CMU AI Remediation Playbook: Safe Accessibility With Human Review

Carnegie Mellon University’s Digital Accessibility Office has issued pragmatic guidance titled “AI as a Remediation Assistant,” urging campus teams to use AI responsibly to reduce barriers and expand participation while maintaining human oversight and conformance to CMU’s Digital Accessibility Policy and WCAG. The guidance recognizes AI as a speed multiplier for common remediation tasks—drafting alt text, generating suggested HTML fixes, producing transcripts and captions, and simplifying dense text—but it is emphatic that automated outputs must be reviewed, integrated carefully, and combined with manual testing to ensure true accessibility and legal compliance. The memo lists practical tools available to CMU users (Microsoft Copilot, Google Gemini, ChatGPT Edu, Siteimprove AI, Grackle Workspace, Zoom AI Companion, NotebookLM) and highlights frequent, high-impact failures (low-contrast text, missing form labels, empty links/buttons, and missing alt text) that remain dominant on the modern web. This feature unpacks that guidance, verifies key claims against public reporting and vendor documentation, and translates the recommendations into an operational remediation playbook for IT leaders, content creators, and accessibility practitioners.

Background / Overview​

Carnegie Mellon’s guidance frames AI as a remediation assistant rather than a replacement for accessibility expertise. The memo emphasizes three core principles:

• Use AI to augment human effort (drafting, batch suggestions, alternate-format generation).
• Keep a human-in-the-loop for verification, context, and policy compliance.
• Prefer building accessibility into code and content rather than relying on overlay or post-hoc band-aids.

Those priorities align with sector-level best practice: start with content creators and code, instrument reviewer workflows, and restrict claims that AI can achieve compliance on its own. Peer institutional guidance and technology governance reporting reinforce the same posture: mandate training, require provenance and auditing for AI outputs, and gate tool access behind role-based controls where sensitive data is involved.

---

What CMU Recommends (short summary)​

• Treat AI as a tool for remediation and format generation, not a compliance guarantee.
• Use institutionally provisioned tools where possible (Copilot, Gemini, ChatGPT Edu, Siteimprove AI, Grackle) and consult the Digital Accessibility Office for questions and tool access.
• Focus remediation effort on high-impact, common errors first: alt text, contrast, form labels, empty links/buttons, and reading order / PDF tags.
• Always review and test AI suggestions (visual inspection, keyboard-only navigation, screen reader testing).
• Produce multiple formats (captions + transcript + slide download) for multimedia to maximize access.

These points are practical and consistent with current accessibility tooling and vendor features; where CMU cites aggregated web accessibility data, those figures are verifiable against independent reports (see next section).

---

Verifying the Big Numbers: WebAIM and the State of the Web​

CMU references WebAIM’s "1 Million" research to show where effort should be prioritized. The headline findings align with the publicly available WebAIM Million analysis: roughly 95% of homepages still had at least one automatically detectable WCAG 2 failure in the 1M sample, and issues like low contrast, missing alt text, empty links, and missing form input labels remain the most common single-error categories. WebAIM’s dataset and summary statistics confirm that image alt-text failures and contrast failures are extremely common and account for the majority of automatically detectable errors. Independent summaries and industry write-ups corroborate the trend (average errors-per-page and the persistence of a small set of frequent errors), underscoring the practical value of prioritizing a short remediation list. These independent analyses reinforce that focusing on alt text, contrast, form names, and empty interactive elements will yield high returns. Practical implication: prioritize the handful of fixes that eliminate the majority of automated failures before attempting costlier, lower-yield tasks.

---

Where AI Helps — and How to Use It​

Alt text and image descriptions​

AI can rapidly draft contextual image descriptions, speeding bulk remediation and serving as a starting point for human edits. CMU gives useful examples showing how different tools phrase alt text differently (ChatGPT, Microsoft Office, Gemini) and stresses that alt text must be contextual—its content depends on the image’s purpose (decorative vs informative vs link target) and the document’s audience.
Best practices:

• Use AI to generate suggested alt text, then edit for context, concision, and privacy.
• Prefer human review when images contain people, sensitive information, or complex data (graphs, charts).
• For linked images, ensure the alt text conveys the link purpose (not just the image content).

This approach mirrors vendor guidance and platform capabilities: Microsoft’s image description features and other LLMs produce drafts that are useful but not infallible. Rely on review workflows and provenance logging when alt text is edited from AI output.

HTML Remediation (forms, links, buttons)​

Tools such as Siteimprove’s AI Remediate can generate code suggestions and help identify missing ARIA roles, unlabeled form inputs, and incorrect semantic structure. Siteimprove documentation describes a workflow where a remediation suggestion is offered in the Page Report, developers test the suggestion in a staging copy, and then deploy once the change passes re-testing. Use AI code suggestions as a developer assist, not an automatic commit path. Practical steps:

• Run automated scanner (Lighthouse/WAVE/Siteimprove).
• Export occurrences and let AI generate candidate fixes.
• Implement in a development branch and run regression/automated tests.
• Manually test with keyboard navigation and a screen reader before production rollout.

Transcripts and captions​

Zoom’s AI Companion and built-in Zoom live transcription provide real-time captions and meeting transcripts; participants can adjust subtitle font size and view a full transcript pane. CMU’s guidance to use live captions and review transcripts is consistent with Zoom’s features — captions are helpful but must be audited for accuracy and corrected before reuse as official records. Key practice: if transcripts are used as accessibility artifacts (e.g., for a public lecture), edit auto-generated transcripts before publishing and attach a corrected caption file (SRT/VTT) where possible.

Simplifying text and alternative formats​

LLMs are excellent at generating simplified versions of dense material, outlines, or study guides. NotebookLM / ChatGPT Edu can rephrase content into plain language or create structured outlines, saving editorial time while improving comprehension for neurodiverse and English-language-learner audiences. Always include a review step to confirm accuracy and to avoid introducing hallucinated claims.

Document remediation (Office, Google Workspace, PDFs)​

• Microsoft Office includes a built-in Accessibility Checker that flags color contrast, heading structure, alt text, and reading order problems; it’s a good first step for Word and PowerPoint remediation.
• For Google Docs/Slides, Grackle Workspace is a widely used add-on that automates checks and can export tagged PDFs or accessible HTML, with guided fixes and table-tagging wizards. Several universities and vendors document Grackle’s capabilities for producing Tagged PDF output and helping with reading order changes. Use Grackle as a productivity booster when your environment is Google-first.
• For legacy PDFs, automated autotagging helps but rarely suffices: Adobe Acrobat and institutional accessibility teams advise that autotagging is a first pass and that reading order, heading structure, alt text, and table tagging typically require manual correction. Plan for manual review after any autotagging pass.

---

Tool-by-tool practical audit: what the docs actually say​

• Siteimprove AI (AI Remediate): provides code suggestions from the Page Report and a conversational assistant for follow-up questions; intended as developer-facing remediation guidance that should be tested in staging before deployment.
• Grackle Workspace: integrates into Google Docs/Sheets/Slides, performs automated checks, offers guided fixes, and exports to Tagged PDF/accessible HTML. Good for high-volume document workflows in Google Workspace.
• Microsoft Office Accessibility Checker: built into Word/PowerPoint and useful for quick, author-facing checks (contrast, alt text, heading structure). It’s not a one-click cure but reduces entry barriers for authors in the Microsoft ecosystem.
• Zoom AI Companion / Zoom Live Transcript: provides live transcription, translation into many languages, and subtitle font sizing controls; transcripts are helpful but should be edited for accuracy if published.
• Large language models (ChatGPT Edu, Gemini, Microsoft Copilot, ChatGPT Enterprise): useful for drafting alternative text, plain-language summaries, and code snippets. But they can hallucinate, omit context, or misinterpret the content’s intended function—necessitating human verification.

---

Known limitations & risks (explicitly verified)​

• PDF Autotagging is imperfect. Automated tagging frequently misorders content and misidentifies decorative elements, requiring manual remediation for reading order, alt text, and table semantics. Institutional remediation guides advise autotagging only as a first pass followed by manual tag tree correction.
• Accessibility overlays (widgets) are unreliable and risky. Multiple accessibility experts and community groups warn that overlays do not produce standards-compliant, durable code fixes and may interfere with assistive technologies; overlays have even attracted legal scrutiny and regulatory action in some cases. The recommendation is to invest in fixing source code and content rather than relying on overlays.
• Hallucinations and content accuracy: generative models can invent facts, misname people or places, or misinterpret the semantic purpose of an element (e.g., treating a navigational icon as a decorative image). Human verification is mandatory for any content that will be published as authoritative. This is a common and accepted limitation across commercial LLM guidance and institutional AI governance documents.
• Data privacy and contractual risk: past procurement reviews and university pilots stress non-training contractual clauses, egress/portability rights, and auditability for vendor models. Do not paste protected or personally identifiable data into consumer models without contractual and technical guarantees. Campus governance and procurement teams should be involved.
• Over-reliance on automated fixes erodes process and accountability. Fixes suggested by AI must be recorded (who accepted/edited/published), and remediation workflows should preserve provenance so that audits and rework are possible.

---

Practical remediation playbook (operationalized)​

Below is a prioritized, sequential remediation plan that teams can implement immediately.

• Triage (1–2 days)
• Run a site- or document-level automated scan (Siteimprove, WAVE, Lighthouse, Grackle).
• Aggregate common, repeatable issues into a short list (alt text, contrast, form labels, empty links/buttons, PDF tagging).
• Estimate effort (hours/per-page, per-document).
• Batch-fix with AI + human review (1–3 weeks)
• For images: generate alt text drafts with an LLM or Copilot, then have an editor verify and approve.
• For HTML issues: run Siteimprove AI Remediate to produce suggested code changes; apply in dev, re-scan, and perform manual accessibility tests before merging to production.
• For Google Docs/Slides-heavy teams: use Grackle Workspace to produce tagged PDF output and fix reading order via the tool’s guided wizards.
• Manual QA and assistive-technology testing (ongoing)
• Keyboard-only navigation checks; basic screen reader pass (NVDA or VoiceOver), and a small set of real-user tests where possible.
• Validate reading order and tab order for PDFs and complex documents.
• Publish alternate formats for multimedia
• For videos or lectures: supply captions (VTT), a corrected transcript, and downloadable slides. Use Zoom live captions during events, then edit the transcript before republishing.
• Operationalize governance and training
• Require short author training on alt text, headings, and the institutional toolset.
• Maintain a remediation log that records the AI outputs, editor changes, and publish approvals (audit trail).
• Preserve procurement and data-use constraints—restrict consumer-grade model use for protected content unless contractually permitted.
• Measurement and continuous improvement
• Track KPIs: % of homepage images with valid alt text, number of contrast failures, mean time to remediate flagged issues, and user feedback.
• Run periodic manual audits and user-testing sessions with assistive-tech users to measure real-world usability improvements.

---

A short checklist for content creators (copyable)​

• Alt text: present for informative images; empty alt (alt="") for purely decorative images.
• Form labels: every input has a visible or programmatic label; associated aria attributes used correctly.
• Contrast: text meets WCAG contrast ratios for normal and large text.
• Links & buttons: no empty link text; link text conveys purpose out-of-context.
• Headings: use semantic headings (h1–h6) to structure content; avoid visual-only headings.
• PDFs: source files exported as tagged PDF where possible; autotag only as a first pass and always review reading order.

---

Governance checklist for IT and procurement​

• Use enterprise/education SKUs (Copilot, ChatGPT Edu, Gemini enterprise features) where non-training and data residency clauses are in place.
• Require human-in-the-loop thresholds and define who must verify AI outputs for different risk classes of content.
• Capture provenance logs: which model, prompt, and edits led to each published remediation.
• Pilot on low-risk content first; measure false-positive/false-negative remediation rates before scaling.

Institutional deployments of AI-driven services should be paired with mandatory training and role-based access, as many campuses and public-sector pilots have recommended.

---

Critical analysis — strengths, trade-offs, and remaining unknowns​

Strengths

• Focused use of AI reduces tedious bulk work (alt text drafts, first-pass autotagging, caption generation).
• Institutional provisioning (Copilot, ChatGPT Edu, Siteimprove, Grackle) reduces the risk of uncontrolled consumer tool usage.
• Prioritizing a short remediation list (contrast, alt, forms, empty interactive elements) promises outsized impact for relatively small effort.

Trade-offs and risks

• Autotagging and overlay reliance are not cures: PDFs often need manual tag-tree editing; overlays can degrade AT experiences and expose sites to legal risk. Invest in source-code fixes.
• Hallucination risk means any factual or contextual content generated by LLMs must be verified; this is especially important for transcripts, alt text that describes people or medical conditions, and any text republished as authoritative.
• Contractual nuance matters: “enterprise” or “edu” plans may state non-training or tenant protections—validate those clauses in procurement and involve legal/privacy teams.

Unverifiable or vendor-claim areas (flagged)

• Any single-vendor claim of “100% compliance via automatic remediation” should be treated with skepticism. Where vendors advertise full compliance from overlays or one-click fixes, empirical evidence and community consensus show those claims are overstated. Institutions should insist on pilot data: false-positive rates, percent of issues requiring manual intervention, and sample audited results.

---

Final recommendations (practical, immediate)​

• Implement a short remediation sprint focused on the top 4 error classes identified by WebAIM and your scanning tools; use AI suggestions to accelerate drafting and developer proposals but require a QA gate before publishing.
• Adopt Grackle Workspace for Google-first document workflows and Siteimprove AI for web remediation where budget and provisioning allow; both provide practical, workflow-integrated assistance.
• Treat PDF autotagging as a staging step—always schedule a manual tag-review pass and prioritize fixing source documents to export accessible PDFs when possible.
• Reject reliance on overlays as a compliance strategy; allocate funding to actual code/content remediation and user testing.
• Build an author QA checklist, a human-in-the-loop verification step, and an audit log of AI-sourced edits to maintain accountability and traceability.

---

Carnegie Mellon’s guidance is pragmatic: AI can reduce friction and expand reach if used as an assistant rather than a shortcut. The most effective programs pair easy-to-use tooling (author checkers, Siteimprove/Grackle automations, Zoom transcription) with minimal-but-rigorous human review, a short prioritized remediation backlog, and institutional governance to manage privacy and procurement risk. That combination—tooling plus human expertise and clear governance—delivers accessibility improvements at scale with manageable risk.
Conclusion: use AI to accelerate remediation, not to bypass rigor. Fix the few things that matter to most users first, require verification, and treat automated suggestions as drafts—not authoritative output.

---

Source: Carnegie Mellon University AI for Accessibility Remediation - Computing Services - Office of the CIO - Carnegie Mellon University