Authoring Tool Accessibility Guidelines (ATAG) 2.0 is a 2015 World Wide Web Consortium (W3C) Recommendation which “provides guidelines for designing web content authoring tools that are both more accessible to authors with disabilities … and designed to enable, support, and promote the production of more accessible web content by all authors” (World Wide Web Consortion, 2015). Since the publication of the ATAG 2.0 Recommendation, the advent of Generative Artificial Intelligence (AI) and its integration into authoring environments has significantly altered the experience and process of authoring web content.
In this research report, we aim to assess the impact and opportunities of these technologies for the existing principles and guidelines of ATAG 2.0, and provide recommendations
As Generative AI is increasingly integrated into authoring tools, new accessibility issues may be introduced by the Generative AI tools themselves (Authoring Tool Accessibility Guidelines CG, 2026a; 2026b; World Wide Web Consortion, 2015). A blind developer identified multiple issues with the web-based interface to Anthropic’s Claude Generative AI tool which prevented them from using screen readers to interact with the tool (Ok-Top-3337, 2025). Blind and low-vision developers also identified issues with the category of Generative AI coding assistants, within the realms of content switching, control and predictability, and cognitive overload caused by the volume of output (Flores-Saviaga et al., 2025). As “current interfaces rely mostly on visual-only indicators… real-time status updates, clear explanations of outputs, and accessible methods for users to verify, understand, and act upon AI-generated results” will be critical to the accessibility of Generative AI-based authoring tools (Chen et al., 2025). One specific accessibility barrier that blind and low-vision users identified with Generative AI tools is keyboard accessibility, specifically scenarios where “shortcuts conflicted with their established keyboard habits” such as “up and down arrow keys in the input box—rather than moving between lines in a multiline-prompt, these keys cycled through previous commands, often without users realizing it” (Chen et al., 2025). The use of “consistent and predictable keyboard shortcuts” can help to “[address] participants’ difficulties with unintended actions that disrupt their workflow” (Chen et al., 2025).
In addition to baseline accessibility of authoring tool interfaces following WCAG (World Wide Web Consortium, 2024) principles, the unique characteristics of natural language interfaces underscore the need to “[position] accessibility… as a fundamental requirement in their design and usage” when Generative AI interfaces are part of an authoring tool (Abu Doush & Kassem, 2025). The W3C’s Natural Language Interface Accessibility User Requirements identify several “cognitively taxing” aspects of natural language interfaces, highlighting the need to quickly call “[speech] input commands… to mind”, to closely observe results “to make sure that the computer has not made a wording mistake”, and to use “the language centre of the brain… [for both] speech input… [and] many of the thought processes that go into things people do on computers such as writing or coding” (World Wide Web Consortium, 2022). As natural language interfaces are integrated into authoring tools, adoption of the Natural Language Interface Accessibility User Requirements will be critical to ensure that accessibility is not compromised.
Other research identifies “metacognitive” demands of Generative AI interfaces, such as the “open-endedness of many current prompting interfaces [which] requires users to have self-awareness of their specific task goals, and be able to decompose their tasks into smaller sub-tasks so as to verbalize these as effective prompts” (Tankelevitch et al., 2024). Additionally, authoring tools which integrate natural language interfaces must support multimodal input and output mechanisms to support a wide range of user needs, and should also ensure that input and output mechanisms can be combined according to user requirements, so that, for example, “[a] user who is deaf or hard of hearing… [can] provide speech input to an application, while having the output presented as text” (World Wide Web Consortium, 2022).
If Generative AI-based authoring tools work with non-text content, it’s essential that they programmatically associate user-supplied text alternatives with that content in the editing view (Authoring Tool Accessibility Guidelines CG, 2026a). It may also be possible for Generative AI-based authoring tools to generate text alternatives for content where the author has not provided them (Oswal & Oswal, 2025).
Another factor to consider is when Generative AI is used to create non-text content such as pictures or videos. For blind or low vision authors, they may be able to provide and/or read the generated text alternatives; however, they may not be able to perceive the actual generated content. A suggestion for this is to have AI act as an intermediary that can describe the generated content to the author .
Generative AI tools can potentially be integrated into authoring tools in such a way as to make the editing views more operable (Authoring Tool Accessibility Guidelines CG, 2026a). For example, section A.3.4 of ATAG suggests that the content structure should be adapted to enhance navigation, for example, to facilitate navigation by structure (World Wide Web Consortion, 2015). This sort of functionality must currently be built into the authoring tool, but a Generative AI agent with access to the content could create and update a table of contents which would facilitate such operability.
Generative AI-enabled authoring tools should support the ability of authors to “personalize the interface according to their needs” (Chen et al., 2025) (see points under Principle A.1 about multi-modal interface options, keyboard customization etc.).
The use of “streamlined and predictable interaction patterns” will be essential in the design of Generative AI-assisted authoring tools (Chen et al., 2025).
Providing “accessible change tracking, including textual or audio cues for the scope (e.g., number of edits), context (e.g., location and status of code blocks), and content (e.g., what was modified)” (Chen et al., 2025) will be critical to ensure that Generative AI-based authoring tools conform to the existing A.4.1 guidelines related to avoiding and correcting mistakes during the authoring process (Authoring Tool Accessibility Guidelines CG, 2026a).
When website builders provide accessibility features and documentation, they aren’t always easily discoverable. As website builders are aimed at novice users, they should actively promote their accessibility features and provide guidance on how to use them through “personalized and accessible learning pathways” (Chen et al., 2025). Furthermore, the user should be suggested ways to improve the accessibility of their work during their authoring sessions in a manner that is obvious and relevant to the content that they are currently authoring (Pillai et al., 2022).
Documentation of authoring tools which incorporate Generative AI will need to include detailed guidance on how to use those tools effectively to produce accessible content. Research has found “several differences in WCAG compliance between accessibility-oriented and accessibility-agnostic prompts, highlighting the importance of explicit accessibility guidance” (Gurita & Vatavu, 2025).
Generative AI tools such as ChatGPT can generate websites; however, they do not tend to produce accessible websites by default nor perfectly compliant sites even with explicit accessibility guidance (Chen et al., 2025), (World Wide Web Consortion, 2015), (Authoring Tool Accessibility Guidelines CG, 2025). This is likely due in part to lack of training on compliant code; as simpler and more common representations of accessible interfaces (e.g. checkboxes, tables) are more often generated correctly5. Some of the accessibility issues that may be produced include lack of contrast, improper labelling or text information, even non-unique IDs. While the model can be instructed to modify or correct its output, due to the limitations in training, it often can still not completely address all issues (Palmer & Oswal, 2025). To improve Generative AI output, a foundation prompt or context can be provided to include more detailed instructions about producing accessible code and markup (Authoring Tool Accessibility Guidelines CG, 2025). However, even with this in place the Generative AI may still require follow up prompts to correct missing, or incorrectly generated output (Doush & Kassem, 2025). It is also important to note that the Generative AI results tend to decline as requests become more complex or ambiguous (Gurita & Vatavu, 2025).
With Generative AI models, we often do not know what data the models were trained on, nor how they may have been finetuned in the applications they are embedded in. For authoring tools with integrated Generative AI, it may not be known if there are already foundation prompts/context or templates that are being used which may affect the accessibility of the generated content (Gurita & Vatavu, 2025).
A deeper look into the failure of Generative AI to consistently produce accessible output suggests this occurs because semantic HTML and ARIA are underrepresented in the training data, AI output is predominantly assessed visually rather than on semantics, it takes fewer tokens to write inaccessible markup, and AI models don’t understand the accessibility tree. The following steps can be taken to mitigate these issues (Pillai et al., 2022):
Rather than allowing the Generative AI to produce any output, have it produce output making use of of existing libraries that already take accessibility into consideration
Perform automated and manual testing of the output
Perform manual testing; possibly comparing to examplars (Gurita & Vatavu, 2025)
However, it’s important to remember that compliant doesn’t necessarily mean accessible. Even if a Generative AI tool is able to produce WCAG compliant output, it may not be accessible to all users (Authoring Tool Accessibility Guidelines CG, 2026c). Evaluation should not just rely on compliance but on their ability to produce meaningful accessible outcomes (Chen et al., 2025; Gurita & Vatavu, 2025). Human expertise is a critical component for identifying and remediating accessibility issues (Palmer & Oswal, 2025).
Some research has explored attempting to generate intentionally inaccessible interfaces through Generative AI. While the Generative AI-based tools would refuse, the output often still contained similar accessibility issues as attempts to create accessible interfaces. It appears that the Generative AI-base tools are locked into following patterns for their output, even if guardrails are in place to limit the prompts to them. This has further implications to both creative expression and inclusive design in general. If existing design patterns are so strictly adhered to, it will limit new designs from emerging; not to mention a lock in to a specific definition of what good design is; which may not be applicable in all contexts (Palmer & Oswal, 2025).
Another approach that Generative AI can strive for is to produce semantically accessible content, rather than trying to replicate the visual surface of compliant interfaces (Authoring Tool Accessibility Guidelines CG, 2026d). This would require the Generative AI to be capable of understanding the relational, labelling and navigational principles that make an interface usable and comprehensible for people with disabilities (Patel & Melcer, 2025).
Section 11.8.3 of CAN/ASC - EN 301 549:2024 Accessibility requirements for ICT products and services (EN 301 549:2021, IDT) specification requires that “[if] the authoring tool provides restructuring transformations or re-coding transformations, then accessibility information shall be preserved in the output if equivalent mechanisms exist in the content technology of the output” (Accessibility Standards Canada, 2024).
Even the most popular website builders, such as Wix and Squarespace, will default to using the file name or image caption as an image’s alt text (Oswal & Oswal, 2025).
Website builders are marketed to support fast and easy, often with no code, solutions for creating a website. However, they may provide little to no support for facilitating accessibility features/requirements. For example, limited or no ability to set alt text on images or properly style headings (Palmer & Oswal, 2025).
Generic and irrelevant strings may pass automated accessibility checks, but fail to provide the required semantic meaning to users. An approach to flag these cases involves using Generative AI-based tools to assess the semantic relevance (Authoring Tool Accessibility Guidelines CG, 2026e). This process should capture most generic strings. However, it is more complex to make these determinations when an understanding of the surrounding context is required (Calò et al., 2026).
For authors without knowledge and experience with web accessibility, using Generative AI-based coding tools often does not provide any improvements in the accessibility of the created content. Authors will often focus on increased productivity and may not be aware of necessary accessibility consideration. Even in cases where the AI tool may suggest accessibility related improvements, the authors may still overlook these if they do not understand the requirement or purpose. A solution would tackle both aspects of improving Generative AI models to produce accessible content by default, and to guide/teach authors about accessibility practices and principles. One option for the latter would be a more guided approach at authoring with an interactive process for implementing accessible content. In this way the author is taught about accessibility as they author (Mowar et al., 2024). When Generative AI-based tools are used, they should “proactively seek clarification by asking [authors] to specify their intent” to support the author’s learning process (Chen et al., 2025). In addition to this, accessibility wizards could be included into the tool to provide a review stage where the generated content is audited with opportunities to explain access issues to the author and provide an interactive approach for remediating the content (Authoring Tool Accessibility Guidelines CG, 2025).
Research with student developers found that they often didn’t fully trust the AI coding assistant and felt the need to review suggestions. There was a feeling that the suggestions could need refinement or could even be completely wrong. Trust increases when the suggestions are accompanied by “references, explanations, or evidence” based on credible sources and guidelines (e.g. WCAG) (Chen et al., 2025). In addition to providing a reference this is also an opportunity to educate the author about accessibility principles and practices. However, the context is still important and currently AI coding tools do not always take into consideration the design nuances or user experience. They also tend to imitate or follow the author’s existing code over following best practices for accessible implementations (Patel & Melcer, 2025).
Generative AI coding assistants shift the author’s workflow from generating content to evaluating it. An author’s self-confidence and ability in evaluating generated content is dependent on their level of domain and AI knowledge/expertise, the Generative AI model’s consistency and quality of output, and even the amount to be evaluated or reviewed. For novice authors knowing how and what to evaluate, as well as how to remediate the issues can be a challenge. This is exasperated by large quantities of content output and even large or complex suggestions from the AI model. The non-deterministic nature of Generative AI can introduce unexpected or erroneous output that may not be easy to detect, even for experienced authors. This problem can be compounded with multiple iterations through Generative AI if the output does not align with previous content. In addition to developing more domain knowledge, even experienced authors will need to develop new strategies for working with AI generated content (Aljedaani et al., 2024; Tankelevitch et al., 2024).
Built-in WCAG checking should run automatically against each generated output, flagging specific violations by criterion number, severity, and the elements that caused them as surfacing the problem to authors often cannot recognize it on their own. Where the tool would otherwise produce non-compliant output, it should warn the designer or refuse, rather than silently generating an interface that looks plausible but fails on accessibility; and where corrective iteration is needed, the interface should support structured component-level refinement rather than forcing authors back into an ad-hoc chat loop and hoping the Generative AI model focuses on the right place (Gurita & Vatavu, 2025).
Section 11.8.2 of CAN/ASC - EN 301 549:2024 Accessibility requirements for ICT products and services (EN 301 549:2021, IDT) specification requires that authoring tools “enable and guide” the production of accessible web content (Accessibility Standards Canada, 2024).
Section 11.8.5 of CAN/ASC - EN 301 549:2024 Accessibility requirements for ICT products and services (EN 301 549:2021, IDT) specification requires that tools which provide templates provide accessible templates (Accessibility Standards Canada, 2024).
Beyond authoring code Generative AI tools can be instructed to remediate existing code. Research has found that instructing a well-trained Generative AI model of the necessary remediations required with instruction based on WCAG principle can fix issues. However, as of yet it isn’t 100% capable which may be related to complexities of the accessibility issues, understanding of the code to remediate, effectiveness of the prompts, and/or training (Aljedaani et al., 2024). While Generative AI may not be able to handle complex remediation, it could be used to handle more routine accessibility issues. Thereby, freeing up time for the authors to focus on the more complex issues manually (Suh et al., n.d.).
Section 11.8.4 of CAN/ASC - EN 301 549:2024 Accessibility requirements for ICT products and services (EN 301 549:2021, IDT) specification requires that the accessibility checking functionality of tools that can detect when accessibility requirements are not met, to also provide suggestions on how to fix them. They could also automate the fixing process where applicable (Accessibility Standards Canada, 2024).
For Generative AI-based authoring tools/environments, users may have varying degrees of experience and may come from different backgrounds (e.g. developer vs designer). The tools need to support interaction patterns with the user, providing real time accessibility feedback and guidance through complex accessibility/design decisions (Gurita, 2025). The tools should also be able to transparently explain its reasoning behind choices (Gurita & Vatavu, 2025). In addition to informing the user about potential requirements that are being followed, they should encourage the user to reflect on the impact of modifications/refinement.