conversation: tutors kept talking about 91.5 percent of the time in every condition. The learner-facing feedback produced only marginal gains in satisfaction and no discourse changes.
Interviews with 17 tutors and 13 learners explained the gap. Tutors did adopt the recommended strategies but often faced learners who preferred to listen, and learners said they would rather see group-level than individual metrics. The authors conclude that automated feedback can change technique, but shifting how much students talk “likely requires changes beyond feedback.” For AI-assisted teaching, the result is encouraging in a specific way: the model is not teaching the student, it is coaching the human teacher, and the human’s practice improves in the directions the evidence says matter.
Source: arXiv
What to Delegate, What to Preserve: Human Preferences for Agentic AI Across Everyday Life and Work

As AI agents move from answering questions to carrying out tasks, the design question becomes how much people actually want handed off. Sarah Wheeler, Kyrie Zhou, and colleagues surveyed 289 US adults with prior AI experience, presenting each with one scenario in each of nine life domains, from household management and travel to medicine, mental health, learning, creative work, communication, and personal finance. For each scenario, participants rated how much automation they wanted, their trust in AI agents versus humans versus human-AI teams, their willingness to pay, and their privacy and fulfilment concerns.
Preferences varied widely by domain. Communication tasks drew the most appetite for full automation, with 25.3 percent of respondents wanting the agent to handle everything, followed by travel at 21.1 percent. Mental health and emotional support sat at the other extreme: only 10 percent wanted complete automation and 31.8 percent rejected automation altogether. Across domains, trust in the AI was the strongest single predictor of wanting more automation, and prior use of AI for that specific task and anticipated financial savings also raised acceptance.
The more striking finding is what lowered it. Participants who worried that automation would reduce their sense of personal fulfilment wanted less of it, independent of any practical benefit. Human-AI teams earned higher trust ratings than standalone agents in every domain. The modal preference the authors describe is “bounded autonomy”: the AI executes, but a person keeps oversight or the final say.
The paper’s conclusion is that delegation is not a dial from zero to full automation but “a configuration of execution, collaboration, leadership, and authority” that should be set differently in different parts of life. For builders of agentic systems, that argues for products that expose those settings rather than guessing at a single level of autonomy for everyone.
Source: arXiv
Today’s Work, Tomorrow’s Workforce: Incentivizing Learning in AI-Assisted Work
If an AI assistant is more accurate than a junior employee, the rational move on any single task is to accept its answer. Repeated over a career, that logic can hollow out the expertise an organisation will need later. Moritz Diener, Philipp Spitzer, Gerhard Satzger, and colleagues formalise this problem by extending the “accept-or-solve” model of AI reliance into a dynamic setting in which a worker’s skill changes with every choice: solving independently raises skill with some probability, while accepting the AI’s output lets it decay.
The model distinguishes myopic workers, who maximise the payoff on the current task, from strategic ones who discount future utility. Several results follow. A strategic worker may rationally solve a task alone even when the AI is more accurate, because the skill gained is worth the present effort cost, a choice that conventional measures would label “underreliance.” Yet even strategic behaviour does not escape two traps: novices can stay stuck at low accuracy while their delivered output looks excellent, and experts drift down toward the AI’s level. In both cases, stable output quality masks degrading human capability.
The authors then ask what incentives fix this. Paying workers for correct independent answers restores appropriate reliance but cannot overcome the accuracy gap that discourages novices from practising. A “learning incentive” that also compensates for the accuracy forgone by not using the AI, scaled to the worker’s current skill, does motivate independent work. Supporting workers up to a target skill level can be temporary if independence becomes self-sustaining above that level, and higher targets can paradoxically lower long-run cost.
The policy implications are concrete: budget explicitly for skill maintenance rather than assuming work naturally develops expertise, track unaided accuracy alongside output accuracy, invest differently in novices and experts, and avoid appraisal systems that reward output while demanding independent practice. The paper reframes the question from whether a worker should rely on AI today to which competencies an organisation chooses to sustain.
Source: arXiv
Complementarity without ground truth?
The promise of human-AI collaboration is complementarity: the pair outperforms either alone. Nearly all empirical evidence for it comes from prediction tasks, where a ground-truth label makes performance easy to score. Andrea Ferrario, Alessandro Facchini, and Federico Cabitza point out that this leaves out much of the territory where decision support actually operates. Choosing a treatment in an emergency department is not a prediction with a right answer; it is a judgement whose quality is “adequacy relative to a practical goal.”
Their proposal is to build a reference where none exists. Drawing on multi-criteria decision analysis as used in healthcare, and on the philosophical framework of computational reliabilism, they define a two-step scoring procedure. In the first step, each candidate option is assessed along explicit adequacy dimensions. For respiratory support in an emergency department, their running example, those dimensions include expected respiratory benefit, risk of deterioration or organ failure, treatment burden and adverse-event risk, and feasibility in the clinical context. In the second step, criterion scores are aggregated into a ranking, and the top-ranked option becomes the adequacy label.
With such labels in hand, complementarity can be computed as it is in prediction settings: the team achieves it when its loss against the adequacy reference is lower than the better of the human alone and the AI alone. The authors illustrate the procedure at two levels of complexity, from fracture detection on X-ray to schizophrenia-spectrum diagnosis, and for a three-way choice among conventional oxygen, advanced non-invasive support, and intensive-care ventilation.
The advantage the authors emphasise is not just that complementarity becomes measurable in more settings but that the normative choices underlying the evaluation are made explicit rather than hidden inside a label. That matters for any field that wants to know whether a human-in-the-loop system is making expert judgement better, rather than merely different.
Source: arXiv
Lea: An Agent Backbone for Mathematician-Controlled Formalization in Lean

Formal verification in the Lean proof assistant is the gold standard for mathematical certainty, but writing Lean is slow and alien to most working mathematicians. Autonomous AI provers can now generate formal proofs, yet they operate as black boxes that a mathematician cannot steer. A group at New York University led by Chinmay Hegde, with collaborators at Fundação Getulio Vargas and Oklahoma State University, built Lea to put the mathematician back in charge of the process.
Lea is an infrastructure layer rather than a single application. It exposes “runs,” persistent “sessions,” and typed events through one HTTP API, so that applications can react to specific state changes, such as a proof compiling or a verification passing, rather than parsing raw tool output. A formalization agent operates on a project of Lean namespaces, instructions, a proof “blueprint,” and accumulated memory, delegating bounded sub-tasks to concurrent sub-agents whose prompts, tools, and budgets are declared in configuration files. A pinned Lean execution environment distinguishes “formalized,” meaning the file compiles without placeholders, from “verified,” meaning it passes kernel replay and an axiom whitelist.
Two applications sit on top. LeaChat is a web client showing the agent’s transcript beside an editable Lean canvas; the mathematician can edit proofs mid-run, approve or deny tool calls, rewrite blueprint nodes, or redirect strategy in chat, and every intervention is recorded and shapes what the agent does next. LeaOverleaf is a browser extension that treats a LaTeX manuscript as the formalization surface: the mathematician annotates theorems in comments, Lea extracts them into blueprint nodes, maintains their status inside the document, and automatically invalidates results when the source text changes. Sessions begun in one tool continue in the other.
In the reported case study, a mathematician outlined a proof about the Durfee square generating function across a series of chat messages. Over three sessions and 38 sub-agent invocations using GPT-5.5 against Mathlib, Lea produced a complete Lean development that followed the intended decomposition and needed substantive correction at a single point. The mathematician sent 28 messages, mostly confirming proposed definitions. Lea and both applications are released as open source, offering a foundation for tools where the expert supplies the mathematics and the machine supplies the rigour.
Source: arXiv
Fluidic Body: Swarm Robots as a Fluid Extension of the Hand

Controlling dozens of small robots at once usually means specifying waypoints or switching among preset formation behaviours. Michikuni Eguchi and Takefumi Hiraki of the University of Tsukuba and Cluster Metaverse Lab, with Shigeo Yoshida and Mai Nishimura of OMRON SINIC X, propose something more bodily: treat the swarm as a fluid that your hand shapes. A hand-tracking camera reads the hand’s outline, height, and openness, and each quantity drives one property of a virtual fluid that the robots then embody.
Under the hood is smoothed particle hydrodynamics, with each robot modelled as a fluid particle subject to three forces: a fluid force combining pressure and viscosity, inter-robot repulsion to avoid collisions, and a shape force that draws robots into the region beneath the hand. Opening the hand lowers viscosity so the swarm becomes soft and deformable, slipping through gaps between obstacles; closing it raises viscosity so robots align into a stiff, cohesive mass that can push an object. Hand height spreads or concentrates the formation.
The team tested the controller in simulation, in a virtual-reality study with 17 participants, and in a physical demonstration with 88 members of the public at SIGGRAPH 2025 Emerging Technologies. In the VR study’s combined obstacle-avoidance and transport task, the fluid controller achieved a 100 percent success rate against 47.1 percent for a standard reciprocal-velocity-obstacle baseline, with mean completion time of 38.5 seconds versus 76.9. Participants reported roughly half the workload on the NASA-TLX scale and a markedly stronger sense of agency, though body ownership did not differ. The physical demo drew similar agency ratings, along with complaints about latency that limited fine manipulation.
The authors frame the result as establishing the swarm as “a deformable extension of the human body.” It is a small but vivid example of augmentation through interface design: the robots gain nothing new, but the person gains an intuitive, low-effort way to direct many of them at once.
Source: arXiv
Goals as a Representational Layer for Human-AI Collaboration

AI systems now sit inside the documents, chats, and calendars where knowledge work happens, yet they mostly see requests without seeing reasons. Gaia Molinaro, Lev Tankelevitch, Sean Rintel, and Scott Counts of Microsoft and Microsoft Research argue in this position paper that goals should become “an explicit representational layer mediating interactions among people, AI systems, and organizations.”
The difficulty they confront is that human goals are hierarchical, partly concealed, and continually shifting. Their proposed representation is multidimensional and organised around seven questions: what outcome and success criteria are desired, who is responsible, who else is involved, when it is due, where the work lives in terms of tools and artefacts, why it matters to larger objectives, and how it decomposes into steps. The representation would be built jointly from what people state and what can be inferred from context, and continuously refined, including by detecting when a person has switched goals.
The payoff, in their account, runs in several directions. An assistant that knows a report summary is for preparing a meeting rather than drafting a proposal can produce a different summary. Goal structure helps organise and retrieve the context that matters, reduces the amount of material an AI must be handed, supports reflection and coordination among colleagues, and could serve as the basis for training and evaluating AI over multi-turn interactions around human-centred objectives rather than isolated prompts.
The authors are candid about the obstacles. Goals are prospective while the evidence for them is retrospective, and ground truth is rare, so inference and validation are hard. Representations go stale, and representing a goal is not the same as acting on it well. Most seriously, a system that models what people are trying to do raises privacy and governance questions, and the paper insists that human agency in setting goals must remain central. As a design agenda it is a clear statement of what intent-aware assistance would require.
Source: arXiv
Brain Stimulation Device Aims to Treat Women’s Health Conditions

Transcranial direct current stimulation, which passes a weak current through the scalp to nudge the excitability of underlying cortex, has spent two decades mostly in laboratories. Greg Uyeno reports for IEEE Spectrum on Samphire Neuroscience, a London company selling a tDCS headband marketed as Nettle in Europe, where it is regulated as a medical device, and as Lutea in the United States, where it is sold as a wellness product. The device uses four saline-wetted sponge electrodes: the front pair targets the dorsolateral prefrontal cortex, a region tied to mood regulation, and the rear pair targets the motor cortex to influence pain pathways.
The news hook is regulatory. The headband was already cleared in the UK and EU for menstrual pain and premenstrual syndrome, and as of 5 November it is also cleared for anxiety symptoms of premenstrual dysphoric disorder, a severe condition with, as Queen Mary University of London psychologist Paul Faulkner puts it, “very few treatment options” that all carry significant side effects. Faulkner led an independent study of 42 participants; after a month of use, PMDD patients reported anxiety symptoms less frequently and had less difficulty regulating emotions. Depressive symptoms showed no measurable difference, possibly because of placebo effects.
The article does not oversell the evidence. The adverse effects reported are minor electrode burns. Medical ethicist Anna Wexler notes that tDCS research has not consistently accounted for diversity in participants, including whether different hair textures change the device’s efficacy. The cost is not disclosed.
The pathway is the notable part: a consumer-grade neuromodulation device accumulating independent trial evidence and indication-specific clearances one at a time. Devices that establish safety and efficacy in a clinical niche are the likeliest route by which electrical modulation of mood and cognition reaches ordinary users.
Source: IEEE Spectrum
Do We Think in Pictures, Words, or… Code?

Whether thinking requires language is one of the oldest questions in cognitive science, and Evelina Fedorenko of MIT has spent decades attacking it with behavioural studies, brain imaging, intracranial recordings, and computational models. In this Nautilus interview with Kristen French, she sets out her conclusion: language is a tool for communication, not the medium in which thought happens.
Two bodies of evidence carry the argument. The first comes from aphasia. People with severe damage to the brain’s language areas can lose the ability to produce or understand sentences while retaining the ability to solve arithmetic problems, navigate space, and make social judgements; what they have lost is the translation between thought and language, not thought itself. The second comes from functional MRI in healthy people, where the language network lights up when people use language but stays quiet during reasoning tasks such as mathematics or inferring what someone else believes.
If not words or pictures, then what? Fedorenko favours a “probabilistic language of thought,” a structured, code-like representation that she describes as closer to a program than to a sentence. On that view, the Sapir-Whorf idea that one’s language determines one’s thinking gets the direction wrong: culture profoundly shapes cognition and shows up in language, but language is the output rather than the engine.
The AI parallel is direct. Large language models demonstrate that fluent language can exist without much reasoning, and Fedorenko observes that the most capable recent models appear to develop separable components for language and for cognition, an architecture that mirrors the separation found in the human brain. That convergence suggests the human brain and its artificial collaborators may be more alike in how they divide the work than the “language model” label implies.
Source: Nautilus
We Attribute Personalities to Chatbots, Even When We Know Better

People who work alongside AI assistants every day develop a feel for their “character,” and that intuition shapes how much they trust the output. Kristen French reports on a study from Johannes Gutenberg University Mainz that put the intuition to the test. Researchers administered standard personality questionnaires, including the Big Five and measures of narcissism, to 70 people and seven chatbots spanning ChatGPT variants, Llama-3, Gemma, Gemini, Phi-3, and Orca-2, then had both humans and bots write self-descriptions in German. Ten human judges and six AI judges rated those texts for personality and tried to tell who had written them.
The chatbots described themselves as more agreeable, more conscientious, and less neurotic than the humans did. Human judges could tell AI text from human text 83 percent of the time, while the AI judges managed only 67 percent, but the human advantage shrank substantially once text length was matched. The telltales were stylistic: AI writing leaned on prepositions, articles, and time-related words, while human writing dwelt on leisure and feelings. Both human and AI judges struggled to detect narcissism.
The authors’ concern is that an interlocutor who comes across as agreeable, diligent, and emotionally steady invites trust, and that people extend it even when they know the personality is a surface produced by training rather than a disposition. Since the same systems remain prone to confident errors and fabrications, that trust can be excessive, and the researchers call for greater transparency about how chatbots are shaped.
The constructive reading is that the mismatch is a calibration problem. Knowing that the pleasant persona is uninformative about reliability lets a user treat it as such and evaluate the content on its merits, which is exactly the discipline that effective human-AI teaming requires.
Source: Nautilus