Peter FriedmanThe next accessibility revolution may begin when software stops waiting for disabled people to discover barriers, diagnose their own difficulty, locate specialist help, and repeatedly explain what has already gone wrong.
Call that missing capacity Realtime Consciousness as a Service, or CaaS: persistent, contextual attention that notices friction, asks before assuming, adapts safely, verifies outcomes, and summons accountable human help. [3]
The name is intentionally provocative, but the claim is practical. CaaS does not mean sentient software; it means supplying the functional attentiveness that fragmented digital services conspicuously lack today.[2]
The newest evidence says “not yet”
On 1 September 2026, Kodandaram and five colleagues reported 1,258 commands from eight blind screen-reader users across twelve desktop applications during a three-week study of computer-use agents. [2]
GPT-5 led the five tested models with 52.5 percent successful completion, yet the researchers concluded that current agents “remain unreliable,” documenting failures in grounding, planning, constraint-tracking, and termination. [2]
Among unsuccessful traces, grounding errors affected 22.6 percent across models, while hidden-path failures affected 20.7 percent. Partial GPT-5 attempts completed 68.3 percent of required steps before breaking down. [2]
This is CaaS’s starting warning. A system that appears observant but fails halfway may create greater danger for a disabled user precisely because it encourages reliance before abandoning the task. [2][17]
Accessible interfaces can now be generated
On 2 July 2026, Jerry, Moreno, Francisco and Hervas published an implemented architecture generating personalised, multimodal interfaces from structured profiles, adaptation rules and validated prompt templates. [3]
Their healthcare example adapted medication instructions for cognitive disability and hearing impairment, while SysML models supplied “explicit traceability” connecting user needs, adaptation rules and formal accessibility requirements. [3]
The architecture could vary language complexity, modality and visual structure, producing plain-language text, pictograms and high-contrast layouts while retaining WCAG 2.2 and EN 301 549 foundations. [3]
That is considerably closer to CaaS than another accessibility toolbar. It describes a governed transformation system whose decisions remain linked to declared needs, rules, standards and inspectable evidence. [3]
Real-time adaptation has proof, but not people
On 9 March 2026, Algamdi published AURA, a voice assistant adjusting speech rate, verbosity and language complexity during a session using observed replays, skips and listening duration. [4]
The simulation reported 63 percent fewer replays, approximately 51 percent fewer skips and 22 percent shorter task-completion time than a static baseline, with convergence exceeding 90 percent. [4]
Yet the author explicitly cautioned that the study “does not involve human participants.” Its results establish technical feasibility, not improved lived accessibility, trust, comfort or cognitive load. [4]
That qualification matters enormously. A convincing simulation can show that feedback loops operate; only disabled people using consequential services can show whether those loops understand, assist or quietly oppress. [4][5]
The participation gap remains serious
On 19 May 2026, Xu, Liu, Xia, Duan and Yu published a review of 117 peer-reviewed studies addressing adaptive human-AI interaction for neurodivergent people. [5]
They found multimodal interaction, adaptive feedback and embodied interfaces improved engagement and usability, but warned that “Most studies lack neurodivergent user participation” and seldom examine long-term consequences. [5]
Their evidence also exposed sensory and cognitive heterogeneity, accessibility barriers and dataset gender bias. Those findings challenge any system that converts a diagnosis into one supposedly appropriate interface profile. [5]
CaaS therefore cannot be disability profiling disguised as personalisation. It must respond to a person’s declared objective and present difficulty, because diagnoses neither describe every need nor remain constant across contexts. [5][8]
Accessibility 1.0 is not even finished
In February 2026, WebAIM tested one million prominent homepages and detected WCAG failures on 95.9 percent, averaging 56.1 detectable errors per page, 10.1 percent above 2025. [6]
Low contrast appeared on 83.9 percent, missing alternative text on 53.1 percent and missing form labels on 51 percent. Six recurring categories represented 96 percent of errors. [6]
WebAIM warned that “absence of detected errors does not indicate that a page is accessible.” Automated testing proves some failures; it cannot certify that a person can complete anything. [6]
CaaS cannot become an excuse for neglecting semantics, keyboards, labels, contrast or captions. Adaptive intelligence layered above defective foundations merely makes elementary failures more complicated, variable and difficult to audit. [1][6][16]
ReplyResearch moves the boundary
ReplyResearch’s Accessibility 2.0 argument distinguishes accessible content from accessible service. A usable form matters, but successful submission proves nothing about whether the promised adjustment, answer or action follows. [1]
On 19 August 2026, the Parliamentary and Health Service Ombudsman described a dyslexic claimant whom HMRC promised to telephone repeatedly, yet the call never came for fourteen months. [9]
The same report described a Deaf woman whose recorded British Sign Language requirement produced no interpreter. Staff then administered a different vaccination from the one she had booked. [9]
Those were not colour-contrast defects. They were failures of organisational attention: known context did not trigger timely action, verification, recovery or human ownership across the complete service journey. [1][9]
Why call the missing thing consciousness?
Ordinary software receives inputs, changes state and returns outputs. A genuinely responsive service must additionally maintain situational awareness: who needs what, what has happened, and what remains unresolved. [1][2]
“Consciousness” names this continuous attending function, not subjective experience. It combines perception, relevant memory, interpretation, choice, action, outcome-checking and escalation around the user’s present purpose. [1][3]
The metaphor earns its keep by exposing organisational unconsciousness. A request may exist simultaneously in a form database, inbox, customer record and queue while nobody recognises its unresolved human meaning. [1][9]
Nevertheless, Hacker News would correctly attack any article smuggling metaphysics into a product label. CaaS must be defined operationally, measured behaviourally and stripped of claims that cannot be falsified. [2][10]
What Hacker News contributes
In an August 2026 Hacker News discussion about interfaces for AI agents, one contributor identified the central tension: make automation disappear “without making its judgment opaque.” [10]
Another argued that agents require “human approvals, validation, training, and tuning,” while a third preferred gates around generated work because chat interfaces cannot provide sufficient control. [10]
That engineering instinct strengthens CaaS. Helpful adaptation should feel immediate, yet every consequential inference, alteration and action needs inspectable state, negative gates, understandable reasons and accessible reversal. [2][3][10]
Hacker News also resists universal interfaces. The thread repeatedly favoured task-specific tools, inspectable artefacts and individual workflows, opposing the notion that one chat box can mediate every complex activity. [10]
Accessibility is already an agent interface
In a June 2025 Hacker News exchange about an agent-friendly document model, a commenter recommended the accessibility tree because it preserves interactive objects and structural cues while filtering presentational noise. [12]
Another contributor, motivated by carpal tunnel syndrome, described building a low-latency computer-use agent through accessibility APIs, hoping to avoid memorising specialised voice-control syntax entirely. [12]
In a March 2026 Hacker News discussion, one commenter called accessibility a “stress test for design assumptions”; another said good semantic structure lets people override presentation safely. [11]
The unexpected implication is powerful: infrastructure built to expose interfaces to assistive technology also supplies machine-readable structure for agents. Accessibility is not technical debt; it may become agentic computing’s control plane. [2][11][12]
The CaaS service loop
A defensible CaaS begins with accessible foundations, then monitors only the signals necessary to support the chosen task. It should never infer disability merely from hesitation, mistakes or unusual navigation. [1][13]
When friction appears, the system offers an understandable option: explain this, simplify it, preserve my place, change modality, slow down, show context, or connect me to someone. [1][8]
After permission, it applies a bounded transformation whose source, purpose and effect remain inspectable. Deterministic rules should govern high-risk changes wherever probabilistic generation adds no genuine advantage. [3][4]
Finally, it verifies whether the task progressed. Failure triggers recovery or a human handover carrying user-approved context, preventing the familiar punishment of beginning again and repeating sensitive information. [1][9]
Adaptation can create fresh barriers
On 26 July 2026, Wanscher, Lorensen, Shafiq, Moghaddam and Alipour reported language-model repairs across twenty live websites, finding 24 improvements alongside 20 regressions in 100 usable trials. [7]
Their warning, “Conformance is not experience,” cuts both ways. Already accessible control pages produced fourteen improvements and fourteen regressions, showing that intervention itself can become an accessibility hazard. [7]
Their verified process detected all 57 seeded violations, produced no false positives and rejected 126 deliberately harmful candidates. Verification materially changed the safety profile of generated adaptation. [7]
CaaS therefore needs a do-no-harm gate before deployment, continuous outcome monitoring afterwards and one-action restoration of the prior state. Personalisation without reversibility is automated coercion. [1][7]
Preferences must outrank predictions
On 5 February 2026, W3C’s Cognitive Accessibility Research Modules described fluctuating needs involving memory, attention, language, communication, ageing, brain injury, neurodivergence, mental health, illness and medication. [8]
The breadth makes a single “cognitive mode” indefensible. One person may need simplification today, fuller detail tomorrow and no intervention whatsoever when familiarity makes the original interface faster. [5][8]
Users should control whether preferences apply once, throughout a journey, locally on one device or across services. Refusing memory must never mean refusing access to ordinary assistance. [1][13]
Prediction may justify an offer, never a diagnosis. “Would you like help?” preserves agency; silently recording that someone appears cognitively impaired creates stigma, privacy risk and potentially discriminatory decision-making. [5][13]
Realtime must not mean rushed
In CaaS, realtime means help arrives while it can still prevent exclusion. It does not mean automatic action occurs faster than a person can understand, refuse or interrupt it. [1][10]
The correct latency depends upon consequence. Enlarging a target can be immediate; changing medication instructions, submitting financial evidence or consenting to treatment requires confirmation, traceability and dependable recovery. [3][9]
A useful service distinguishes reversible presentation changes from semantic transformations and committed transactions. Each class requires different permissions, evidence thresholds, retention rules, monitoring and human-approval gates. [3][10][17]
Speed should therefore be measured alongside comprehension, completion, error, abandonment, reversal and confidence. A faster journey that conceals misunderstanding is not improved accessibility; it is accelerated failure. [1][2]
Human support is infrastructure
ReplyResearch’s defining contribution is refusing to treat human responsiveness as external to UX. If adaptation fails, the reachable and empowered person is part of the accessibility architecture. [1][9]
The human must receive only context the user approves, understand what automation attempted and possess authority to continue the journey. A powerless chatbot handoff merely relocates the barrier. [1][13]
Service teams consequently need response deadlines, ownership rules, escalation routes and outcome monitoring. Availability alone is insufficient: an unanswered accessible channel remains functionally inaccessible despite perfect markup. [1][9]
This is where CaaS differs from an overlay. It connects interface behaviour with operational fulfilment, following the request across departmental and technological boundaries until the intended human outcome occurs. [1]
Privacy is part of accessibility
In June 2025, the Royal Society reported evidence from more than 800 disabled respondents and roughly 2,000 British adults, identifying privacy, bias, affordability, obsolescence and infrastructure as adoption barriers. [13]
Its conclusion was direct: “Inclusive design (or ‘co-design’) practices are essential.” Disabled people must shape systems throughout development, receive usable information and see their feedback produce change. [13]
The more attentive CaaS becomes, the more intimate its data may become: confusion, fatigue, sensory preferences, reading patterns, communication needs, failed attempts and moments of vulnerability. [5][13]
Data minimisation must therefore precede personalisation. Prefer explicit choices and on-device state, separate preferences from identity, prohibit advertising reuse, publish retention periods and make deletion genuinely accessible. [1][13][17]
The law points towards service outcomes
On 5 August 2026, the Equality and Human Rights Commission updated its services Code, stating that equal service “does not necessarily mean” treating everybody identically. [14]
The Code describes anticipatory duties and reasonable adjustments involving practices, auxiliary aids, services and physical features. It supports preparation before disadvantage, although reasonableness remains contextual and legally specific. [14]
CaaS could help organisations operationalise anticipation by connecting declared needs with tested options, assigned actions and verification. It cannot certify legal compliance or replace case-specific human judgement. [1][14]
The legal danger is equally clear. Inaccurate inference, inaccessible consent, discriminatory profiling or failed automation may produce new disadvantages while creating persuasive but misleading records of apparent assistance. [7][13][14]
From compliance badge to observability
Accessibility programmes currently emphasise conformance evidence: audit reports, defect lists, statements and certificates. CaaS requires service observability showing whether real people reach outcomes across complete journeys. [1][6]
Useful measures include successful completion, time to effective help, repetition burden, adaptation acceptance, reversals, abandonment, comprehension, human-response latency, fulfilled adjustments and remaining unresolved cases. [1][2][9]
Metrics must be disaggregated by functional need, device, task and context without converting small groups into identifiable records. Average improvement can conceal severe exclusion affecting fewer people. [5][13]
Independent mystery shopping becomes especially valuable because internal dashboards confirm recorded process, not lived experience. A complete test should travel from first contact through response, adaptation, action and recovery. [1]
A research programme, not a product promise
CaaS should begin as falsifiable hypotheses. Does context-aware assistance improve completion without increasing errors, disclosure pressure, dependency, surveillance, cognitive burden or inequitable outcomes compared with accessible static controls? [2][4][7]
Trials need disabled co-designers, realistic consequential tasks, accessible consent, published adverse outcomes and long-term follow-up. Simulation is useful for engineering, but insufficient for judging dignity or trust. [4][5][13]
Every experiment should include already accessible control journeys, because testing only obviously broken interfaces rewards needless intervention. Successful adaptation means outperforming non-intervention without harming participants who needed no change. [7]
Researchers should publish system versions, models, prompts, adaptation rules, participant characteristics, assistive technologies, latency, failures, reversals and human interventions. Reproducibility is essential when interfaces change dynamically. [2][3][7]
What should be built next
First comes a portable, user-controlled context wallet storing preferences and active-task requirements without requiring diagnostic disclosure. It should reveal every stored item and support selective, immediate deletion. [1][13]
Second comes an adaptation contract: machine-readable boundaries describing what may change, what meaning must remain invariant, which evidence validates success and when human approval becomes mandatory. [3][7]
Third comes a service-resolution layer connecting interface events with accountable operational actions. Interpreter bookings, promised calls, alternative formats and deadline changes must become monitored outcomes, not passive notes. [1][9]
Fourth comes independent observability: external tests, accessible incident reporting, public limitations and auditable version histories. Organisations should prove improvement through completed journeys, not screenshots of adaptive features. [1][2][7]
The Hacker News test
Hacker News would ask whether CaaS is merely an inflated name for adaptive interfaces, agents, workflow engines and customer service. Without measurable boundaries, the criticism would be correct. [10][11]
The answer must be architectural integration. Existing components become CaaS only when perception, permission, adaptation, verification, accountable action and recovery operate as one inspectable service around user intent. [1][3]
Hacker News would also ask who controls the system, what happens offline, how state migrates, how failures surface and whether an open protocol can prevent another proprietary dependency. [10][12]
Those are not peripheral objections; they are the design brief. The most credible CaaS will be boringly interoperable, locally controllable, semantically grounded, observable and exceptionally easy to escape. [3][10][13]
What comes after disability UX
Disability UX traditionally improves interfaces for recognised impairments. CaaS shifts attention towards situated capability: whether this person can complete this particular task, under these conditions, with agency intact. [1][5]
That shift benefits people whose needs are permanent, temporary, episodic, undisclosed or simply situational. Yet the curb-cut effect must never erase disabled people’s leadership or redistribute resources away from them. [5][11][13]
The deeper change is from designing artefacts to maintaining relationships. Accessibility becomes a continuing promise that the service will notice difficulty, respond appropriately, verify progress and remain answerable when wrong. [1][9]
In that sense, CaaS is not consciousness sold by an API. It is institutional attentiveness rebuilt as accountable technical and human infrastructure, available at the precise moment exclusion begins. [1][2]
The disciplined version of a radical idea
Realtime Consciousness as a Service deserves survival only if it remains metaphorically bold and operationally modest. Current agents are useful, adaptive systems are promising, and neither is reliably trustworthy yet. [2][4]
The evidence supports bounded adaptation, semantic foundations, inspectable state, disabled co-design, data minimisation, verification, reversibility and prompt human recovery. It does not support autonomous benevolence or frictionless universal personalisation. [3][5][7][13]
ReplyResearch supplies the missing unit of analysis: the complete service journey. Hacker News supplies the missing attitude: show the state, expose the failure, build gates and distrust magic. [1][10]
Put them together and CaaS becomes a defensible proposition: not software that claims consciousness, but services finally engineered to keep paying attention until the person can genuinely proceed. [1][2][9]
Sources and relevant reading
[1] ReplyResearch, ‘Accessibility 2.0 Because Today’s Accessibility 1.0 Just Isn’t Accessible Enough’. Full URL: https://replyresearch.com/accessibility-2-0-because-todays-accessibility-1-0-just-isnt-accessible-enough/. Essay use: supplies the article’s distinction between accessible content and accessible service, and its service-loop framing. Source passage: the Accessibility 2.0 argument on adaptation, outcome testing and human recovery. Locate: read the article’s main sections on service delivery and accountability.
[2] Kodandaram et al., ‘Are We There Yet?’. Full URL: https://arxiv.org/html/2609.00524v1. Essay use: supports the opening warning about computer-use-agent reliability, blind screen-reader users and failure modes. Source passage: study results reporting 1,258 commands, model completion rates and the error taxonomy. Locate: HTML paper, abstract, Results and failure-analysis sections.
[3] Jerry et al., ‘LLM-Driven Accessible Interface’. Full URL: https://link.springer.com/chapter/10.1007/978-3-032-22937-3_4. Essay use: supports traceable multimodal adaptation, structured profiles and rules-based safeguards. Source passage: implemented architecture, SysML traceability and healthcare adaptation example. Locate: Springer chapter abstract, architecture figures and case-study sections.
[4] Algamdi, ‘A behaviour-adaptive AI assistant’. Full URL: https://www.nature.com/articles/s41598-026-43320-2. Essay use: supports the AURA simulation metrics and its limitation that no human participants were involved. Source passage: reported replay, skip and task-time changes; limitation statement. Locate: Scientific Reports article, Results and discussion of study limitations.
[5] Xu et al., ‘A scoping review of inclusive and adaptive human–AI interaction’. Full URL: https://researchportal.hw.ac.uk/en/publications/a-scoping-review-of-inclusive-and-adaptive-humanai-interaction-de/. Essay use: supports the review of 117 studies, heterogeneity of need and the participation gap. Source passage: review findings on multimodality, participation, bias and long-term evaluation. Locate: publication page and linked article abstract, findings and discussion.
[6] WebAIM, ‘The WebAIM Million’. Full URL: https://webaim.org/projects/million/. Essay use: supports the claim that basic accessibility defects remain widespread and automated testing is not certification. Source passage: 2026 homepage error rates, common failure categories and testing caveat. Locate: report’s 2026 results, ‘Home page complexity’ and conclusion.
[7] Wanscher et al., ‘From Blind Edits to Verified Repair’. Full URL: https://arxiv.org/html/2608.24913v1. Essay use: supports evidence that generated repairs can improve and regress accessibility, making verification and reversal essential. Source passage: live-site trial counts, seeded violations and harmful-candidate rejection. Locate: HTML paper, evaluation tables and verified-repair results.
[8] W3C Accessible Platform Architectures Working Group, ‘Cognitive Accessibility Research Modules’. Full URL: https://www.w3.org/TR/coga-research-modules/. Essay use: supports the account of fluctuating, intersecting cognitive accessibility needs and preference-led support. Source passage: research modules on memory, attention, language, communication and fluctuating conditions. Locate: module overview and the linked user-needs research sections.
[9] Parliamentary and Health Service Ombudsman, ‘Spotlight on accessible communication’. Full URL: https://www.ombudsman.org.uk/publications/spotlight-accessible-communication-your-stories-your-rights. Essay use: provides the service-failure examples involving an undelivered call and a missing BSL interpreter. Source passage: anonymised case stories and the report’s recommendations on recorded communication needs. Locate: ‘Your stories, your rights’ report, individual case-study pages.
[10] Hacker News, ‘What should the GUI for AI agents look like?’. Full URL: https://news.ycombinator.com/item?id=49119274. Essay use: informs the Hacker News perspective on inspectability, approvals, validation gates and task-specific tools. Source passage: comments debating opaque automation, chat limits and human control. Locate: linked discussion; use in-page search for ‘opaque’, ‘approvals’, ‘validation’ and ‘chat’.
[11] Hacker News, ‘Accessibility Issues Are Often Usability Issues’. Full URL: https://news.ycombinator.com/item?id=47254624. Essay use: supports the framing of accessibility as a stress test for design assumptions and semantic override. Source passage: contributor comments on design assumptions, semantics and usable controls. Locate: linked discussion; use in-page search for ‘stress test’ and ‘semantic’.
[12] Hacker News, discussion of accessibility trees and agent-friendly interfaces. Full URL: https://news.ycombinator.com/item?id=44331222. Essay use: supports the link between semantic interface structure, assistive technology and computer-use agents. Source passage: comments on accessibility trees, carpal tunnel syndrome and accessibility APIs. Locate: linked discussion; use in-page search for ‘accessibility tree’, ‘carpal’ and ‘APIs’.
[13] The Royal Society, ‘Disability technology’. Full URL: https://royalsociety.org/-/media/policy/projects/disability-technology/disability-technology-report.pdf. Essay use: supports the privacy, bias, cost and co-design discussion. Source passage: survey evidence from disabled respondents and recommendations for inclusive design. Locate: PDF executive summary, evidence chapters and recommendations.
[14] Equality and Human Rights Commission, Equality Act services Code. Full URL: https://www.gov.uk/government/publications/equality-act-2010-code-of-practice-for-services-public-functions-and-associations-2026/. Essay use: supports the account of anticipatory duties and reasonable adjustments in service provision. Source passage: Code guidance on equal service, anticipatory adjustment duties and contextual reasonableness. Locate: GOV.UK publication page and the downloadable Code’s services chapters.
[15] W3C Accessibility Guidelines Working Group, ‘WCAG 3.0 Working Draft’. Full URL: https://www.w3.org/TR/wcag-3.0/. Essay use: provides context for developing non-binary accessibility conformance and requirements. Source passage: framework description of outcomes, assertions and conformance. Locate: draft’s Introduction, Requirements and Conformance sections.
[16] World Wide Web Consortium, ‘Web Content Accessibility Guidelines 2.2’. Full URL: https://www.w3.org/TR/WCAG22/. Essay use: supports the foundational requirement to preserve perceivable, operable, understandable and robust content. Source passage: four accessibility principles and success criteria. Locate: recommendation overview and principles section.
[17] National Institute of Standards and Technology, ‘Generative AI Profile’. Full URL: https://doi.org/10.6028/NIST.AI.600-1. Essay use: supports the governance risks of confabulation, privacy, bias and over-reliance. Source passage: risk taxonomy and lifecycle-management guidance. Locate: PDF executive summary, risk tables and human-AI configuration guidance.
Footnote Zone for Realtime Consciousness as a Service
Disclosure: The diagnostic tools referenced below were developed by NokNok, a specialist in online responsiveness tool design.
This Footnote Zone uses NokNok’s diagnostic toolkit to examine how the responsiveness failures and inaccessible service journeys described in this article can be identified, measured, and addressed.
- Email Finder: Identifies the friction of user requests existing simultaneously in form databases without accountable human contact options, and scans an organization’s website and related public-facing materials for published email addresses, then reports on structural deficiencies, discrepancies, missing contact routes, or other contactability gaps.
- Reply Radar: Targets the failure of service teams to meet response deadlines and manage human queues, such as the dyslexic claimant whose promised contact never came for fourteen months, and deploys targeted test emails and quantitatively measures reply rates, latency, response consistency, and related responsiveness benchmarks.
- Compliance Sniffer: Addresses the unreliability of current automation agents, including grounding errors and failures to meet legal service outcome expectations, and analyzes incoming responses for objective quality, clarity, relevance, escalation, and compliance benchmarks.
- Mystery Shopper: Audits the complete service journey from first contact through response, adaptation, action, and recovery to prevent systemic end-to-end contact failures, and executes a comprehensive end-to-end responsiveness UX audit, testing how a real user experiences the organization’s contact, response, and escalation pathways.
Disclosure: The diagnostic tools referenced in this Footnote Zone were developed by NokNok, a specialist in online responsiveness tool design. ReplyResearch may use NokNok tools, resources, or analysis when preparing coverage, while retaining responsibility for its editorial decisions, including what topics to cover, what sources to cite, and how stories are presented. Read the full ReplyResearch Collaborative Disclosure Policy here.
