AI as Both Ramp and Staircase: When ‘Easy Technology’ Excludes Those Who Need It Most

The demo is flawless. A citizen asks a chatbot a question in natural language. The system finds the right form, simplifie s the instructions and fills several fie lds. The audience applauds. Somebody says, ‘This will make the service accessible to everyone.’

Now change the user.

She navigates by keyboard, not mouse. The chat window traps the focus. Its answer disappears before her screen reader finishes. When she asks for a person, the bot repeats the help article. The old telephone number has been removed because the new service is ‘easier’.

The same AI that looked like a ramp in the presentation has become a staircase placed across the only entrance.

Accessibility is not a property a product acquires by adding a microphone icon. It is a relationship between a person, a task, an environment and available support. The right question is not ‘Does the feature exist?’ but ‘Can this person achieve what matters, with dignity, control and a real way out?’

The promise is genuine: AI can be a ramp

AI can expand practical independence. Speech-to-text may help a person who cannot type comfortably. Text-to-speech and image description can make visual material available. Automatic captioning can open a live event. Plain- language rewriting may help somebody facing cognitive overload or an unfamiliar bureaucracy. Predictive input can reduce the physical effort of composing a message.

For journalists and educators, these functions can widen both production and participation. A reporter with a motor impairment can dictate notes. A student can ask for an alternative explanation without performing confusion in front of a class. A reader can move between audio, text and visual summary according to the day, device and context.

Piotr Plichta’s chapter in the Polish EU Kids Online 2026 report captures the double character of digital technology for young people with special educational needs. Assistive technologies can support development, communication, independence and social participation, while low accessibility, limited competence and insufficient support can deepen existing difficulties (2026, pp. 163–165). The report describes this as a context of potentially high reward and high risk, not a simple story of rescue or harm (p. 164).

That balance matters. If we speak only about risk, disabled and otherwise marginalised users become people to be protected from technology. If we speak only about opportunity, they become grateful figures in a product launch. In reality, they are users, workers, creators and citizens entitled to shape the conditions of use.

Why the ramp turns back into stairs

The first transformation occurs when an option becomes a gate. Voice input is helpful when it supplements a keyboard; exclusionary when voice is the only route in a noisy home, with a speech impairment or without privacy. A chatbot is convenient when it supplements a form and staffed line; coercive when it replaces them.

The second occurs through interface instability. Conventional services can be badly designed, but generative interfaces add unpredictability. The same request may produce different wording, order and length. A person with cognitive or learning difficulties may have to rediscover the route each time. ‘Conversational’ does not necessarily mean simple.

Third comes error asymmetry. A confident user may notice that a summary is wrong and repair it. A person relying on the system precisely because the original is inaccessible may not have an independent way to check. The supposed accommodation becomes a single point of epistemic failure.

Fourth is hidden cognitive labour. A system may use plain words but require the user to formulate a perfect prompt, remember several conditions, interpret uncertain advice and recover from a loop. The interface has removed visible buttons and replaced them with an invisible syllabus.

Fifth is dependency by design. Assistance can build capability, but it can also bypass the learning or judgement a person wanted to develop. Chodak and Filipek identify over-reliance and deskilling as a risk of generative AI (2025, p. 45). In the EU Kids Online discussion, Plichta gives the warning sharper human form: “Those who may lose most are pupils with weaker learning skills” (author’s translation, 2026, p. 166).

The question is not whether help is allowed. It is whether help enlarges the person’s room to act or quietly removes the furniture.

Capability: stop counting features and look at freedom

Amartya Sen’s capability approach distinguishes resources from the real opportunities people have to do and be what they value (1999, pp. 74–110). Two people can receive the same resource and have radically different freedom to convert it into an outcome. A ramp is useful only if it reaches the door, remains unblocked and leads somewhere the person has reason to go.

Applied to AI, this changes evaluation. A product team may count supported languages, accessibility settings or completed conversations. A capability audit asks whether a person can understand the decision, complete the task,

protect private information, correct a mistake, seek help and choose another route.

It also reveals conversion factors. A speech assistant may work technically but fail in a shared room where speaking reveals a health condition. A simplified text may be readable but remove the legal detail needed for informed consent. A free AI tool may require a modern device, stable connection and account that a user does not have. The feature is present; the capability is absent.

This perspective avoids the mythical average user. Disability, age, language, income, education and temporary circumstances intersect. Somebody may use a screen reader, speak a minority language and rely on an old phone. Another person may have excellent technical skills and experience fatigue that changes by the hour. Accessibility is not a small department serving a fixed category after the main product has been built.

The Polish AI-exclusion report makes a related institutional point: services should be designed from the ‘lowest threshold of entry’, involve end users in co-creation and provide human contact, appeal and understandable explanations (Kaniecki, 2026, pp. 37–39). Lowest threshold does not mean lowest ambition. It means the door is designed from the edge rather than widened later with a screwdriver.

Who pays for ‘easy’ technology?

The burden often shifts to the person least able to absorb it.

Consider a hypothetical student with dyslexia using an AI summary because the original article is difficult to process quickly. If the summary invents a citation, the student may submit it and be accused of carelessness. The institution gains efficiency; the student carries the credibility risk.

Consider an older resident whose benefit application is rejected after a chatbot misunderstands a response. The official service log records completion. A daughter spends an evening finding a route to appeal. The system has not removed labour; it has privatised it into the family.

Or consider a deaf journalist relying on live captions during a press conference. Incorrect names can be corrected later if the recording remains available and a colleague helps. If the newsroom treats the captions as final and deletes the source, accommodation has become evidence destruction.

These scenarios are hypothetical, but the distribution mechanism is familiar: when a ‘frictionless’ system fails, repair is performed by users, relatives, teachers, frontline staff and community organisations. Their time is rarely included in the efficiency calculation.

The EU Kids Online 2026 sample is a useful reminder not to imagine a tiny edge group. In Plichta’s aggregated analysis, 348 of 1,502 respondents — 23.2 per cent — reported at least one disability, health problem or learning difficulty included in the study’s special-educational-needs indicator. The author explicitly notes that this umbrella category is heterogeneous and analytically limited (2026, p. 167). Good design begins by retaining that diversity, not converting 348 people into one persona called ‘the vulnerable user’.

Human-centred means dignity, autonomy and appeal

Dignity requires help without humiliation. A user should not have to disclose a diagnosis to unlock a readable format, repeat a distressing account to successive bots or accept childish language in the name of simplicity.

Autonomy requires more than successful task completion. The person should know when AI is involved, what happens to their data, which parts can be corrected and whether the system is advising or deciding. They should be able to use assistance without surrendering the original, and decline assistance without losing access.

Responsibility requires a reachable human with authority. ‘Contact support’ is meaningless if support can only repeat the model’s answer. High-stakes public, educational, employment and health services need a route to a person who can inspect context, change the outcome and explain the remedy.

A route without AI is not nostalgia. It is resilience. Systems fail, devices break, models drift and people have legitimate objections to automated processing.

The alternative must be equivalent: no extra fee, punitive delay or lower-status treatment. A staircase beside a locked lift is not choice.

The Lowest-Barrier Gate: a practical accessibility audit

The Lowest-Barrier Gate is a release test for one real outcome. It should be run with people who use the service, not simulated by a design team taking turns to close their eyes.

1. Name the outcome and the stakes. Write ‘submit an appeal before the deadline’, not ‘use the chatbot’. Identify what happens if the person fails: inconvenience, lost learning, financial harm, public misrepresentation or danger. The higher the consequence, the stronger the fallback.

2. Recruit varied end users and pay them. Include people with sensory, motor, cognitive and learning needs; different languages and digital skills; older devices and slow connections. Work with disability organisations, but do not expect unpaid testimony. Record who was missing from the test.

3. Map every route, including no AI. Draw the journey for keyboard, touch, voice, screen reader, assisted use, conventional form and human contact. Mark where AI is mandatory, data leave the organisation or the user must switch modes. A path exists only if participants can find it without insider knowledge.

4. Test perception and operation. Check headings, focus order, labels, contrast, captions, transcripts, zoom, timeouts and compatibility with assistive technologies. Offer text and voice where useful, but never force one. Make generated updates predictable and announce them to screen readers without interrupting the user.

5. Test cognitive accessibility. Use plain language without deleting necessary meaning. Break tasks into stable steps, show progress and allow pause, review and return. Do not rely on memory of earlier chat turns. Explain errors specifically: what happened, what remains saved and what the person can do next.

6. Test truth and recovery. Seed realistic model failures. Can the user detect a wrong answer, view the source, correct the input and reverse the action? Preserve originals. Never let generated simplification become the only record of consent, testimony or evidence.

7. Measure capability, not applause. Record whether users independently reach the outcome, how long it takes, where assistance is required and whether they understand the consequence. Ask whether the tool increased control. A fast task completed by a facilitator is not independent success.

8. Protect privacy and choice. Explain AI involvement and retention in accessible language. Minimise sensitive data, provide a way to delete or correct it and avoid inferring disability when the person can simply choose a format. Accessibility preferences should not become advertising profiles.

9. Verify the human and non-AI routes. Put the telephone number, staffed desk, email or conventional form before the loop begins. Test response time, authority and handover. The user should not repeat the full account. Allow an authorised supporter while preserving the person’s own voice and consent.

10. Set a release threshold and publish evidence. Critical tasks fail the

gate if any tested group cannot complete them safely through at least one equivalent route. Name the owner, repair date and temporary alternative. Retest after model or interface changes and report unresolved exclusions.

There is no single lowest barrier

Accessibility needs can conflict. Animation may guide one person and overwhelm another. Plain language may aid comprehension while reducing legal precision. Voice can increase independence and destroy privacy. The answer is usually adaptable routes, not one ‘accessible mode’ banished to a settings basement.

Participatory testing can also become tokenism. A few familiar consultants cannot represent every disability or context. Organisations must combine

standards, expert review, user research, complaints and ongoing observation. They must be ready to hear that a celebrated feature should not launch.

Human alternatives can be fictional if underfunded. A staffed telephone line with a two-hour queue is not equivalent access. Conversely, assuming every person wants human help can be patronising. Some users prefer the privacy and pace of a machine. The ethical requirement is viable choice.

Nor can a local audit repair inaccessible operating systems, poor connectivity or poverty. The second level of digital exclusion concerns skills and use, but material access still matters. Collaboration, public funding and procurement standards are needed beyond one product team (Kaniecki, 2026, pp. 17–20).

Return to the flawless demo. Keep the chatbot if it genuinely helps. Add keyboard stability, source access, a conventional form and a human who can act. Pay the people testing it. Measure whether the resident completes the task, not whether the animation looks futuristic.

Then invite the woman with the screen reader back before launch and ask the only question that matters: can she enter, decide and leave on her own terms?

Technology becomes a ramp not when it looks easy from the stage, but when the person at the foot of the stairs has more than one real way through the door.