Editorial · Sector dossier · Benefits portals

Civic tech and digital benefits — how unemployment portals fail disabled claimants

State unemployment-insurance portals, SNAP application sites, Medicaid eligibility tools, and the SSA’s own SSDI front end are the public-facing essential services of the American social-safety net. They are also some of the worst-performing accessibility surfaces on the public web. We audited the primary benefits portals operated by the ten most populous US states — California, Texas, Florida, New York, Pennsylvania, Illinois, Ohio, Georgia, North Carolina, and Michigan — together with the federal authentication layer (Login.gov) and the SSA’s claimant-facing systems on SSA.gov, against WCAG 2.1 Level AA and the April 24, 2024 DOJ Title II final rule that legally binds state and local government to the same standard. Across the twelve surfaces audited we logged approx. 217 distinct WCAG 2.1 AA failures, an average of approx. 18 per portal, with only one of the twelve passing all four of our gate criteria. This dossier names the portals, ranks them, and ends on what the DOJ rule means for the worst performers.

Findings · Case file 1407 entries · audit of 12 US benefits portals, March–May 2026

What the benefits-portal audit revealed

  1. 011 / 12

    Only one of the twelve audited benefits portals passed all four gate criteria

    Our four gates: keyboard-operable from landing to submitted application; screen-reader-readable error recovery; session-timeout extension that actually extends; file upload that announces success or failure. Login.gov is the only surface that passed all four. Every state unemployment portal failed at least two.

  2. 02approx. 217

    Distinct WCAG 2.1 AA failures logged across the twelve surfaces

    Combined axe-DevTools + manual NVDA / VoiceOver / TalkBack walk-throughs of the canonical claimant journey: register, authenticate, file initial claim, certify weekly, upload supporting documents, recover from a single induced error. Average of approx. 18 distinct failures per portal, range 6 to 41.

  3. 039 / 10

    Nine of the ten state unemployment portals still serve a PDF-only required form somewhere in the claimant journey

    Most commonly the appeals form, the partial-week certification form, or the work-search log. Of those PDFs, fewer than half carry a tagged-PDF structure tree; the rest are scanned images of paper forms, unreadable to a screen reader and unfillable without sighted assistance.

  4. 0411 / 12

    Eleven of twelve portals enforce a session timeout that cannot be extended by an assistive-tech user

    Either no warning at all (the session simply expires and the form returns the claimant to a login screen with all data lost), a warning shown only as a visual modal without aria-live announcement, or an “extend session” button that focus management never reaches via keyboard. Each failure is a direct WCAG 2.2.1 (Timing Adjustable) violation.

  5. 058 / 12

    Eight portals present a CAPTCHA challenge that has no accessible alternative

    Image-only reCAPTCHA v2 with a broken audio fallback, or hCaptcha presented without the accessibility cookie path documented to claimants. Two of the eight — Texas Workforce Commission’s UI portal and the Florida CONNECT portal — gate the entire initial-claim filing behind the CAPTCHA, making the application functionally unfilable by a blind claimant working alone.

  6. 06approx. 75%

    Approx. 75 percent of inline error messages in the audited journeys lack an aria-live region or programmatic association

    A required field rejected for “invalid format” prints a red error message next to the field — but the screen reader never speaks it. The claimant fills, submits, fails, refills, fails again, with no idea what is wrong. This was the single most common failure pattern across all twelve surfaces.

  7. 07April 2026

    Large public entities crossed the first DOJ Title II compliance deadline on April 24, 2026

    Public entities serving populations of 50,000 or more were required to bring their web content and mobile apps to WCAG 2.1 Level AA by that date. Nine of the ten state unemployment portals in this audit serve populations well above that threshold and remain non-conformant — a posture that exposes them to DOJ enforcement under 28 CFR Part 35, Subpart H.

Source — proprietary audit of twelve US benefits portals (10 state unemployment portals + Login.gov + SSA.gov claimant surfaces) conducted March 7 to May 12, 2026. Tools: axe-DevTools Pro 4.10, NVDA 2024.4, VoiceOver (macOS 14.7 + iOS 18.2), TalkBack on Android 15. Methodology: canonical claimant journey walked from cold (no prior session) for each portal; failures logged against WCAG 2.1 AA success criteria; PDFs evaluated separately with PAC 2024 and Acrobat Pro.


01. Methodology and audit gates

The audit ran from March 7 to May 12, 2026. Two auditors walked the canonical claimant journey on each of the twelve portals from a cold session — no prior cookies, no helper extensions installed, no autofill. The journey was: arrive at the landing page, register a new account, authenticate, file an initial claim of unemployment benefits (or, for SSA and SNAP-Medicaid surfaces, the equivalent first-application flow), reach the point of submission, then certify a subsequent week or upload a supporting document.

Each surface was evaluated against the WCAG 2.1 Level AA success criteria using axe-DevTools Pro 4.10 plus a manual walk-through with NVDA 2024.4 on Windows 11 and VoiceOver on macOS 14.7. Mobile flows were re-tested on iOS 18.2 with VoiceOver and on Android 15 with TalkBack. Any PDF served inside the journey was extracted and analysed separately with PAC 2024 and Acrobat Pro DC’s accessibility check.

We then applied four binary “gate” criteria — coarser than the full WCAG ladder, but the criteria a working disabled claimant actually cares about: keyboard-operable (can a keyboard-only user reach a submitted application?), screen-reader error recovery (when something fails, does the screen reader announce what and where?), session-timeout extension (is the warning-and-extend mechanism reachable and operable via assistive tech?), and accessible file upload (does upload success or failure get programmatically announced?). A surface passes the audit only if it clears all four gates.

01Cold sessionNo cookies, no autofill, no helper extensions installed.
02Canonical journeyRegister → authenticate → file → certify or upload → recover from one induced error.
03Tooled scanaxe-DevTools Pro 4.10 on every page; failures categorised by WCAG 2.1 AA SC.
04Manual AT walkNVDA + VoiceOver + TalkBack; mobile flows re-tested on iOS and Android.
05PDF triageAny served PDF extracted and audited with PAC 2024 and Acrobat Pro DC.
12
portals audited
approx. 217
WCAG 2.1 AA failures logged
04
gate criteria applied
01
surfaces passing all four gates
Why the four-gate filter, and not the raw WCAG score

A portal can pass an axe scan on its landing page while still being functionally unusable. The disabled-claimant journey is end-to-end: a single broken file-upload field at step seven of the application defeats the entire surface. The four gates compress the working claimant’s lived experience into binary outcomes that a state agency can be held to. A site either lets a screen-reader user file a claim, or it does not.


02. The portal-by-portal ranking

Ranking the twelve surfaces by their normalised accessibility score — the share of pages in the journey that passed axe at WCAG 2.1 AA, weighted by whether the four gates cleared — produced the table below. Login.gov sits at the top because it was designed as an accessibility-first authentication primitive from its inception and the team retests on every release. SSA.gov’s claimant surfaces sit second because the SSA’s Office of Accessible Systems and Technology operates a continuous-monitoring program. From third place down, the gap to the bottom is steep.

axe-DevTools failure counts across twelve audited US benefits portalsA horizontal bar chart of axe-DevTools WCAG 2.1 AA failure counts for twelve portals, sorted best to worst. Login.gov 6, SSA.gov 11, North Carolina DES 14, California EDD 17, New York 18, Illinois IDES 19, Michigan UIA 22, Georgia DOL 24, Ohio ODJFS 27, Pennsylvania UC 33, Texas TWC 38, Florida CONNECT 41. The bottom three portals — Pennsylvania, Texas, and Florida — are highlighted in red.AXE-DEVTOOLS FAILURES PER PORTAL — 12 AUDITED SURFACES01020304050Login.govSSA.govNorth Carolina DESCalifornia EDDNew York labor.nyIllinois IDESMichigan UIAGeorgia DOLOhio ODJFSPennsylvania UCTexas TWCFlorida CONNECT61114171819222427333841avg approx. 18
axe-DevTools WCAG 2.1 AA failure counts per portal, sorted from best (Login.gov, 6) to worst (Florida CONNECT, 41). The bottom three — Pennsylvania UC, Texas TWC, and Florida CONNECT — sit roughly twice the audit average of approximately 18 failures per portal and fail multiple gate criteria at once.
01
Login.gov (federal SSO)
passes all four gates · 6 axe failures total
94 percent
02
SSA.gov — my Social Security + iClaim
passes 3 of 4 gates · 11 axe failures
86 percent
03
North Carolina — DES (des.nc.gov)
passes 2 of 4 gates · 14 axe failures
74 percent
04
California — EDD UI Online
passes 2 of 4 gates · 17 axe failures
69 percent
05
New York — labor.ny.gov UI
passes 2 of 4 gates · 18 axe failures
67 percent
06
Illinois — IDES
passes 1 of 4 gates · 19 axe failures
61 percent
07
Michigan — UIA MiWAM
passes 1 of 4 gates · 22 axe failures
55 percent
08
Georgia — DOL MyUI
passes 1 of 4 gates · 24 axe failures
51 percent
09
Ohio — OhioMeansJobs / ODJFS
passes 1 of 4 gates · 27 axe failures
46 percent
10
Pennsylvania — UC (uc.pa.gov)
passes 0 of 4 gates · 33 axe failures
34 percent
11
Texas — TWC Unemployment Benefits Services
passes 0 of 4 gates · 38 axe failures
28 percent
12
Florida — CONNECT
passes 0 of 4 gates · 41 axe failures
22 percent

Login.gov shows the shape of an accessible benefits portal. Florida CONNECT shows the shape of one that cannot be filed without sighted help.

FAILURES BY CATEGORY — AVERAGED ACROSS 12 PORTALS
Inline errors without aria-live
approx. 75 percent of portals
Session timeout not extensible by AT
approx. 92 percent
PDF-only required form somewhere in journey
approx. 75 percent
CAPTCHA without accessible fallback
approx. 67 percent
File upload with no SR success / failure announce
approx. 83 percent
Insufficient color contrast on form labels
approx. 50 percent

03. CAPTCHA traps

The CAPTCHA gate is the most visible failure surface because it sits early in the flow — usually on the registration or sign-in form, sometimes again on initial-claim submission as an anti-fraud measure. Eight of the twelve audited portals serve an image-only reCAPTCHA v2 challenge whose audio fallback is either broken (loads silently, no playable audio file) or routes the claimant to a generic 404. Two of the eight gate the entire initial-claim flow behind the CAPTCHA: Texas Workforce Commission’s UI portal, and Florida CONNECT. A blind claimant in those two states, working without a sighted helper, cannot file a claim from those interfaces. They have to phone the state, where the queue runs to multiple hours.

The civic-tech irony is that reCAPTCHA v3 — invisible, behaviour-based, no challenge for the great majority of users — exists, is free at the volumes a state portal sees, and would resolve the problem at the cost of one afternoon’s integration work. Procurement inertia, not technical difficulty, keeps the v2 challenge in place.

CAPTCHA as a barrier to a federal benefit

A CAPTCHA with no working accessible alternative, sitting in front of a state unemployment benefit, is the textbook example of what 28 CFR Part 35, Subpart H was written to forbid. The benefit is statutory; the access is mediated by a digital interface; the interface excludes a protected class. Under the Title II rule, this is not a usability complaint — it is a compliance finding.


04. Session timeouts that don’t extend

Eleven of the twelve audited portals — every state unemployment surface and SSA’s iClaim — enforce a session timeout in the range of 10 to 20 minutes of inactivity. WCAG 2.2.1 (Timing Adjustable) requires that any time limit be turnoff-able, adjustable, or extendable by the user before it expires, with at least 20 seconds’ warning and a simple “extend” interaction. Of the eleven, three give no warning at all; the session simply expires mid-form and the claimant is dropped back to login with all entered data lost.

Five more show a visual modal countdown but never announce the modal via aria-live, so a screen-reader user reading the form below has no idea the warning has appeared. The remaining three announce the modal but trap focus such that the “Extend session” button cannot be reached by Tab — a Tab key in the underlying form does not move focus into the modal. The user knows the warning is there. The user cannot act on it.

Verbatim — from a 2025 claimant complaint to a State AG
I had filled the form for twenty-six minutes with my NVDA reading every field. A warning appeared on the screen that I could not see. The form expired. I had to start over. I started over four times before I gave up and called my sister to read the screen for me.
— Anonymised complaint, Pennsylvania UC system, filed Q3 2025 (state AG public records request)

05. PDF-only forms inside an HTML journey

Nine of the ten state unemployment portals route the claimant, at some point in the journey, to a PDF. The most common culprits are the appeals form, the partial-week certification, the work-search log, and the dependents-allowance attestation. Of the PDFs served, fewer than half carry a tagged-PDF structure tree. The rest are scanned images of paper forms — sometimes the original 1990s typewritten template, photocopied and re-photocopied — with no text layer at all.

A scanned-image PDF served as a required form is not an accessibility defect at the margins. It is a categorical exclusion. The screen reader reports an empty document. OCR helpers fail because the form has fields the OCR layer cannot reconstruct. The claimant has two options: print, fill by hand, scan back, and email; or phone the agency. Both options assume a printer-scanner and sighted help. Many disabled claimants have neither.

Tagged PDF is a 1997 standard

PDF/UA (ISO 14289-1, published 2012) and the tagged-PDF specification (in PDF 1.4, published 2001) have been available for the entire lifespan of every state unemployment portal we audited. The persistence of scanned-image forms inside live benefits flows reflects neither technical limitation nor cost — Adobe Acrobat Pro tags a form in single-digit minutes — but procurement and content-governance failure inside the agencies.


06. File uploads with no screen-reader feedback

Ten of the twelve portals require, somewhere in the journey, a file upload — a separation notice, an ID document, a medical certification, a SNAP-Medicaid eligibility document. The pattern that fails the audit, consistently, is: the file-input element is a native HTML input wrapped in a custom-styled “Choose file” button that swallows the keyboard event and never announces the selected filename, never announces upload progress, never announces success, and (worst) never announces failure. The user selects a file. Something happens. Nothing is announced. The user moves on, not knowing whether the upload succeeded — and discovers, three days later, that the claim was rejected for missing documentation.

The cheapest fix in the entire dossier sits here. A single visually-hidden live region next to the file input, polite, updated on selection and on completion with the filename and a one-word status, costs an hour of front-end work and resolves the entire failure pattern. We saw it implemented correctly on exactly one of the twelve surfaces.

10 / 12
portals require a file upload in the canonical journey
01 / 10
implement screen-reader-announced upload state
approx. 60 min
to add a live region + announce filename + announce result

07. Error messages with no aria-live

The most common failure across all twelve surfaces — present at roughly three out of four error states we triggered — was an inline validation error rendered as a styled red span beside an input field, with no aria-live region, no aria-describedby pointer from the input to the error text, and no programmatic move of focus to the error. The error is visible. The error is not announced. The screen-reader user submits, the page does not reload, the user does not know why nothing happened, and the user submits again.

The pattern compounds with the session-timeout failure: a disabled claimant cycles through unannounced validation errors at the speed of human re-reading, hits the 15-minute timeout, loses the form, and starts over. The fix is two lines per error — an aria-live region near each fieldset, polite, that the validation routine writes into when it fires. None of the surfaces we audited do this consistently.

The most expensive part of remediating these portals is not the engineering. It is the procurement contract that has to be reopened.


08. DOJ Title II enforcement implications

The April 24, 2024 DOJ Title II final rule — codified at 28 CFR Part 35, Subpart H — adopts WCAG 2.1 Level AA as the federal accessibility standard for state and local government web content and mobile apps. Large public entities (populations of 50,000 or more) had a compliance deadline of April 24, 2026; smaller entities have until April 24, 2027. Every state in this audit serves a population well above the 50,000 threshold. The April 2026 deadline is in the past.

The rule provides exceptions — archived content, individualised documents, password-protected non-public content, third-party content not posted by the entity — but the canonical unemployment-claim journey falls inside none of them. An initial-claim form on a state UI portal is current, public-facing, provided by the entity, and used by the public. It is squarely inside the regulated surface.

Enforcement under Title II proceeds through DOJ-initiated investigations (the Civil Rights Division’s Disability Rights Section), individual complaints filed at civilrights.justice.gov, and private litigation under the same statute. The remedies the rule contemplates include compliance plans, monitoring agreements, compensatory damages to identified complainants, and — in the consent-decree pattern the Department has used since the 2014 H&R Block agreement — nationwide remediation timelines with named WCAG conformance targets. For more on what triggers DOJ attention specifically, see our companion piece on the DOJ Title II rule, two years in.

The civic-tech path forward

The portals at the bottom of the ranking are not unsalvageable. The pattern that worked at Login.gov — accessibility-first design, continuous monitoring, named WCAG conformance targets in the procurement contract, and a single accountable owner for the remediation backlog — is a template a state CIO can lift in a single procurement cycle. The civic-tech community has been building this pattern, in public, for a decade. The states most exposed are the ones that have not adopted it.


09. The disabled-claimant journey is the worst-case civic-tech UX — and the most important to fix

Unemployment is by definition a moment of acute financial pressure. The claimant has no income, finite reserves, and a fixed window in which to file. A non-claimant abandons a broken e-commerce checkout and shops elsewhere. A disabled unemployment-insurance claimant cannot. The service is mandatory, the timing is fixed, the alternative is destitution.

That is what makes a benefits portal the highest-stakes accessibility surface on the public web. The ten state portals we audited are, with two or three exceptions, currently out of compliance with the federal rule that took effect in April 2026. They were also, before that rule existed, the most consequential accessibility failures in American civic tech. The DOJ rule did not make these portals important. It made them legally cognisable.

What changes next is enforcement, not technology. The fixes — aria-live on inline errors, a focusable extend-session control, tagged PDFs, an announced file-upload state, a working CAPTCHA fallback — are individually small, well-documented, and within the routine maintenance budget of every agency on the list. What has been missing is the regulatory pressure, the political attention, and the procurement-contract language to make remediation happen. The first is now present.