Public-Sector AI and Accessibility

By /

SWANK AI Guidance Note 07
Accessibility · Public-Sector AI · Human-Centred Systems

Core Standard

ACCESSIBILITY SHOULD SURVIVE AUTOMATION

AI should expand the ways people can participate.

It should not become a new condition of participation.

Where technology creates complexity, a meaningful human route should remain available.

Purpose

Artificial intelligence is increasingly entering public administration at the same time that members of the public are using AI to navigate increasingly complex administrative systems.

Accessibility therefore needs to be considered on both sides of the interaction.

AI can:

  • remove barriers;
  • create new barriers;
  • improve access to information;
  • increase cognitive burden;
  • support communication;
  • or make participation more dependent upon technology.

The relevant governance question is not simply whether AI is being used.

It is whether people can still participate effectively, understand what is happening and obtain meaningful human assistance where necessary.

AI as an Accessibility Tool

Generative AI may help people who experience difficulty with:

  • speech;
  • writing;
  • reading complex administrative language;
  • organising large quantities of information;
  • translation or language barriers;
  • cognitive load;
  • fatigue;
  • formulating questions;
  • navigating multiple policies;
  • understanding unfamiliar procedures;
  • or structuring correspondence.

For some individuals, AI may make participation substantially easier.

Its use should therefore not automatically be treated as evidence that a communication is:

  • less authentic;
  • less credible;
  • less personally understood;
  • or less deserving of substantive review.

The fact that technology assisted the communication does not determine the quality of the underlying information.

Accessibility Works in Two Directions

Public-sector AI governance should consider both:

AI Used by the Institution

For example:

  • chatbots;
  • automated triage;
  • summarisation;
  • decision-support tools;
  • translation systems;
  • case-management assistance;
  • document classification;
  • automated correspondence.

and:

AI Used by the Individual

For example:

  • drafting assistance;
  • translation;
  • organising evidence;
  • simplifying complex language;
  • structuring questions;
  • summarising their own records;
  • preparing correspondence.

A system may be technically accessible while institutional policy surrounding it creates a participation barrier.

Both sides should be examined.

Accessibility by Design

Public bodies introducing AI should consider whether the system:

  • makes information easier to understand;
  • preserves alternative communication channels;
  • supports written communication;
  • accommodates different communication styles;
  • increases or reduces cognitive burden;
  • explains consequential outputs clearly;
  • permits clarification;
  • provides meaningful human escalation;
  • creates barriers for people unable or unwilling to use AI;
  • and preserves appropriate routes for reasonable adjustment.

Technology should increase options.

It should not unnecessarily narrow them.

Do Not Create a Mandatory AI Interface

Public services should be cautious about making AI the only practical route to assistance.

People may be unable or unwilling to use AI because of:

  • disability;
  • digital exclusion;
  • literacy;
  • language;
  • privacy concerns;
  • lack of appropriate technology;
  • lack of confidence;
  • complexity of the issue;
  • safeguarding concerns;
  • or personal preference.

An automated interface may be useful.

It should not become an obstacle between an individual and a service that requires meaningful human engagement.

The AI Participation Divide

AI can create a new form of administrative inequality.

People who understand generative AI may become substantially better able to:

  • interpret policies;
  • draft formal correspondence;
  • organise evidence;
  • locate relevant information;
  • prepare questions;
  • navigate procedures;
  • and challenge unclear decisions.

People without those skills or tools may become comparatively disadvantaged.

Public bodies should therefore avoid assuming that everyone possesses equivalent:

  • AI literacy;
  • digital access;
  • confidence;
  • language ability;
  • or technical resources.

Good governance should seek to prevent technology from becoming an additional prerequisite for effective participation.

Do Not Penalise Legitimate AI Assistance

The opposite problem also arises.

A person who uses AI to communicate more effectively should not automatically be treated as:

  • deceptive;
  • excessively sophisticated;
  • inauthentic;
  • difficult;
  • unreasonable;
  • or less entitled to participate.

The substantive questions remain:

What issue is being raised?

What evidence supports it?

What response or action is required?

AI assistance does not remove the need for verification.

But neither does it justify dismissing otherwise legitimate engagement.

Written Communication

For some people, written communication may be substantially more accessible than:

  • telephone contact;
  • live meetings;
  • rapid verbal exchange;
  • or unstructured conversation.

Public-sector systems should therefore consider whether AI-enabled workflows preserve meaningful written routes.

An institution should not assume that:

more immediate communication

is necessarily

more accessible communication.

Accessibility depends upon the person and the context.

Cognitive Load

AI systems may reduce cognitive burden by:

  • simplifying language;
  • grouping information;
  • explaining procedures;
  • identifying next steps;
  • or helping users navigate complex documents.

They can also increase cognitive burden where they:

  • generate excessive information;
  • repeatedly redirect users;
  • require complex prompting;
  • produce inconsistent answers;
  • conceal escalation routes;
  • or make people repeat information already supplied.

Accessibility review should therefore consider the whole interaction, not simply whether an automated interface exists.

Human Escalation

Where accessibility, disability, safeguarding or complex individual circumstances materially affect an interaction, people should be able to obtain meaningful human review.

A useful human escalation pathway should answer:

  • Who can I contact?
  • How do I reach them?
  • What information should I provide?
  • Can they see the previous interaction?
  • Do they have authority to act?
  • Can they correct an automated interpretation?
  • Can they reconsider the matter?

A nominal contact route that cannot meaningfully change the outcome may not provide effective human escalation.

Accessibility and Triage

AI-assisted triage may unintentionally disadvantage people whose communication does not match expected patterns.

For example, urgency may be under-recognised where:

  • language is calm;
  • a person communicates indirectly;
  • important information appears deep within a long document;
  • unusual terminology is used;
  • communication is highly structured;
  • or a person has difficulty expressing urgency.

Conversely, certain words or communication patterns may produce excessive escalation.

Accessibility should therefore be considered during triage design and testing.

Communication Style Is Not Risk

Institutional systems should be cautious about inferring matters such as:

  • credibility;
  • emotional state;
  • motivation;
  • intent;
  • risk;
  • cooperation;
  • or capacity

from communication style alone.

Language may be influenced by:

  • disability;
  • education;
  • culture;
  • stress;
  • translation;
  • AI assistance;
  • neurodevelopment;
  • literacy;
  • professional support;
  • or personal communication preference.

A communication style is not, by itself, a reliable proxy for a consequential human characteristic.

AI-Assisted Correspondence

Where people use AI to prepare complaints, requests or representations, public bodies should distinguish between:

technology used to assist communication

and

the substantive content requiring administrative review.

Relevant questions include:

  • What issue is being raised?
  • Has it been substantively answered?
  • What remains unresolved?
  • What factual claims require verification?
  • Is the person seeking clarification?
  • Does the communication indicate an accessibility need?
  • Is the institutional response proportionate?

The tool used to draft the communication should not silently replace analysis of its substance.

Accessible Explanation

Where AI materially contributes to a consequential public-sector process, affected people should be able to understand enough about the process to participate meaningfully.

That may require explaining:

  • what information was considered;
  • what role AI played;
  • whether the system produced a recommendation or decision;
  • who made the final determination;
  • what uncertainty existed;
  • how inaccurate information can be corrected;
  • and how the outcome can be challenged or reconsidered.

Technical complexity should not become administrative opacity.

Correction Pathways

Accessible systems should make factual correction practical.

A person should not need to understand the architecture of an AI system before they can say:

This information is wrong.

A correction pathway should make clear:

  • where the correction should be sent;
  • what evidence may be required;
  • who reviews it;
  • whether the underlying record will be amended;
  • whether downstream records will be checked;
  • and whether the corrected information can affect the outcome.

Accessibility includes the ability to correct what the institution believes about you.

Alternative Channels

Responsible AI deployment should preserve alternative routes where appropriate.

These may include:

  • email;
  • written correspondence;
  • telephone;
  • in-person assistance;
  • accessible digital forms;
  • representative or advocate support;
  • human case review.

Not every service needs every channel.

But organisations should understand the consequences of removing an existing route.

Automation should not unintentionally convert administrative convenience into exclusion.

Testing Accessibility

Public bodies should test AI-assisted services with real-world diversity in mind.

Testing may consider:

  • long communications;
  • short communications;
  • indirect language;
  • second-language users;
  • accessibility-assisted writing;
  • people with low digital literacy;
  • different reading levels;
  • conflicting information;
  • incomplete information;
  • unusual communication styles;
  • users seeking human escalation.

Technical performance under ideal conditions is not enough.

The question is whether the system remains usable when reality becomes complicated.

Questions for Public Bodies

Where AI interacts with public participation, organisations may ask:

  1. Does AI improve or reduce accessibility?
  2. Can people participate without using AI?
  3. Are meaningful alternative channels preserved?
  4. Can AI-assisted correspondence be assessed on its substance?
  5. Are staff trained not to equate polished AI-assisted writing with unreliability?
  6. Could triage disadvantage unusual communication styles?
  7. Can people obtain meaningful human review?
  8. Can inaccurate information be corrected easily?
  9. Are consequential outputs explained clearly?
  10. Are accessibility needs considered when systems are designed?
  11. Can someone challenge an automated interpretation?
  12. Does the technology expand participation or restrict it?

SIAAF Relevance

This Guidance Note principally relates to:

Domain 04 — Communication & Feedback Integrity

Does important information move through the institution accurately and effectively?

Domain 05 — Escalation, Challenge & Contestability

Can people obtain meaningful human review and challenge inaccurate institutional interpretations?

Domain 07 — AI Literacy & Organisational Readiness

Do staff understand how AI can function both as an institutional tool and as an accessibility aid?

It may also engage:

Domain 02 — Decision Integrity & Human Oversight

Where automated systems contribute to consequential administrative outcomes.

Domain 06 — Risk, Harm & Operational Resilience

Where inaccessible design may affect access to essential services or significant rights and interests.

SWANK AI Standard

ACCESSIBILITY SHOULD SURVIVE AUTOMATION

AI should:

expand participation

preserve meaningful alternatives

support legitimate assistance

permit human clarification

retain correction pathways

and

avoid making technology a new condition of access.

Where automation increases complexity, a meaningful human route should remain available.

Related SWANK AI Guidance

Guidance Note 01 — Handling AI-Assisted Complaints and Correspondence

Guidance Note 04 — AI Literacy and Academic Integrity

Guidance Note 06 — AI Detection: Indicators Are Not Proof

Guidance Note 12 — AI and Administrative Fairness

Guidance Note 16 — AI Triage and Prioritisation

Guidance Note 19 — Correction Rights in AI-Assisted Systems

SWANK AI

Independent AI & Institutional Assurance

We do not just review AI. We review the institutional systems responsible for governing it.

Scroll to Top