Intelligence Augmentation Weekly Review 2026-09-07
Week In Review
The research that landed this week shares an unusual amount of common ground. Almost none of it asks whether an AI system can perform a task — that question is largely settled for the domains under study. It asks instead what condition the human is left in afterward: whether they can still catch the machine’s errors, whether they are becoming more capable or merely more dependent, and whether the interface between the two is costing more attention than it saves. That is a shift in the center of gravity of the field, and it showed up independently in a dozen unrelated labs within four days.
The sharpest version of the question came from a pair of studies on oversight and dependence. A large field experiment with 640 customer-facing employees found that whether people catch a language model’s mistakes turns less on their skill or diligence than on whether the relevant knowledge is retrievable at the moment of review — and that a lightweight onboarding exercise, asking users to explain the task in their own words, measurably improved detection weeks later (Knowing Is Not Enough: Information Retrievability as a Precondition to Effective LLM Oversight). A very different piece of work, a multi-agent simulation of a virtual classroom, found that an AI counselor’s style — affirming versus solution-oriented — drove simulated students toward dependence or self-reliance while producing similar surface satisfaction (How a Chatbot’s Response Style Shapes a Classroom). Both point at the same uncomfortable fact: the properties that make an assistant pleasant to use are not the properties that keep its user sharp.
A second cluster concerned agents that operate software directly, and it produced the week’s most grounded corrective. A three-week diary study put a computer-use agent in the hands of eight blind screen-reader users doing real desktop work; the best model succeeded on just over half of 1,258 commands (Are We There Yet? Assessing Computer-Use Agents for Blind Users). A companion line of work argues that the gap is not raw model capability but missing procedural knowledge, and shows that recording expert demonstrations into explicit standard operating procedures makes agents materially more reliable inside professional software (OmegaUse-SOP). Two studies of the conversational interface itself round out the picture — one attacking the retype-and-rescan loop that makes extended AI sessions exhausting (RecalibrateGPT), another mapping what happens when analysts and models jointly search unstructured data for structure that neither can define alone (Exploratory Unstructured Data Analysis).
The neurotechnology results were quieter but arguably more consequential, because they were about timing and bandwidth rather than accuracy. A virtual-reality study found that EEG can signal whether an operator’s decision is correct before the decision is committed — and that weighting a team’s votes by that signal helps substantially under high workload and actively hurts under low workload (Decoding Decision Correctness from EEG Under High Cognitive Workload). A separate group improved non-invasive visual decoding by abandoning the assumption that EEG maps onto a fixed high-level semantic layer (EEG-based Visual Retrieval and Reconstruction), while a prosthetic-vision study found that when visual bandwidth is scarce, sound and touch guide search better than adding more to the eye (Cross-Modal Guidance for Out-of-View Object Search). The learning-science entry fits the same frame: an algebra tutor that added an emotional layer did not raise test scores, but got students to the same place with less time and fewer hints (Beyond Problem Solving). Across all ten, the interesting variable is never the model. It is the channel, the timing, and the state the human is left in.
Items
Computer-Use Agents Meet Blind Users, and Succeed Half the Time
Computer-use agents — systems that look at a screen, reason about it, and click and type on the user’s behalf — have an obvious constituency among blind screen-reader users, for whom navigating a visually organized interface is the daily bottleneck. This study is the first sustained field test of that premise, and its value lies in refusing to run the evaluation in a benchmark harness.
Eight blind participants used OLLA, a screen-reader-accessible agent prototype, for three weeks across twelve desktop applications, issuing 1,258 commands in the course of their actual work. The researchers captured screenshots, UI trees, model responses, and full action traces for each one, then re-executed the same command set against four additional models to separate what the interface contributed from what the underlying model did. GPT-5 posted the highest success rate at 52.5 percent.
That number deserves to be read in both directions. A little over half of real commands completed successfully is a striking capability for a technology this young, and it is a genuine expansion of what these users could do unassisted in visually complex software. It is also a failure rate that no one would accept from a screen reader, and the failures are not random. The trace analysis sorts them into four recurring categories: grounding errors, where the agent misidentifies what an on-screen element actually is; planning errors; constraint-tracking failures, where it loses track of a condition the user specified; and termination failures, where it does not know when it is finished.
The interviews surface something the success rate cannot. Participants described needs the researchers characterize as “beyond automation” — not simply wanting the task done, but wanting to know what the agent saw, what it decided, and where in a document or application they now are. An agent that completes a task while leaving its user without a mental model of the resulting state has solved a smaller problem than it appears to have solved. That framing generalizes well past accessibility.
Source: arXiv
Teaching Agents the Procedures That Professionals Never Wrote Down
General-purpose computer-use agents perform respectably on public benchmarks and then struggle in real professional software. This paper offers a specific diagnosis: professional work runs on standard operating procedures that encode implicit domain knowledge, software-specific conventions, and verification steps that nobody has ever written out, because the expert doing the work never needed them written out.
OmegaUse-SOP is a human-in-the-loop system for converting expert demonstrations into reusable agent skills, in a process the authors call SOP Engineering by deliberate analogy to prompt engineering — an iterative refinement of demonstrations, execution rules, and domain knowledge rather than a one-shot capture. It runs in four modules. Observe records the expert’s operations as multimodal traces of the graphical interface. Reason abstracts raw click-and-keystroke events into semantic, step-level instructions. Configure folds in domain rules and task-specific parameters. Execute runs the resulting skill in a live interface with step-wise grounding, action generation, and verification.
The test case is refreshingly unglamorous. Working with a client in the power sector, the researchers applied the system to photovoltaic simulation workflows in PVsyst 7.2 — a piece of engineering software with its own conventions that no foundation model has meaningfully internalized. Reliability on professional SOP tasks improved.
What makes this worth attention is the division of labor it proposes. The expert is not asked to write documentation or specify a policy; they are asked to do their job once while the system watches, and then to correct the abstraction it produces. The knowledge that transfers is procedural and tacit, which is exactly the kind that resists being typed into a prompt. If agents are going to enter specialized professional software rather than the generic web, something with this shape has to exist.
Source: arXiv
Why People Miss AI Errors, and a Cheap Fix That Sticks
The standard account of why human review of AI output fails has two branches: reviewers lack the capability to evaluate the output, or they lack the motivation to try. This paper proposes a third mechanism that has been hiding in plain sight — reviewers may know perfectly well how to check the work, but not have that knowledge available to them at the moment they are looking at the screen.
The authors call this information retrievability, and treat it as a distinct precondition for effective oversight rather than a subspecies of skill. They tested it across two randomized lab-in-the-field experiments with 640 customer-facing employees, a scale and setting that put this well above the usual undergraduate-sample study of AI reliance.
Two interventions worked. Asking employees to generate their own explanations — putting the verification-relevant reasoning into their own words at onboarding — improved error detection and strengthened their later recall of that reasoning. And short cues that reactivated the reasoning helped sustain detection through repeated model use, which is where oversight normally decays. The authors name the mechanisms generative encoding and cue-supported reactivation, borrowing from a well-established memory literature that AI-oversight research has largely ignored.
The practical implication is unusually cheap. Organizations worried about rubber-stamped AI output tend to reach for heavyweight remedies: approval workflows, second reviewers, confidence thresholds. This work suggests that a self-explanation exercise during onboarding plus a lightweight daily cue can make human review meaningfully more resilient — not because the reviewer becomes more diligent, but because the knowledge they already have is where they can reach it.
Source: arXiv
An EEG Signal That Arrives Before the Decision Does
Collaborative brain-computer interfaces aggregate neural signals across a team to improve group decisions. The approach has a structural limitation that this paper attacks directly: existing systems rely on evidence that only exists after a decision has been made and reported — reaction time, or stated confidence. Those signals can explain or discount a decision in retrospect. They cannot inform one that is still in progress.
The researchers asked whether spatial-covariance features in EEG could provide a genuinely pre-emptive indicator of whether an operator is about to be right, available inside the response window itself. Twenty-three participants performed a continuous target-detection task in virtual reality under a within-subject manipulation of cognitive workload, high versus low.
The team-level results are strong and sharply conditional. On contested trials — those where the group was evenly split — weighting votes by the pre-emptive neural signal under high workload raised accuracy from 57 percent to 88 percent as simulated team size grew from two to sixteen. Under low workload, the same weighting was actively detrimental. The neural signal also held its own against post-hoc behavioral measures: stated confidence was the strongest single team-level signal overall, but confidence, by construction, cannot inform a decision that has not yet been made.
The authors draw the right conclusion, which is a narrowing one. This is not a general-purpose team augmentation technology; it is a workload-conditional one, useful precisely when operators are loaded enough that their conscious confidence has decoupled from their accuracy. That specificity is what makes it deployable. A system that knows when it should not be trusted is more valuable than one that claims to work everywhere.
Source: arXiv
Reading Images From EEG by Asking Which Layer the Brain Is Actually Visible In
Non-invasive visual decoding — reconstructing what someone is looking at from scalp EEG — has settled into a standard recipe: align EEG features with the semantic features of a pretrained vision model, usually taken from that model’s final layer. This paper argues the recipe contains an unexamined assumption, and that dropping it improves results substantially.
The assumption is that EEG carries information at one level. It does not. Different visual components are neurally visible to different degrees in the EEG signal, and forcing all of them into alignment with a single high-level semantic representation creates cross-modal mismatches and throws away usable information. The authors’ alternative is to search, per subject, for the intermediate layer of the CLIP vision model that maximizes retrieval performance — what they term the Neural Visibility Optimal Layer, or NVOL.
Building on that, they couple retrieval and image generation through a shared intermediate representation. The retrieval branch fuses multi-layer NVOL features, aligns them to image embeddings by contrastive learning, and applies a test-time correction to counter the tendency of a few embeddings to act as universal near-neighbors. The generation branch reconstructs subject-specific NVOL features from EEG using a conditional diffusion prior, maps them into CLIP space through a lightweight adapter, and drives a pretrained Stable Diffusion XL model.
On the THINGS-EEG dataset, NVOL-based retrieval reached 78.1 percent mean top-1 accuracy in a 200-way task, rising to 86.4 percent with the test-time correction, and the two-stage reconstruction beat single-stage final-layer diffusion on both semantic and structural measures. The broader lesson is methodological and portable: when bridging a neural signal to a machine representation, the question of where in the model to connect is a scientific question about the signal, not an implementation detail.
Source: arXiv
Sound and Touch Beat Vision When Vision Is the Scarce Resource
Guiding someone toward an object outside their field of view is a solved problem in virtual and augmented reality: draw an arrow. This study asks what happens when the display itself is the bottleneck — specifically under simulated prosthetic vision, where any visual guidance cue has to be rendered in the same sparse array of phosphenes the user is trying to inspect the scene with.
Nineteen participants searched for objects under two simulated prosthetic vision conditions differing in electrode density and phosphene spread, at 10x10 and 20x20 resolution, crossed with four guidance conditions: none, visual, haptic, and audio. All three cue types were driven by the same underlying variable — horizontal offset to the target — which makes the comparison a clean test of modality rather than of information content.
All three modalities reduced search time and head movement relative to no guidance. But the auditory and haptic cues produced roughly 25 percent faster overall search and 11 to 13 percent faster target acquisition than the visual cue, despite all three producing similarly direct orienting trajectories toward the target. Sound and touch also shortened the search that continued after the target was first acquired.
The most striking result came at the lower resolution. In the 10x10 condition, all three cues reduced vertical localization error by roughly 45 to 58 percent — even though the cues encoded only horizontal offset and carried no elevation information whatsoever. Reducing the user’s uncertainty along one axis apparently frees enough attention to resolve the other. For anyone designing sensory-substitution or low-bandwidth assistive displays, the operational takeaway is direct: when the visual channel is precious, route the guidance somewhere else and spend the pixels on the world.
Source: arXiv
The Interaction Cost of Talking to a Model All Day
Anyone who works with a language model for hours knows the pattern: type, read, retype, scroll back to find what it said before, rephrase the thing you already said, lose the thread. This paper’s contribution is to take that experience seriously as a measurable interaction cost rather than a complaint about model quality, and then to design against it.
A formative qualitative study with twelve advanced language-model users produced a taxonomy of four distinct fatigue types: retyping, scanning, decision paralysis, and context drift. Each has a different cause and, the authors argue, needs a different remedy — which is why a single “regenerate” button does not fix any of them.
RecalibrateGPT introduces five cross-turn operators, each targeting one fatigue type: Anchor, Replay, Delta, Scope, and Steer. Crucially, they act on the full conversation history rather than the last message, and they are invoked with a single click through a structured panel rather than by composing a new prompt. The team tested three layouts for the operator palette — Vertical, Arc, and Tablet — treating the physical arrangement of the controls as a design variable in its own right.
In a follow-up quantitative evaluation with the same twelve users, perceived cognitive workload fell by roughly half on the NASA-TLX scale, with high perceived usability at a System Usability Scale score of 86.5. The sample is small and the studies are explicitly pilots, so the effect sizes should be treated as directional. The framing, though, deserves wider adoption: a substantial share of what users experience as an AI’s limitations is an interaction-flow cost that the interface imposes and the interface could remove.
Source: arXiv
An AI Tutor That Attends to Anxiety Gets Students There Faster
Intelligent tutoring systems have concentrated their adaptive machinery on the cognitive side of learning: diagnose the misconception, deliver the right hint. That works. This study asks what a language model adds when it is pointed at the part of learning that tutoring systems have historically ignored — how the student feels while doing it, which in mathematics is a well-documented predictor of outcomes.
“Math with Matt” is an algebra tutor built around two additions. The first is a language-model-based mindful chat that offers context-sensitive emotional support through a pedagogical agent named Matt. The second is a rewriting of the system’s feedback and hint messages so that they are mindful rather than purely evaluative — a change to the register of routine interaction, not just an extra feature bolted on the side.
The researchers ran a classroom study with seventh graders, comparing the mindful version against an otherwise identical version offering cognitive support only. The headline result is a null one: both versions reduced students’ in-the-moment math anxiety and improved learning, with no significant difference between conditions on learning outcomes.
The secondary results are where it gets interesting. Students in the mindful condition reached comparable learning gains with less time spent and fewer hints requested — higher learning efficiency and what the authors describe as more balanced problem-solving behavior. They also rated the pedagogical agent as more supportive and caring. A tutor that produces the same learning in less time while students lean on it less is not a null result in any practical sense; it is a different and arguably better shape of assistance, and it suggests the socio-emotional layer works by changing how students engage rather than by teaching more.
Source: arXiv
When Analyst and Model Search Together for Structure Neither One Has
Exploratory data analysis has a well-developed toolkit for tables. For large unstructured collections — a million images, say — the analyst’s task is stranger: not to answer a question about known categories, but to work out what the categories are. This paper proposes a conceptual framework for that activity, combining the classical elements of querying and visualization with what the authors call active knowledge construction in the search for structure.
Their formative study watched users conceptualize a structure for an image dataset while exploring it, and produced two findings that constrain how a tool should be built. Users construct faceted classifications from the bottom up, accumulating dimensions as they go rather than starting from a taxonomy. And they rarely produce meaningful spatial categorization during the process — which matters, because spatial layout is what most dimensionality-reduction interfaces offer them.
The team also evaluated CLIP for zero-shot assignment and semantic categorization on the same data. The verdict is split and worth stating plainly: the model remains unreliable at assigning user-defined concepts to images, but it does usefully support semantic grouping. An honest negative result about where the model cannot yet be trusted is more useful for design than another demonstration that it sometimes can.
From this the authors derive four opportunities for human-AI collaboration: intelligent sampling and visualization to maximize how much of the data the analyst actually sees; incremental and few-shot learning to reduce the labeling effort needed for reliable assignment; automatic suggestion of categories, concepts, and facets during the search for structure; and — given the reliability gap they measured — effective methods for calibrating the analyst’s trust. The last is the one the field keeps deferring, and this study makes concrete why it cannot be.
Source: arXiv
Simulating What Happens to a Classroom That Keeps Consulting an AI
Chatbots optimized for user satisfaction tend toward excessive affirmation. The individual-level psychological effects of that have begun to be studied; the population-level effects — what happens to the relationships and mental states of many people in a shared social setting who all keep consulting the same agreeable AI — are close to unobservable in a real classroom over a real semester. This paper’s response is to build the classroom in simulation and be scrupulous about what that does and does not license.
Twenty student agents interact in a virtual classroom across four daily phases, each carrying five state variables: stress, happiness, self-reliance, AI dependence, and sociability. When stressed, an agent consults either a friend or a counselor AI. The counselor is instantiated in six response styles via system prompts — affirming, listening, solution-oriented, reality-redirecting, inciting, and blaming — and a second model call acts as an evaluator that converts each consultation into parameter updates without being shown which style prompt produced it. Seven conditions including a no-AI control were run over 15 days across three classrooms, then over 50 days, then again with a lowered threshold for seeking consultation.
The pattern that emerged is the one worth reporting. The solution-oriented style kept AI dependence low while raising self-reliance and maintaining happiness. The affirming style markedly increased dependence — as did the inciting style, which additionally raised stress and school non-attendance. The listening style, notably, did not relieve accumulated stress. Warmth and helpfulness came apart.
The authors are unusually disciplined about the epistemic status of all this, stating outright that the results describe the simulated system and not measured effects on humans, publishing a complete specification of the agent dynamics, identifying the built-in mechanisms that shape the outcomes, and listing the validation steps — repeated runs, sensitivity analyses, human data — required before psychological conclusions could be drawn. Read as what it is, it is a hypothesis generator for a question the field urgently needs to answer with real people: whether the styles that make an assistant feel supportive are the ones that leave its users less able to cope without it.
Source: arXiv