World models promise the foresight accessibility lacks

World models and disability access overlap heavily in substance and almost not at all in name. A research review of where prediction could help most, and where a model of "normal" could hurt.

As of October 2026, world models and disability accessibility overlap heavily in substance and almost not at all in name: searches found no peer-reviewed paper that frames an assistive system as a world model, and none of the flagship launches (Genie 3, V-JEPA 2, Cosmos, Marble, Muse) mentions disability. Yet the failures disabled users report most often in today's AI assistants — going silent when a scene starts to move, misjudging distance and direction, forgetting where the user has already walked, alerting too late or too often — are exactly what world models are built to fix by forecasting how a scene will change, simulating the result of an action, remembering space and planning before acting. The working intersection is therefore implicit, scattered across anticipatory navigation prototypes for blind pedestrians (best warning accuracy 67%), spatial sound models for Deaf and hard-of-hearing users, conversation-aware AAC that raised ALS eye-gaze typists' text-entry rates by 29–60%, routine-learning dementia glasses, shared-control wheelchairs and brain-computer-interface copilots, and GUI world models that predict how an interface will change. Shipped products deliver continuous perception rather than prediction, and the most proactive offering, Google DeepMind's Project Astra Visual Interpreter built with Aira, runs only under live human oversight. The central risk is built into the method: a world model is a learned prior over what normally happens, trained on footage shot mostly by sighted, hearing, walking people, so disabled bodies and behaviors are both underrepresented and predicted as improbable — a hazard that matters most in autonomous vehicles and robots, and that persistent memory of a person's home and routines compounds with surveillance. The opportunity is to treat the edges of the distribution as the real test of a world model, to build consented, disability-representative data and benchmarks that reward "no" and "I can't tell", and to deploy prediction as a copilot the user controls rather than an autopilot that replaces human help.

The overlap is large, but nobody has named it

Six families share the label, and assistive technology needs the least glamorous ones

"World model" has no settled definition; the common core is an internal model that predicts how an environment evolves, usually in response to actions (Wikipedia (opens in a new tab))). At least six families now sit under the umbrella. Latent-dynamics agents in the Dreamer line learn a model and improve by imagining futures; DreamerV3 outperformed specialized methods on more than 150 tasks with a single configuration (arXiv 2301.04104 (opens in a new tab)). Joint-embedding predictors such as Meta's V-JEPA 2 forecast embeddings rather than pixels after pretraining on more than 1M hours of internet video (arXiv 2506.09985 (opens in a new tab)). Generative interactive worlds — DeepMind's Genie 3, World Labs' Marble, Odyssey, Decart's Oasis, Microsoft's Muse — render explorable environments from prompts (DeepMind (opens in a new tab); World Labs (opens in a new tab)). Physical-AI platforms such as NVIDIA Cosmos generate synthetic data for robots and vehicles, and Cosmos 3 (June 2026) folds forward dynamics, inverse dynamics and policy generation into one model (NVIDIA (opens in a new tab); Hugging Face/NVIDIA (opens in a new tab)). Video generators framed as world simulators, such as Veo 3, show emergent zero-shot visual reasoning (arXiv 2509.20328 (opens in a new tab)). And internal world models in LLMs remain contested: Othello-GPT develops a board representation that interventions can steer (arXiv 2210.13382 (opens in a new tab)), while Vafa and colleagues find such models far less coherent than they look and brittle on subtly different tasks (arXiv 2406.03689 (opens in a new tab)). The field is now venture-scale, with Yann LeCun's AMI Labs raising $1.03B in March 2026 (NYT (opens in a new tab)). For accessibility the most useful sort is by purpose — models for acting, for generating experiences, and for understanding and answering. Assistive technology mostly needs the first and third, running from the user's own vantage point and in real time; the most visible world models are the second kind.

Current capability limits fit assistive needs poorly. V-JEPA 2-AC, post-trained on under 62 hours of robot video, planned zero-shot pick-and-place of new objects in unseen labs with 65–80% success, but it took about 16 seconds per action against roughly four minutes for a Cosmos-based planner (arXiv 2506.09985 (opens in a new tab)) — orders of magnitude slower than the sub-second guidance a blind pedestrian or powered-wheelchair user needs. The generators that do run in real time, such as Genie 3 at 24 fps and 720p, hold a world together for only a few minutes and cannot reproduce real geography (DeepMind (opens in a new tab)). Across the board, models are weakest at the forward-looking judgment assistance demands: video models score at or near chance on IntPhys 2 while humans score 85–95% across Meta's physical-reasoning benchmarks, and on CausalVQA they describe what already happened far better than they anticipate what comes next (Meta (opens in a new tab); arXiv 2506.09849 (opens in a new tab)). Video diffusion models trained on simulated physics generalize by imitating the nearest training case rather than by learning laws (arXiv 2411.02385 (opens in a new tab)), which is bad news for anyone whose body or movement sits far from the training cases.

Explicit work is nearly absent, and the labs speak of care rather than disability

Targeted searches of arXiv and of ASSETS-, CHI-, UIST- and ICRA-indexed work found no peer-reviewed paper that frames an assistive system as a world model in the machine-learning sense. WorldScribe, the University of Michigan's UIST 2024 Best Paper, has "world" in its name but is a live describer that adapts to the user's intent and current context without forecasting (arXiv 2408.06627 (opens in a new tab)). The flagship announcements for Genie 3, V-JEPA 2, Cosmos, Cosmos 3, Muse and Marble never mention accessibility or disability, and a 58-page July 2026 roadmap for the field lists model-based reinforcement learning, video generation, robotics and physical AI with no assistive application at all (arXiv 2607.06401 (opens in a new tab)). This is absence from web-indexed literature, not proof that no such work exists; the October 2026 ASSETS proceedings could not be searched in full text.

The most concrete bridge is organizational rather than technical. In May 2025 Demis Hassabis called making Gemini a world model the critical step toward a universal AI assistant (Google (opens in a new tab)), while DeepMind separately co-developed Project Astra's Visual Interpreter with the visual-interpreting service Aira for blind and low-vision users, advertising multimodal memory and proactive responses (DeepMind (opens in a new tab)). The Astra page never uses the term world model, and no public document says one powers the Visual Interpreter. Fei-Fei Li's November 2025 essay on spatial intelligence imagines a home assistant helping an older person cook without eroding their autonomy and robots helping seniors who live alone, yet never uses the words disability or accessibility (Li (opens in a new tab)); LeCun ties world models to smart-glasses assistants (NYU CDS slides (opens in a new tab)). When world-model leaders invoke human benefit, they speak of aging, illness and care — a medical framing, not the rights-based social model disability organizations use. Meanwhile the term is drifting toward modeling people themselves, with an October 2026 startup pitching a world model of human behavior for market research (TechCrunch (opens in a new tab)). That drift matters because every assistive system needs a model of its user's abilities, intentions and timing, and that user model is where ableist defaults most easily enter.

Prediction adds six things perception cannot

Perception maps the present to a label or description; a world model adds time and action, and six assistive functions depend on that addition. Anticipation warns about the cyclist who is about to cross your path rather than the one already there. Timing decides when to interrupt, which means predicting both the scene and the user's attention. Counterfactual consultation answers what would happen if the user took the left door instead of the right, the capability CausalVQA shows to be weakest. Object permanence and memory answer where the keys went or whether this corridor has already been walked. Safe action on someone's behalf lets a feeding robot or GUI agent simulate consequences before acting, which matters most when the user cannot easily intervene or undo. And synthetic rehearsal generates rare cases, as Waymo already does with Genie 3 for tornadoes and elephants on the road (PCMag (opens in a new tab)) — an approach that could in principle produce the wheelchair users and guide-dog teams driving systems rarely see. Anticipation cuts both ways. ForeSightGuide, a 2026 blind-navigation system, treats cutting redundant alerts as a headline metric alongside catching hazards (arXiv 2608.18993 (opens in a new tab)), so a good world model earns as much by staying quiet about harmless futures as by warning about dangerous ones.

Disabled people already build world models, through other channels

Cognitive science points to a design target that vision-first AI misses. Blind people who learn layouts by touch show the same encoding biases and spatial updating as sighted people who learn by sight, which supports a shared amodal spatial representation rather than recoding into imagery (UMaine VEMI (opens in a new tab)). Reviews of navigation by congenitally blind people favor a convergent model in which differences shrink with experience, note that about 75% of people considered blind keep some vision, and warn that experiments often force sighted strategies on blind participants (Schinazi et al. (opens in a new tab)). Visual knowledge also travels through language: congenitally blind adults judge the meaning of verbs about light and sight just as sighted adults do (Bedny et al. (opens in a new tab)), and text-only word embeddings recover those meanings at correlations of 0.59 with sighted and 0.63 with blind judgments (Lewis et al., PNAS (opens in a new tab)). Early deafness selectively enhances peripheral and motion attention rather than vision across the board (Bavelier et al. (opens in a new tab)), and contested predictive-processing accounts describe autistic perception as weighting priors and prediction errors differently (van Boxtel & Lu (opens in a new tab); HIPPEA (opens in a new tab)). The implications are concrete even though the autism theories are unsettled and should not be read as deficit accounts. An assistive world model's output need not be visual and its internal state should not be vision-only; blind users' verbal knowledge is a rich channel rather than a degraded substitute for sight; and what counts as salient, and how much forewarning someone wants, should be set per person rather than fixed to sighted, hearing, neurotypical defaults.

The table summarizes where each area stood in October 2026; the evidence for each entry follows in later sections.

AreaShipped or deployedPilot or emergingWorld-model-like researchLargest unmet need
Blind and low visionBe My AI photo chat; Meta Live AI; Oko signal reading; WeWALK overhead alertsAstra Visual Interpreter with Aira (human-supervised); Glide; Shenzhen's Xiaosuan guide robotPedestrian forecasting (HEADS-UP, 67% collision warnings); ForeSightGuide; WalkVLM; StreetReaderAI session memoryWell-timed hazard anticipation with low miss rates
Deaf and hard of hearingSignapse digital signers for scripted content; caption glassesLive automatic signing for streamsSignGPT sign-to-sign; SpeechCompass localization; SoundTrace episodic memorySign recognition; anticipating who speaks next
Speech and AACApple Personal Voice; ElevenLabs Impact; phrase prediction; ASR improved by the Speech Accessibility ProjectModels personalized on a user's own logsSpeakFaster context expansion; vocabulary generated from scenesModels of the partner and setting that preserve authorship
Cognitive and neurodivergentTiimo AI plannerCrossSense dementia glasses (prize-winning, not peer-reviewed)Grounded VR rehearsal; AIroad; AI-CarePeer-reviewed outcomes; consent for models of the person
Motor and mobilityLUCI shared control; Wandercraft Eve (no AI claims)Stretch 4 pilots; 1X NEO (teleoperated); BCI feasibility trialsADA feeding; HHI-Assist; Georgia Tech exoskeleton; UCLA BCI copilotHome autonomy that hands control back well; contact care tasks
Autonomous vehiclesScreen-reader-accessible Waymo app for blind ridersDriverless wheelchair-accessible rides essentially absentWaymo World Model and Tesla simulator (no disability disclosure); PMMA datasetAudited detection and prediction of mobility-aid users
Digital (GUI, games, XR)TalkBack Ask Gemini; designer-authored game accessibilityAgentic browsers with poor screen-reader accessWMA state-change descriptions; A11y-CUA; OLLA; audio world modelsAgents that work through assistive tech; semantic state from generated worlds

Blind users expose the gap between describing and predicting

Shipped assistants watch continuously but rarely look ahead

By late 2026 every major platform ships an assistant for blind and low-vision users, and nearly all of them still wait to be asked. Be My AI remains photo-plus-chat, and Be My Eyes' 2026 growth came from human video calls launched hands-free from Ray-Ban Meta and Oakley Meta glasses, with Meta AI merely routing the call (Be My Eyes (opens in a new tab)). Meta's Live AI holds a continuous conversation about what the camera sees, but Meta's own accessibility page documents no navigation feature and no proactive hazard alerts, and lists battery life of four to eight hours (Meta (opens in a new tab)). Apple's May 2026 features, including Live Recognition through the Action button, start only when the user asks, and Apple's footnote warns against relying on VoiceOver or Magnifier for navigation or anywhere harm could occur (Apple (opens in a new tab)). Specialized devices stay reactive: Oko reads pedestrian signals on-device (EU Road Safety Charter (opens in a new tab)), and the WeWALK Smart Cane 2 uses ultrasound to warn of overhead obstacles (RNIB (opens in a new tab)). The one offering that combines memory with proactive speech, Astra's Visual Interpreter, ships through a trusted-tester waitlist in which Aira's human interpreters can check answers or take over (Aira (opens in a new tab)). Practitioners draw the same line: an American Foundation for the Blind review found Ray-Ban Meta glasses neither fast nor accurate enough for street crossings and no replacement for a cane (AFB (opens in a new tab)). The commercial frontier is continuous perception, not prediction, and providers that attempt proactivity do not yet trust it unsupervised.

Measured failures cluster where foresight would help

Recent studies show how far description-only systems fall short. In a two-week CHI 2026 diary study, 20 blind and low-vision people using a GPT-4o app got correct follow-up answers only 56.6% of the time, and 22.2% of responses contained false information, yet participants reported high satisfaction and trust (Cornell Chronicle (opens in a new tab)). When eight blind participants tested ChatGPT's live video at ASSETS 2025, every one tried to get continuous feedback and none succeeded: the model promised to alert users and then fell silent, one participant asked 18 times whether it could see an umbrella, and the model placed a nonexistent staircase behind one user and forgot that another had already turned around (arXiv 2508.03651 (opens in a new tab)). Participants said they would rather receive no information than misleading information, and the authors called for systems that learn when to intervene instead of waiting their turn in conversation.

Other studies converge on the same weaknesses. A CHI 2026 analysis of 418 interactions found weak estimates of distance and time and little help aiming the camera (arXiv 2602.16930 (opens in a new tab)). The EgoBlind benchmark, built from 1,392 videos filmed by blind people, put the best of 16 multimodal models at about 60% against 87.4% for humans, with failures concentrated in navigation, temporal reasoning and tracking orientation as the wearer moves (arXiv 2503.08221 (opens in a new tab)). A direct comparison of a voice agent with a remote sighted assistant found the agent lacked the unprompted, sight-triggered actions human helpers rely on (arXiv 2602.05671 (opens in a new tab)), and an ICRA 2026 benchmark found that off-the-shelf models give badly timed navigation instructions that distract users and add cognitive load, while training them to predict the reason behind each instruction helped (arXiv 2609.05596 (opens in a new tab)). Spatial, metric and temporal reasoning is the most consistent weakness across this literature, which suggests world models aim at the right problem. The same evidence exposes a new hazard. An assistant that implies it is watching for hazards creates an implied safety promise, and breaking that promise through silent omission is worse than offering no service at all.

Forecasting prototypes work modestly, and silence is half the problem

Explicit forecasting for blind mobility exists, mostly without the world-model label. Early systems estimated time to collision from a suitcase-mounted camera (BBeep (opens in a new tab)) or projected nearby pedestrians four seconds ahead at constant velocity (Kayukawa et al. (opens in a new tab)). HEADS-UP (2024) contributed the first head-mounted dataset built for trajectory prediction in blind assistance, with a system that runs in real time on a Jetson (arXiv 2409.20324 (opens in a new tab)). A 2025 collision-risk model adapted from driver assistance reached 67% warning accuracy against 51% for distance and time-to-contact baselines, though it was never tested with blind users (arXiv 2506.16219 (opens in a new tab)). ForeSightGuide (August 2026) reports 0.299 redundant alerts per guidance output and a missed-hazard rate of 0.112 (arXiv 2608.18993 (opens in a new tab)), and the ICCV 2025 WalkVLM system exists largely to decide whether a reminder should fire at all (ICCV (opens in a new tab)). General embodied AI has meanwhile moved on to generative navigation world models that forecast several possible futures, such as NavWM and WorldMAP, none applied to blind users (NavWM (opens in a new tab); WorldMAP (opens in a new tab)). This forecasting gap between AV-grade prediction and blind mobility aids is one of the clearest unclaimed opportunities, but 67% warning accuracy cannot serve as the only safety layer, and false alarms consume the same audio channel blind travelers use to hear traffic.

Robotic guides suggest a different output channel. Glidance's Glide steers the user with unpowered wheels and a haptic handle and remains in waitlist and demo-tour mode with no published evaluation (glidance.io (opens in a new tab)). Shenzhen Metro's Xiaosuan guide robot served 87 visually impaired passengers with no safety incidents over a six-month pilot, but inside a structured station with tactile paving (NewsGD (opens in a new tab)). The Co-Ego quadruped cut mean collisions from 3.33 to 1.33 by fusing a chest-mounted phone's view of head-height hazards with the robot's map, though only six blindfolded sighted participants tested it (arXiv 2603.20121 (opens in a new tab)), and guide-dog handlers in a separate study found quadruped noise and jerky gait especially disruptive because blind people navigate by environmental sound (arXiv 2505.11808 (opens in a new tab)). A steering handle conveys a forecast without spending speech bandwidth, yet the evidence base across all of this work remains thin, built on small samples, sighted stand-ins and lab routes.

Street imagery already works as a pre-travel world model

The most practical world model blind people can use in 2026 is a long-context multimodal model running over street-level panoramas. Google's StreetReaderAI let 11 blind screen-reader users explore more than 350 panoramas; of 816 questions, 86.3% were answered correctly and 3.9% incorrectly, and a session memory of up to about a million tokens let it answer questions about earlier views (Google Research (opens in a new tab)). Apple and Columbia's SceneScout produced descriptions that were 72% accurate, with subtle, plausible errors that are hard to catch without sight (Apple ML Research (opens in a new tab)). The safety detail that matters most is that 62.5% of StreetReaderAI's errors were false negatives, such as a missed bike rack; an unmentioned object leaves a blind user nothing to probe. Street imagery is geo-registered and covers whole cities, but it goes stale, so construction and other temporary hazards vanish. No group was found capturing specific indoor venues for blind users with Gaussian splats or NeRFs, an open niche given that metric geometry is precisely what current models get wrong. Audio description faces a parallel timing problem, since automated live description must predict gaps in commentary to avoid talking over it (CSUN (opens in a new tab)). The 2026 FIFA World Cup's haptic boards, which rendered ball position by touch (FIFA (opens in a new tab)), offer a precedent for presenting a world model's state through a channel other than speech. No video world model was found anticipating on-screen events for description.

Sound, speech and memory need models of people, not only places

Sign technology models moving bodies, not the situations that give signs meaning

Commercial sign generation in 2026 mostly applies photorealistic digital signers and avatars to scripted, predictable content. Signapse says it generates more than 5,000 British Sign Language train announcements daily across the UK (Signapse (opens in a new tab)) and has begun live automatic signing for streams with SyncWords (PR Newswire (opens in a new tab)). Research is moving toward large pose-based corpora such as SignVerse-2M, roughly 2 million clips across 55+ sign languages gathered from public video with no stated consent arrangements (arXiv 2605.01720 (opens in a new tab)), and toward unified models such as SignGPT, whose sign-in, sign-out pipeline was explored with 12 Deaf ASL signers (arXiv 2609.21709 (opens in a new tab)). Evaluation lags far behind: the CVPR 2025 sign production challenge drew 231 entries scored by BLEU and pose error on German weather-forecast glosses with no Deaf evaluators (arXiv 2508.06951 (opens in a new tab)), and a 2026 diffusion baseline performed about as well with shuffled prompts as with the correct ones (arXiv 2607.13164 (opens in a new tab)). Where world models appear in sign technology, they model how a signer's body moves. The World Federation of the Deaf and the World Association of Sign Language Interpreters hold that word-for-sign translation is impossible without context and cultural norms, and oppose avatars replacing human signers for live, complex or high-stakes information (WFD (opens in a new tab)). That context is exactly what a situation-aware world model would have to supply, and nothing in current benchmarks measures it. Photorealism can also outrun meaning: a convincing signer saying the wrong thing can do more harm than an obviously synthetic avatar, because viewers trust it more.

The direction of effort is also inverted. Generation leads commercially, but most of the value for Deaf people lies in being understood. Google announced SignGemma for sign-to-text translation in May 2025; in January 2026 staff said there was no release timeline, and users were still asking on Google's forum in late October 2026 (Google AI forum (opens in a new tab)). Lara Draper of Deaf Aotearoa called sign-to-text too complex for now and warned that projects fail without Deaf involvement from the start (NZ Herald (opens in a new tab)), and a Georgia kiosk maker scaled back its AI after Deaf community pushback (Hoodline (opens in a new tab)). A Deaf-led review of 101 sign-language AI papers found an overemphasis on perceived communication barriers, unrepresentative data and little meaningful Deaf input (Desai et al. (opens in a new tab)).

Sound awareness is the clearest case of world modeling beyond transcription

For Deaf and hard-of-hearing users, the most world-model-like work models the acoustic environment and its history rather than just the words. Google's SpeechCompass (CHI 2025 Best Paper) adds real-time speaker direction and separation to mobile captions with 11–22° localization error, roughly human level, while using no video or speaker embeddings and letting users mute chosen directions; a survey of 263 frequent caption users found current tools struggle to tell speakers apart (Google Research (opens in a new tab); arXiv 2502.08848 (opens in a new tab)). Michigan's Soundability Lab has built sound recognition with temporal context and episodic memory (SoundTrace, IMWUT 2026) alongside rich auditory scene descriptions (Soundability Lab (opens in a new tab)), and caption glasses already ship, including Meta's Ray-Ban Display (HearingTracker (opens in a new tab)). What is missing is anticipation. No system was found that predicts who is about to speak, even though listeners' head movements signal approaching turn ends (Frontiers in Psychology (opens in a new tab)). A critical ASSETS 2026 review of 53 XR studies adds a caution: most frame accessibility as helping Deaf people cope in spaces designed for hearing people, placing the burden of adaptation on the Deaf user's device rather than on the environment (arXiv 2609.11706 (opens in a new tab)).

AAC predicts the conversation but not the partner

Predictive AAC is moving from conversation history toward scene and personal context. Google and Team Gleason's SpeakFaster expands abbreviations using dialogue context, saving 57% more motor actions than conventional predictive keyboards in simulation and raising text-entry rates 29–60% for two ALS eye-gaze users (Nature Communications via PMC (opens in a new tab)). GPT-4V can generate visual-scene-display vocabulary from a photo, though not personalized to the user (arXiv 2408.11137 (opens in a new tab)), and design probes with 11 people with aphasia included tools for preparing future conversations but sometimes misread intent (arXiv 2504.09435 (opens in a new tab)). An AAC user who fine-tuned a model on seven months of his own communication found that it amplified or muted parts of his identity and resurfaced private details out of context (arXiv 2509.13671 (opens in a new tab)). Shipped AI in AAC is mostly phrase prediction and neural voices (Tobii Dynavox (opens in a new tab)). The missing world model is one of the communication partner and setting — who the partner is, what they already know, what is happening around them — which is what skilled human partners and speech-language pathologists track.

Speech recognition shows that data, not model scale, is the binding constraint. The Speech Accessibility Project's Interspeech 2025 challenge reached a best word error rate of 8.11% (arXiv 2507.22047 (opens in a new tab)), Microsoft reported 18–60% accuracy gains from the project's data (EurekAlert (opens in a new tab)), and Zoom self-reported 5.72% WER on that speech in July 2026 against more than 20% for one competitor (Zoom (opens in a new tab)). But an Interspeech 2026 study found that leaderboard models' error rates rise 2–5× on stuttered and dysarthric speech, with hallucination bursts up to 24× (Interspeech 2026 (opens in a new tab)), and fluent but wrong output is especially hard to catch when the speaker cannot easily correct it. Voice restoration is the most mature consumer area, shifting from long voice-banking sessions to cloning from minutes of audio, as in ElevenLabs' free program for cloning a voice before mouth or head-and-neck cancer takes it away (ElevenLabs (opens in a new tab)) and Apple Personal Voice, whose setup was reportedly cut to 10 phrases (TechRadar (opens in a new tab)).

Cognitive assistants model routines, but the evidence is tiny and unreviewed

The strongest signal for a world model of a person's daily life comes from dementia care. CrossSense's Wispy smart glasses recognize household objects, learn the wearer's routines as dementia progresses and prompt the next step while leaving choices to the user; the system won the £1M Longitude Prize on Dementia in March 2026 (Med-Tech Insights (opens in a new tab)). Its headline evidence — object naming rising from 46% to 82% across 23 person-caregiver pairs — is reported secondhand and not peer-reviewed (Optimist Daily (opens in a new tab)). Other assistants rest on even less: a voice agent for Alzheimer's care that asks clarifying questions instead of guessing and grounds medication answers in caregiver-verified records was piloted with four people (arXiv 2605.08480 (opens in a new tab)), and an in-home reminder study warned that reminder logs are not neutral records of behavior and should keep uncertainty visible (arXiv 2604.19574 (opens in a new tab)).

Rehearsal simulations show that grounding the AI in a physical task helps. In a VR café, autistic trainees and job coaches preferred an LLM partner that knew which objects the trainee was holding, though trainee preference was borderline at p = .050 with nine trainees (Li & Peiris (opens in a new tab)), and 14 autistic children got better at reading drivers' social intent in an LLM-simulated traffic environment (arXiv 2502.03447 (opens in a new tab)). Previews of new places, which reduce anxiety for some autistic children, exist as VR tours of real venues (The Transmitter (opens in a new tab)), but no generated preview of a specific real place was found. Simulation is used more to model the disabled person for caregivers than to give disabled people a model of the world: one LLM simulator of a person with dementia was rated plausible by experts who nonetheless rewrote its replies on 54.5% of turns (arXiv 2603.29856 (opens in a new tab)). The Autistic Self Advocacy Network has reportedly urged organizations to stop using generative AI for plain-language translation after finding meaning-changing errors (ASAN (opens in a new tab)), a reminder that the people to whom simplification is most marketed are least represented in its evaluation. Tiimo's AI planner for neurodivergent users was Apple's 2025 iPhone App of the Year, an editorial honor with no efficacy evidence behind it (Yahoo/CNET (opens in a new tab)), and the memory features of mainstream assistants have not been studied with people who have dementia, brain injury or intellectual disability.

In robots, wheelchairs and cars, prediction succeeds as a copilot

Care robots use foundation models for planning, not physics

Physically assistive robotics remains research with very few disabled participants. Feeding is furthest along: the University of Washington's ADA system ran a five-day in-home deployment in which a community researcher fed himself 10 meals with about 80% success at acquiring the main dish, and the lesson was that customization and variable autonomy made home use work (UW HCR Lab (opens in a new tab); UW News (opens in a new tab)). Cornell's FLAIR plans bite sequences with vision-language models but was tested on 42 people without mobility limitations and one care recipient (FLAIR (opens in a new tab)), and the RABBIT bed-bathing study included one participant with severe mobility limitations among 12 (RABBIT (opens in a new tab)). Foundation models enter at the planning layer, while physical contact with bodies still depends on task-specific learning and force sensing. The most world-model-like caregiving work is HHI-Assist, a diffusion model that jointly predicts caregiver and care-receiver poses from motion capture (arXiv 2509.10096 (opens in a new tab)); no deployed assistive product was found using an explicit learned-dynamics planner.

Platform companies invoke eldercare far more than they demonstrate it. Elon Musk said at Davos in 2026 that humanoids will handle eldercare (iProUP (opens in a new tab)), and Apptronik raised $350M partly to expand into eldercare and healthcare while its deployments stay industrial (DC Velocity (opens in a new tab)). What has been shown is narrower: Gemini Robotics 1.5 scores 0.54 on action generalization (arXiv 2510.03342 (opens in a new tab)), Physical Intelligence concedes that π0.5 does not always succeed on the first try in unseen homes (PI (opens in a new tab)), and 1X's $20,000 NEO relies on remote operators in VR for unfamiliar chores, with its CEO tying improvement to collecting users' data (BusinessWire (opens in a new tab); Notebookcheck (opens in a new tab)). No commercial humanoid has shown transfers, toileting, bathing or dressing. Hello Robot's Stretch 4, priced at $29,950 and sold only for pilots, keeps humans in the loop; its designers worry about humanoid safety in homes, and paralyzed design partner Henry Evans points out that a wheeled robot freezes on emergency stop while a two-legged one can fall (IEEE Spectrum (opens in a new tab)). The UW lesson implies that world models will add most value here by predicting failure and handing control back, not by removing the user from the loop.

Shared control is where prediction has actually shipped

The predictive mobility technology people use today arbitrates between user and machine rather than replacing the user. LUCI, a power-wheelchair add-on, fuses stereo, infrared, ultrasonic and radar sensing to block collisions and drop-offs while the user steers, and its makers stress that it is not a self-driving car (Rehab Management (opens in a new tab)). Georgia Tech's exoskeleton controller models the wearer's internal state, estimating hip and knee joint moments in real time across 28 activities with no per-user calibration (Georgia Tech (opens in a new tab)), and prosthesis research uses cameras with uncertainty estimates so a leg can hold back when it is unsure of the terrain ahead (Futurity (opens in a new tab)). Wandercraft's Eve became the first self-balancing exoskeleton cleared by the FDA for personal home use in August 2026, supervised by a trained companion and marketed with no AI claims (Wandercraft (opens in a new tab)).

The clearest case of combined intent and environment prediction comes from UCLA, where a camera-based AI copilot inferred the user's goal from noisy EEG decoding plus the scene; with it, a paralyzed participant finished a robotic-arm task in about 6.5 minutes, and without it he could not complete the task (UCLA (opens in a new tab)). Synchron's plan to pair a foundation model of neural signals with Cosmos-generated simulated homes remains, by the company's own account, speculative (MassDevice (opens in a new tab)), and implanted BCI users number only in the dozens — Neuralink reported 21 participants in January 2026, Synchron has implanted 10 patients, and Paradromics performed its first implant in June 2026 (WHBL/Reuters (opens in a new tab); BioWorld (opens in a new tab); BusinessWire (opens in a new tab)). User studies agree on the governing principle. Assistive-arm users generally want to keep control (Pascher et al. (opens in a new tab)), and blind participants guided by a museum robot over repeated sessions felt that full autonomy reduced their sense of control and gradually formed clearer preferences about when to rely on it (arXiv 2601.19851 (opens in a new tab)). A trickle of intent suffices when a scene model fills in the detail, which is the bargain shared autonomy offers.

Autonomous vehicles serve blind riders but cannot show they see wheelchair users

Robotaxis already work as accessibility technology for blind riders: Waymo's app supports screen readers, and NFB members have ridden fully driverless vehicles (Waymo Help (opens in a new tab); NFB (opens in a new tab)). For wheelchair users, driverless service barely exists. Waymo's accessible vehicles are human-driven (Waymo Help (opens in a new tab)); a powered-wheelchair user petitioned in August 2026 against Nevada's approval of 1,000 Waymo vehicles in Las Vegas because the company has no accessible-vehicle contract there (Streetsblog (opens in a new tab)); Tesla's robotaxi fleet is not wheelchair-accessible and its Cybercab requires riders to transfer out of their wheelchair (Electrek (opens in a new tab); Fast Company (opens in a new tab)); and May Mobility's accessible vans keep onboard operators because wheelchair securement is still unsolved (May Mobility (opens in a new tab)). That blocker is hardware and regulation, not AI.

Prediction is the AI problem, and nobody has audited it. Waymo's World Model, announced in February 2026 and built on Genie 3, generates camera and lidar data for rare scenarios and can control how other road users behave, but Waymo has published no benchmarks and nothing about simulating wheelchair, cane or guide-dog users (The Decoder (opens in a new tab)); Tesla's neural world simulator, described mostly in secondary reports, likewise discloses nothing about vulnerable road users (iTiger (opens in a new tab)). The disability coalition's February 2026 letter to the Senate warned that "not all AVs are being taught to detect people seated in their wheelchairs" and proposed a rule testing how automated driving systems sense and respond to disabled people outside the vehicle (DREDF/CCD (opens in a new tab)). The research base supports the worry without measuring it. The PMMA dataset's authors note that KITTI, Waymo Open, nuScenes and JAAD rarely give mobility-aid users fine-grained labels (arXiv 2602.10259 (opens in a new tab)); eight pedestrian detectors proved biased against side views and hidden lower bodies, which plausibly affects seated wheelchair users although mobility aids were not tested (arXiv 2509.26166 (opens in a new tab)); another fairness test of eight detectors measured a 20.14% higher miss rate for children than adults but never examined disability (arXiv 2308.02935 (opens in a new tab)); and a DREDF brief recounts a model that failed to recognize a woman propelling her wheelchair backward with her feet and indicated the vehicle should proceed (DREDF (opens in a new tab)). No public, peer-reviewed audit of any production AV's detection or trajectory prediction for mobility-aid users was found. Generative world models cut both ways here. Trained on logs where such road users are rare, they may generate too few of them or animate them with walking-pedestrian dynamics, carrying the gap into the validation loop itself; yet controllable scene generation is exactly the tool that could synthesize reversing wheelchairs, cane sweeps and guide-dog pairs for testing, if companies chose to build such suites and disclose them.

Generated worlds and GUI agents arrive without accessibility semantics

Pixel-stream worlds have no accessibility tree

Generative interactive worlds in late 2026 are keyboard-driven pixel streams lasting seconds to minutes. Project Genie, the consumer prototype built on Genie 3, caps generations at 60 seconds for US Google AI Ultra subscribers (Google (opens in a new tab)), and no shipped world model was found offering screen-reader output, audio description, a semantic state interface or alternative inputs. The problem is structural. Accessible games work because designers author semantic hooks: The Last of Us Part II's 60+ accessibility settings and its blind and low-vision preset depend on the engine knowing which objects are enemies, where the critical path runs and which ledges can be climbed (PlayStation (opens in a new tab)). A world model whose consistency emerges from learning rather than from stored state has no object list to expose, so accessibility would need either a decoder that turns latent state into a scene graph or a describer running on output frames — and frame describers are already noisy enough that some low-vision players switch them off (GamerAstra (opens in a new tab)).

Two trends point the other way. Systems that export meshes or Gaussian splats, such as Marble and Tencent's HunyuanWorld, move generated worlds into game engines where navigation meshes, audio sources and labels can be attached, turning accessibility into a build step (World Labs (opens in a new tab); The Neuron (opens in a new tab)). Spatial sound that responds to the user's actions is also emerging in Helix World (self-reported results), WorldSonus and BinauralVAE, whose authors name scarce dynamic binaural recordings as the bottleneck (Noiz (opens in a new tab); arXiv 2609.06837 (opens in a new tab)); audio-first worlds cannot exist without it. Spatial-computing headsets offer a template, since they already maintain explicit, queryable scene meshes and anchors and Vision Pro runs VoiceOver in 3D space, though generative world models are not part of any headset's accessibility stack (Disability World (opens in a new tab)). On the motor side, a Sony patent application published around September 2026 describes AI that completes partial controller inputs for players who cannot fully press a button or move a stick through its range (Let's Data Science (opens in a new tab)). World models trained on typical players' controller logs, such as Muse's seven years of recorded gameplay (Microsoft Research (opens in a new tab)), may fail to map slow or partial inputs to their outcomes, an inference nobody has tested.

GUI world models already describe state changes, but agents fail under assistive constraints

The digital world model most immediately useful to disabled people may be the least publicized. WebDreamer uses an LLM to imagine the outcome of each candidate web action before committing (arXiv 2411.06559 (opens in a new tab)), and WMA (ICLR 2025) works over accessibility trees and predicts plain-language summaries of what will change after an action (arXiv 2410.13232 (opens in a new tab)). That output is functionally the verification aid a screen-reader user lacks — a spoken preview of what a click will do and a spoken summary of what it did — yet no paper frames it that way. The agents built on such models perform poorly under assistive conditions. On the CHI 2026 A11y-CUA dataset, a state-of-the-art computer-use agent completed 78.3% of everyday Windows tasks by default, 41.67% when restricted to the keyboard and 28.3% under magnification, because agents learn from sighted, mouse-centric behavior (arXiv 2602.09310 (opens in a new tab)). In a three-week diary study, eight blind screen-reader users issued 1,258 commands through an accessible layer over several agents; the best, GPT-5, completed 52.5%, failing through actions on the wrong interface element, paths hidden in menus and skipped final saves, while users asked for explanation before action and confirmation before consequential steps, and some stopped delegating sensitive tasks such as banking (arXiv 2609.00524 (opens in a new tab)). Mainstream products sit at both ends of the spectrum: TalkBack's Ask Gemini answers questions about on-screen content (Google (opens in a new tab)), while one accessibility specialist judged ChatGPT Atlas virtually unusable with a screen reader (Double Tap (opens in a new tab)).

The asymmetry is sharp. A sighted user can glance to check an agent's work, while a blind user must re-traverse the interface with a screen reader, so agents that act quickly and silently leave behind an audit burden the user cannot easily discharge. Accessibility testing shows which kind of simulation deserves trust. ScreenAudit drives a real TalkBack screen reader and caught 69.2% of errors against 31.3% for a common checker (arXiv 2504.02110 (opens in a new tab)), whereas LLM personas of disabled people flatten and misrepresent them, behaving more like an outsider's imitation than an insider's account (Wang et al. (opens in a new tab)). Simulating the assistive technology can be checked against reality; simulating the disabled person cannot.

A learned prior over "normal" is the central risk

The model learns that disabled people are improbable

A world model is by construction a learned prior over what normally happens, which makes the disability-studies critique that AI treats disabled people as outliers apply twice over (AI Now (opens in a new tab)). The training corpora are silent on disability: Ego4D's 923 participants come with no public breakdown by disability or assistive-device use (Ego4D (opens in a new tab)), and the sources describing V-JEPA 2's million hours of video or Cosmos's pipeline for curating 20 million hours never say who appears in the footage. Where disabled people's own data exists, it reveals a sharp distribution shift. Blind creators' video in EgoBlind is tilted, off-center and poorly lit, and models answered it with hallucinated objects, sycophantic yes answers and false all-clear replies about blocked paths (EgoBlind (opens in a new tab)). Generative models already default to stereotypes: across 17 text-to-image models and 119,000 images, mobility-impairment prompts mostly produced wheelchairs and the stereotyped images scored higher on text-image alignment (INCLUDE-BENCH (opens in a new tab)), disabled focus-group participants found wheelchair users rendered as sad and lonely (Mack et al. (opens in a new tab)), and AccessEval found systematic disability bias across 21 LLMs and nine disability types (AccessEval (opens in a new tab)). No disability audit of text-to-video or world models was found, though they share backbones with these systems. Jutta Treviranus's argument that statistical reasoning harms people at the edges of a distribution (W3C keynote (opens in a new tab)) and Shari Trewin's observation that disability's extreme diversity resists group-fairness methods (arXiv 1811.10670 (opens in a new tab)) both predict the failure mode: a cane user's sweep, a guide dog's pace or a reversing wheelchair will read as noise unless the model's action space and priors come from those users' own ways of moving. The cleanest open gap is a study measuring whether trajectory prediction is worse for disabled pedestrians; none exists.

A hallucinated future is worse than a hallucinated description

Users act on predictions, and they cannot easily check them. Every disability group faces a version of this verification gap: a hearing official cannot check a photorealistic signer, a listener cannot check a cleaned-up transcript of dysarthric speech, a person with an intellectual disability cannot check a simplified text, a person with dementia cannot audit a reminder log, and a blind user cannot glance at an agent's work. Trust does not track accuracy. One 2024 diary study found low trust in a scene-description app (arXiv 2403.15604 (opens in a new tab)), while the 2026 study found high trust despite 22% false information, and autistic workers preferred an LLM's workplace advice to a human's even as a job coach flagged that advice as questionable (Jang et al. (opens in a new tab)). Predictive systems inherit the hallucination and sycophancy already documented in live assistants and add an implied promise to keep watching, which makes the ability to answer blocked, no or I cannot tell, and to display uncertainty, a core product requirement rather than polish.

A world model of a person is also a dossier

The features that make predictive assistance useful — always-on cameras, persistent spatial memory, voice capture, learned routines — are the same features whose privacy defaults moved against users in 2025–2026. Meta switched AI features on by default for its glasses and began storing US voice recordings with no opt-out (TechCrunch (opens in a new tab)), its next-generation glasses are reportedly planned with an always-on sensing mode (eMarketer (opens in a new tab)), and advocates quoted by Mother Jones describe accessibility as a Trojan horse for privacy-invasive products even as Meta funds 30 disability organizations and gives glasses to every blind US veteran (Mother Jones (opens in a new tab); Meta (opens in a new tab)). Disabled people carry a double burden. A blind user who depends on glasses to read mail cannot leave them at home, so opting in is a thin form of consent, and many disabled people already live under heavy surveillance through benefits investigations and Medicaid's electronic visit verification, which DREDF says creates an atmosphere of house arrest (DREDF (opens in a new tab)). For people with fluctuating capacity, proposed dementia-AI safeguards shift consent authority progressively to caregivers (PMC13529824 (opens in a new tab)), so the caregiver may end up authorizing the model of the user, and monitoring can push people into self-disciplining conformity with standards an AI sets (PMC13377237 (opens in a new tab)). Fixing underrepresentation creates its own tension: the European Disability Forum opposes the 2026 AI Omnibus provision letting developers process disability data without consent to detect bias (EDF (opens in a new tab)). Disability organizations want more data, but only on consented, community-governed terms.

Rules are loosening just as these systems arrive

Most assistive AI in daily use falls outside medical-device rules, and the FDA loosened its wellness and clinical-decision-support guidance in January 2026 (Foley (opens in a new tab)), leaving a US gap in which navigation claims rest on consumer disclaimers. The Justice Department pushed its ADA Title II web-accessibility deadlines to 2027 and 2028, citing among other reasons the limits of generative AI for remediation (Reed Smith (opens in a new tab)), and a December 2025 executive order targeted state AI laws including Colorado's anti-discrimination statute (DLA Piper (opens in a new tab)). The EU is moving the other way on liability: the revised Product Liability Directive treats software and AI as products under strict liability for anything placed on the market after 9 December 2026 (Norton Rose Fulbright (opens in a new tab)). Yet the AI Omnibus (Regulation 2026/1744) delayed high-risk obligations to December 2027 and August 2028 (Cuatrecasas (opens in a new tab)), and most wearable assistive AI will not count as high-risk anyway, leaving it with transparency duties and the AI Act's ban on exploiting disability-related vulnerabilities (sota.io (opens in a new tab)). The best-documented harm from predictive AI to disabled people is administrative rather than robotic: a class action alleges that UnitedHealth's nH Predict cut off post-acute care and that about 90% of denials appealed to federal administrative judges were reversed, claims the company disputes (HealthExec (opens in a new tab)). Pressure toward replacement completes the picture. Deaf organizations insist AI must never replace qualified interpreters in medical, legal and educational settings (Irish Deaf Society (opens in a new tab)), and Ashley Shew's critique of technoableism warns against technology framed as eliminating disability rather than serving disabled people (Norton (opens in a new tab)).

The opportunities sit at the edges of the distribution

The building blocks for most high-value applications already exist in separate places; what is missing is integration, evaluation with disabled users and governance. The table maps the strongest opportunities against what exists and what each still lacks.

OpportunityBuilding blocks that existWhat is missing
Anticipatory assistant on glasses with tuned alert thresholdsAstra-style memory and proactivity; WorldScribe-style adaptive verbosity; HEADS-UP and EgoTraj trajectory data; multi-future forecasting from NavWM-type models; Aira human fallbackBudgets for false alarms and misses; real-world evaluation with blind users; acceptable latency and battery on glasses
Forecasts delivered by steering rather than speechGlide, Xiaosuan, Co-Ego, quiet quadruped gaitsPerformance outdoors and in crowds; published user outcomes
GUI world models as screen-reader verificationWMA state-change summaries; A11y-CUA interaction traces; OLLA accessible layerSpoken previews and change summaries evaluated with blind users; agents trained on keyboard-only and magnifier behavior
Anticipatory captions and sound awarenessSpeechCompass localization without speaker embeddings; SoundTrace memory; head movements that signal turn endsNext-speaker prediction running on glasses
AAC that knows the partner and settingSpeakFaster; vocabulary generated from scenes; models trained on personal logsA partner model with user-controlled memory scopes and clearly marked AI authorship
Grounded rehearsal and previews of real placesVR role-play that knows what the user is holding; AIroad; mesh-exporting generatorsA generated preview of a specific venue, tested for anxiety reduction and real-world transfer
Disability-aware simulation for AVs and robotsWaymo World Model's control over road-user behavior; PMMA; WheelPose; HHI-AssistDisclosed test suites validated against real road users and care-receivers
Audio-first, semantic generated worldsHelix World, WorldSonus, BinauralVAE; Marble mesh export; patterns from The Last of Us Part IISemantic metadata generated alongside each frame
Shared-autonomy copilotsUCLA BCI copilot; LUCI; Georgia Tech exoskeletonOne model of user, device and environment that predicts its own failures

Five enablers cut across every row. The most important missing input is consented, disability-representative embodied data — first-person video, motion and depth streams from blind, low-vision, wheelchair, cane, walker and guide-dog users — gathered under agreements that pay contributors, allow revocation, place governance with the community and bar use by insurers and benefits agencies. The Speech Accessibility Project shows the model can scale, reaching about 2,000 participants by September 2025 (SAP (opens in a new tab)); new egocentric trajectory corpora such as EgoTraj give forecasting work a foothold (arXiv 2605.19004 (opens in a new tab)); and synthetic data has narrow wins, with WheelPose's Unity-generated wheelchair users improving pose estimation in the authors' own tests (arXiv 2404.17063 (opens in a new tab)). Crawling public video, as EgoBlind and SignVerse-2M did, proves demand but not consent. Benchmarks are the second enabler: they should reward correct negative answers and honest abstentions, penalize missed hazards, measure trajectory accuracy for mobility-aid users and atypical gait, test robustness to footage shot by blind wearers, and score sign output by Deaf comprehension rather than BLEU.

The remaining enablers concern control, privacy and power. Interaction patterns should follow the community's own principles — user-set autonomy levels, design around interdependence rather than independence from other people (Bennett et al. (opens in a new tab)), and seamless handoff to humans in the Aira mold. Privacy should be built into the architecture through on-device processing, expiring memory and bystander-protective defaults, as SpeechCompass does without speaker embeddings and as the ear-worn assistant called The Pen does by processing bounded sessions locally instead of capturing continuously (arXiv 2603.06564 (opens in a new tab)), with facial recognition off the table. And governance and funding should be led by disabled people: the lessons from Deaf co-creation projects call for redistributing power through Deaf leadership (De Meulder et al. (opens in a new tab)). Timing favors action now. The EU's harmonized standards and codes of conduct, drafted in 2026–2027, will decide whether disability testing becomes a real obligation; strict product liability from December 2026 will reward vendors that can document disability-inclusive testing; and NIDILRR's new research center on AI-driven assistive technology, with applications due in July 2026, offers a rare public counterweight to corporate grants (Grants.gov (opens in a new tab)).

The argument most likely to move mainstream world-model developers is scientific rather than charitable. Video world models are known to generalize by resemblance to training cases rather than by learning laws, which makes edge cases diagnostic. A model that correctly predicts a person propelling a wheelchair backward with their feet has learned something general about agents and motion; a model that fails has memorized typical motion. Treviranus's inversion — treating the edges of the distribution as the real test of a model — is the curb-cut case world-model labs are best placed to hear.

Conclusion

The first world-model capabilities most disabled people encounter will not carry the label. They will arrive as proactive features inside assistants and smart glasses, as simulators that certify robotaxis, and as agents that operate computers, so accountability has to attach to behaviors rather than architectures: whether a system promises to watch, how often it stays silent when it should speak or speaks when it should stay silent, what its miss rate is for cane and wheelchair users, and who can see the map it keeps of someone's home. The evidence also shows that the bottleneck has moved from perception to judgment — when to speak, when to act, when to hand back control — and that judgment requires a model of the user as much as of the world, which is exactly where disabled people's expertise is irreplaceable and where ableist defaults most easily enter.

Four signals will show whether the intersection matures responsibly. The experiences of Astra's trusted testers are the likeliest first contact between proactive world-model features and blind users. Ireland's cohort of 15,000 donated Meta glasses (PPC Land (opens in a new tab)) is a ready-made natural experiment that no one has yet committed to evaluate. The EU's strict-liability regime will test whether hazard-detection failures carry a price. And the clearest signal of all will be the first world-model lab to publish a disability data card or an evaluation on mobility-aid users — something none has done so far.

  • research
  • accessibility
  • AI