Is Automated Accessibility Remediation Reliable?

Automated remediation is reliable only within the scope demonstrated by its checks and evaluation. There is no published Aelira benchmark establishing that 70–80% of issues can be resolved without human intervention.

Separate mechanical changes from interpretation

Adding a known language value or title differs from inferring the language or a meaningful title. Building bookmarks from an approved heading tree differs from deciding the intended heading hierarchy. Reading order and table semantics can require context even when a change looks structural.

Generated descriptions and recognized mathematics must be compared with the source and teaching purpose. A confidence value is not a calibrated accuracy percentage unless a relevant evaluation establishes that relationship. An item without a warning can still be wrong.

Inspect the actual deliverable

Check whether the tool returned a saved candidate, applied only partial changes, refused output or left manual work. Open the downloaded file and compare content, structure and navigation with the original. Run applicable validators and manual or assistive-technology checks. A structurally modified file is not automatically accessible in every viewer.

Measure reliability in a representative pilot

Define acceptance criteria and use authorized files reflecting the actual mix of formats and complexity. Record all inputs, accepted outputs, errors, unavailable artifacts, review effort and corrections. Evaluate the saved result independently where possible. Include unsuccessful inputs in the denominator before making an accuracy or time-saving claim.

Aelira Core is beta software. Supported workflows can apply eligible partial fixes and record outcomes, but some inputs need manual work or cannot produce a downloadable remediated file. Review the actual saved output, unresolved findings and content preservation before publication. Automated scores describe implemented checks rather than accessibility conformance. Inspect the public source and release documentation and evaluate a representative authorized sample before planning a rollout.

Automation can assist the work; it does not prove that manual remediation is impossible, that every institution must buy AI tooling, or that only flagged items need review.