Accessibility Standards for Indic AI Interfaces

A department launched a Hindi agent on the portal homepage. The widget opened only on click, stole focus, and could not be closed from the keyboard. The Hindi text used a webfont the popular screen-reader voice skipped on conjuncts. The accessibility statement on the site still claimed GIGW alignment from a 2022 static-page audit.

The agent was new. The duty was not.

This guide maps GIGW 3.0, the Rights of Persons with Disabilities Act, 2016, WCAG 2.1, and IS 17802 onto interfaces where some of the text is generated. It is field guidance, not a certificate and not legal advice.

The duties do not pause for a generated sentence

The RPwD Act requires accessibility of information and communication in the terms the statute and its rules set. GIGW 3.0 is the government web and app instruction set — quality, accessibility aligned with WCAG 2.1, cybersecurity, lifecycle — including apps and multilingual publishing. IS 17802 is the ICT accessibility standard you can put in a tender.

Generated text is still text. Generated controls are still controls. A model that writes a low-contrast caption, or a widget that has no name for a button, is an accessibility defect. 'The model did it' is not a defence the audit party will enjoy.

Where Indic AI interfaces specifically break

Fonts that look fine and do not speak. Line-heights that clip matras so a screen-magnifier user loses a vowel and changes a word. Language tags that say en on a Hindi reply, so the wrong voice speaks. Mix that the screen reader language-switches into nonsense. A voice agent with no transcript for a deaf user. A transcript with no audio for a user who cannot read.

Captchas in English in front of a Hindi agent. OTP flows that are only visual. Timeouts in a chat that a keyboard user cannot extend. Images of circulars without an alternative, offered as the only copy of a decision.

Tests to add when the interface is an Indic agent.

Test	How	Fail if
Keyboard	Open, type, read history, close, without a mouse	Focus trap or no close
Screen reader + Indic font	NVDA / TalkBack / VoiceOver on your real font	Skipped conjuncts on names or amounts
Language tag	Inspect the live reply	en spoken over Hindi, or the reverse
Contrast and clip	Zoom and high contrast on matras	Clipped vowel changes a word
Equivalent path	Same task on web, voice, human	A benefit exists only inside an inaccessible widget

Rules for generated text

Lock language tags to the language of the reply. Do not emit images of text as the only form of an answer. Keep headings and lists as structure, not as ASCII art. Do not use colour alone to mark a rejected application. Offer a plain-language alternative when the authentic text is a statute — labelled as alternative.

When the model is unsure, the accessible move is a short, structured clarification, not a long paragraph of hedging that a screen-reader user must hear twice.

What to put in the buy

Require a GIGW-and-WCAG test on the integrated property, not on a vendor's marketing site. Require the Indic screen-reader test on your fonts. Require a keyboard path. Require that voice, if present, has a text equivalent and the reverse. Do not accept a 2022 static audit as evidence for a 2026 widget.

If the agent speaks a language the rest of the portal does not, the skip-link, the language toggle and the human path still have to exist in that language. An English-only 'contact us' under a Hindi widget is a second door that the person who needed Hindi cannot open.

Objections you will hear — and what to do with them

These are the lines that stall the file. Answer them in the room, then put the answer in the note. A spoken answer without paper will be forgotten by the next officer.

We will fix accessibility after the language model is tuned.

The widget's name, focus and contrast do not depend on the checkpoint. Do them first. Generated-text rules can follow, but the shell cannot wait.

Screen-reader users are a small percentage.

The RPwD Act is not a traffic threshold. Neither is GIGW. Build the path.

Indic screen readers are the weak link, not us.

Then choose fonts and structure that the existing readers can speak, and offer voice plus human. Blaming the reader is not a file.

A third-party chatbot SaaS is outside our GIGW boundary.

If it is on the government property and it is how a citizen completes a task, it is in the boundary. If it cannot be made to pass, do not embed it.

Generated is still published

Officers sometimes treat a chat reply as ephemeral and therefore outside GIGW. Citizens screenshot it. Journalists quote it. An inquiry will ask why the only copy of a rejection was an inaccessible bubble. Treat every generated sentence that leaves the widget as published text.

That is why language tags, contrast, and an equivalent path are not polish. They are how a published sentence remains a public service.

Two weeks to an accessible Indic shell

Bring one keyboard-only officer and one screen-reader user, paid for their time, not a volunteer from IT.

1. Day 1: read the current GIGW statement. Date it. If it predates the widget, mark it stale.
2. Day 2–3: keyboard the entire agent path. Log traps.
3. Day 4–5: screen-reader pass on Hindi or the year-one language, real font, names and amounts.
4. Day 6: fix language tags on generated replies.
5. Day 7: contrast and matra clipping at zoom.
6. Day 8: confirm an equivalent non-widget path for the same task.
7. Day 9: write generated-text rules into the system prompt and the QA checklist.
8. Day 10: put the integrated test into the tender or the go-live gate.
9. Day 11–12: retest. File the new statement with a date.

How this shows up in the file

The note should say: the agent is part of the GIGW property. RPwD duties apply. We tested keyboard, Indic screen reader on our fonts, language tags, and an equivalent path. A prior static audit does not cover this widget.

Attach the test log and the dated statement.

This article is informational field guidance for Indian public institutions, not legal, procurement, security-accreditation, linguistics or engineering advice. Confirm against the current Gazette, Official Languages Act and Rules, state official-language law, GIGW, RPwD Act, DPDP text and Rules, CERT-In directions, departmental manual and your counsel before you file it.

How to test this with real speech, not staff English

“Accessibility Standards for Indic AI Interfaces” fails in the field if you only tested officers. A P4 Programme/Implementation should hear a first-generation student, a rural caller, or a Hinglish grievance before claiming “accessibility Indic interface”.

An Indic agent that a screen reader cannot speak, or that traps a keyboard in a chat widget, fails GIGW and the RPwD Act no matter how many languages the deck lists. Twenty-two scheduled languages is a Constitution fact, not a model fact. Script support is not language support. Official language rules may require bilingual output even when the model prefers one script.

• Name the languages and scripts in the eval set.
• Include code-mix and scheme-name tests.
• Measure comprehension, not BLEU alone.
• Design a human fallback when language fails.

Close this loop before the next CAB

Put “Accessibility Standards for Indic AI Interfaces” on the next change-advisory or bid-opening agenda as a single line item with an owner. If it cannot earn a line item, it will not earn a control. The owner should be a P4 Programme/Implementation, not “the vendor.”

Revisit the item when the model, the GeM term, the region, or the SI changes. “accessibility Indic interface” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.

What must be true before you file this

If “Accessibility Standards for Indic AI Interfaces” is only a heading, it will not survive a file inspection. A P4 Programme/Implementation should be able to attach one artefact that proves “accessibility Indic interface”: a log export, a clause, a scored row, a dated notice, or a refusal rule.

Write three dated sentences: what was decided, who owns it, and when it will be re-checked. Unsigned sentences are souvenirs. Dated sentences are controls.

• Name the owner of “accessibility Indic interface” inside the institution.
• Attach one artefact a stranger can open next year.
• Revisit when the model, the notice, or the SI changes.
• Do not treat a vendor slide as evidence.