Independent analysis · prepared to support adoption

// IBM SkillsBuild — Artificial Intelligence Certificate

A strong certificate, and a workable path to adopting it.

The IBM SkillsBuild Artificial Intelligence Certificate is 72 hours of self-paced learning across 13 modules and 4 capstone projects, in four microcredentials. It is free to students, carefully built, and structurally well matched to California’s noncredit rules. This analysis sets out what a California college adds around it to make it credit-bearing — and the short list of questions that IBM and the college can close together.

AI fluency Responsible AI Governance Prompt engineering Human-in-the-loop Enterprise AI

Free to learners. Access is arranged through IBM — the college registers and IBM adds it to a partner group space.

73.7 hrstotal learner effort
13 + 4modules + capstones
5Credly credentials
1.0–1.5 unitsCalifornia equivalent
01 / what this is

What the certificate actually is

Thirteen modules and four embedded capstone projects, grouped into four microcredentials, each carrying one learning outcome and issuing one Credly badge, plus a culminating certificate. Delivery is asynchronous and self-paced on skills.yourlearning.ibm.com. The final capstone additionally requires IBM watsonx Prompt Lab.

A precision point. IBM’s 72 hours is total learner effort, in IBM’s own words — not contact hours. At the Chancellor’s Office recommended divisor of 54 total student-work hours per unit, that is 1.33 units; at the 48-hour minimum divisor, 1.5. Presenting it as contact hours would inflate the unit calculation roughly threefold, and is the single most likely technical error to sink a curriculum submission. Summing every unit in IBM’s own syllabus gives 73.7 hours, not 72 — cite the summed figure.

// scored work

The modules are completion-based. The certificate contains 6 scored assessments in total — 4 capstone projects, 1 final assessment, 1 final capstone project — each with unlimited attempts at a 80% pass mark. IBM’s faculty playbook confirms the capstones and final assessment are auto-scored.

02 / what the college is signing up to support

What the delivery chain means for adoption

Sampling 21 lesson views of one module revealed something more consequential than the module itself. Delivering this certificate involves six hostnames across three organisations. The learning platform is Adobe Learning Manager on a European regional endpoint, and the interactive activities come from H5P, a third company. Only the first host is named anywhere in IBM’s public materials, its faculty playbook, or the certificate documentation.

  1. skills.yourlearning.ibm.com IBM

    Catalogue, learning plan, progress tracking. The only host a college sees named anywhere.

  2. alm.ibm.com IBM

    IBM’s Adobe Learning Manager tenant. Launches the module and frames it as course:4068807.

  3. learningmanagerfr.adobe.com Adobe

    Adobe Learning Manager, EU regional endpoint. Referenced as the frame source; redirects onward.

  4. cpcontents.adobe.com Adobe

    Adobe Captivate prime-player, build tagged prodfr. Three further nested documents inside it before the course itself.

  5. video.ibm.com IBM

    Embedded video, iframed into lesson views.

  6. csr.h5p.com H5P

    Embedded interactive activities. A third organisation entirely, named nowhere in IBM’s materials.

  7. the authored course IBM content

    Reached three frames deeper still, via Module Player and a frame correctly titled AI Impact. The content itself is well structured — one h1, clean heading order, lang set, image alt text present.

Consequence

Accessibility review covers the whole chain

A conformance report for any single layer would not describe what a student experiences. Each layer contributes its own controls, focus behaviour, and markup.

Consequence

The sub-processor list is longer than IBM

The learning platform is Adobe Learning Manager, on a European regional endpoint, and the interactives come from H5P. Whatever the institutional agreement says about IBM, it has to answer for both — and for where learner data sits.

Consequence

Fixes have to come from IBM

The content layer is cross-origin from IBM’s wrapper, so the college has no technical means to fix a defect locally. That makes a working support channel to IBM part of the adoption plan, alongside an alternative path for students where needed.

// and the authored content is genuinely well built

This deserves as much prominence as the defects. Across the 21 lesson views of the module examined: 67 images, none missing alt text. Exactly one <h1>, clean heading order, lang set, 49 focusable elements. The START MODULE control renders as an <a> rather than a button — a common keyboard trap in authored courseware — but it carries a real href and is properly focusable. A second module matches on every measure.

The pattern is careful authoring and defective integration. That points remediation at IBM’s wrapper and embed configuration, not at the instructional designers.

// every embedded media and interactive frame is unnamed

Sweeping 21 lesson views found four embedded frames — one video, three interactive activities — and not one of them carries a title. A 100% failure rate. These are not decorative. A screen reader user navigating by frame reaches four unnamed regions and is told nothing about what any contains. WCAG 4.1.2 Name, Role, Value.

The wrapper defect is systematic too: module 1 is framed as title="course:4068807", module 2 as title="course:4068808". Two for two — this is how IBM’s integration names every module, not one bad frame. Adobe’s own layers name theirs correctly, including marking decorative ones Intentionally blank, so the defect sits in the layer IBM controls.

Deliberately not a finding

One result that looks alarming and is not

The browser accessibility tree read from IBM’s wrapper returned essentially nothing — a single generic node, no content beneath it. Read carelessly that says the course is invisible to assistive technology.

It almost certainly means the inspection tool did not traverse a cross-origin frame. It is recorded here specifically so nobody later reports it as a defect. Settling it requires a real screen reader, which is the only authoritative check.

For the security review, not this one

Two things worth a second pair of eyes

Every H5P activity frame carries allow="autoplay *; geolocation *; microphone *; camera *", delegating those permissions to a third-party origin. Whether H5P exercises any of them is not established, and it may simply be an embed-template default — but it is a concrete question for IBM.

Separately, the Captivate player is invoked with an access_token in the URL query string; the value is recorded nowhere in this analysis. Tokens in URLs are exposed via history, referrers, and logs. Neither of these is a vulnerability claim.

Observed 27 July 2026 on a learner account, viewing without completing. 21 lesson views of module 1 (MC1 · AI Impact) were sampled — an incomplete subset, not the whole module. The Rise player unmounts off-screen sections, so a DOM count at one moment undercounts; the module contains further lessons that were not examined. Module 2 was sampled and found consistent. The capstones, final assessment, watsonx Prompt Lab and Credly remain unexamined, as do the screen-reader, reflow and contrast passes. Keyboard focus was verified to reach the content frame, but whether focus is visibly indicated inside it could not be resolved across the cross-origin boundary and needs a sighted human or a screen reader. Nothing in this section is a conformance determination.

04 / adoption pathway

Five routes, ranked

Ranked by speed to students against approval burden. Routes 1 and 4 both require the accessibility evaluation first, because both make a third-party authenticated platform a required element of instruction.

#RouteApproval body
1 Embed as courseware in an existing course
Fastest credit-bearing route — term scale. Requires the RSI wrapper and the accessibility evaluation.
Department and discipline faculty. No new curriculum action if outcomes and hours are unchanged.
2 Not-for-credit community education
Fastest overall, but no credit, no apportionment, and typically a student fee. Approval mechanics unverified.
Local administration only.
3 Credit for Prior Learning (§55050)
Awards credit for the credential without teaching it, sidestepping interaction rules. Needs a host course outline.
Board CPL policy, curriculum-committee procedures, discipline faculty.
4 Noncredit CDCP Short-term Vocational sequence
Slowest, and the only route both free to students and revenue-positive. The four microcredentials map onto the required sequence.
Curriculum committee, governing board, Chancellor’s Office, EDD determination.
5 Standalone credit Certificate of Achievement
Weakest fit — 1.0–1.5 units is far below a viable standalone certificate.
Curriculum committee, governing board, Chancellor’s Office.
// no ACE recommendation

ACE’s National Guide lists two IBM SkillsBuild entries — Cybersecurity (IBM-0028, 12 credits) and Data Analytics (IBM-0029, 8 credits), both expiring 30 June 2027. There is no Artificial Intelligence row. IBM’s faculty playbook states a recommendation is “in progress” but attaches no date. Title 5 §55050(d), which obliges colleges to consider ACE recommendations, is therefore an empty hook for this product today.

05 / labour market

What the labour market actually shows

California EDD projections for occupations an AI-literacy certificate could plausibly contribute toward, led by the Los Angeles metro division rather than statewide averages. U.S. Professional & Business Services — employment: 22,507K (June 2026, BLS CES).

// read this before the wage columns

9 of 12 of these occupations list a bachelor’s degree as typical entry-level education — and the 3 that do not are the 3 that are flat or declining. The growth and the wages sit entirely on the far side of a degree.

That is the empirical case for embedding this certificate in a credit pathway rather than offering it as a terminal credential. Presented as a standalone route to an AI job, the data does not support it. Presented as one input that stacks toward credit, it does.

Source: California EDD via data.ca.gov — Long-Term Occupational Employment Projections + OEWS wages. Regional figures are Los Angeles-Long Beach-Glendale MD (Los Angeles County), 2023-2033. Statewide is California, 2024-2034. Wage bands are OEWS 2025 1st Qtr for Los Angeles-Long Beach-Glendale MD.
Occupation (SOC) LA median wage LA growth LA annual openings CA growth
Reachable without a degreeEDD lists entry at high school or some college. These are the occupations this certificate could contribute to directly — and all three are flat or shrinking.
Computer User Support Specialists
15-1232
$73,698+4.4%9,930-2.2%
10th $48,49625th $60,960median $73,69875th $90,81790th $106,737
Secretaries and Administrative Assistants, Except Legal, Medical, and Executive
43-6014
$53,120+3.2%54,230+1.4%
10th $38,81725th $46,453median $53,12075th $63,74490th $76,789
Customer Service Representatives
43-4051
$48,057-5.7%55,830-5.6%
10th $38,05825th $40,982median $48,05775th $60,02490th $74,586
Bachelor’s degree is typical entryWhere the growth is. A 72-hour non-credit certificate does not qualify a learner for these; it is one input among several on a path that runs through credit.
Data Scientists
15-2051
$129,699+33.7%7,300+37.1%
10th $68,09325th $85,363median $129,69975th $168,68890th $205,718
Software Developers
15-1252
$161,781+12.1%25,050+18.4%
10th $97,62225th $127,057median $161,78175th $184,11490th $218,984
Management Analysts
13-1111
$109,712+10.8%32,500+11.5%
10th $60,28125th $81,249median $109,71275th $165,33290th $190,475
Computer Occupations, All Other
15-1299
$104,996+9.8%12,130+11.1%
10th $47,12925th $70,511median $104,99675th $157,75190th $195,332
Market Research Analysts and Marketing Specialists
13-1161
$81,066+9.3%27,840+9.4%
10th $47,97625th $62,685median $81,06675th $119,55790th $159,808
Marketing Managers
11-2021
$165,938+8.9%12,690+8.1%
10th $89,12625th $120,172median $165,93875th $220,47690th $239,200+
Computer Systems Analysts
15-1211
$130,668+8.7%7,710+9.7%
10th $79,22225th $101,970median $130,66875th $144,58590th $172,781
Project Management Specialists
13-1082
$105,744+8.4%20,890+6.8%
10th $61,75725th $80,382median $105,74475th $136,77190th $172,357
Technical Writers
27-3042
$103,172+6.6%6,240+1.9%
10th $103,17225th $103,172median $103,17275th $103,17290th $103,172
// the two columns are different vintages

EDD publishes the Los Angeles metro division at 2023-2033 and California at 2024-2034. There is no LA 2024-2034 series and no California 2023-2033 series, so the regional and statewide columns are different vintages and are not strictly comparable. They are shown together because the regional figure is the one a local labour-market question actually asks about, but the comparison is indicative rather than exact and should be described that way in any submission.

06 / what the college adds

The faculty wrapper California requires

Federal law (34 CFR §600.2) and California law (Title 5 §55204) both require that a distance course include substantive interaction between the instructor and students, comprising at least two of five enumerated activities. Every one of those activities is framed as conduct by the instructor — so a vendor platform, however good, supplies none of them.

IBM delivery model

Independent

“Fully self-paced with minimal faculty involvement”

Fails interaction requirements. Not viable for credit.

IBM delivery model

Hybrid

“Self-paced with periodic check-ins or facilitated sessions”

Marginal. “Periodic” is not evidently predictable and scheduled.

IBM delivery model

Fully embedded

“Integrated into a course with ongoing faculty support”

The only defensible model, if support is scheduled and documented.

IBM’s own playbook is candid about the model: “Faculty support learner progress rather than deliver traditional instruction,” with faculty acting as facilitators. Of the five faculty contributions IBM lists, only one — clarifying concepts and answering questions — maps cleanly onto a required element, and it is reactive rather than scheduled. IBM’s model supplies at most one of the required two.

This is not a criticism of IBM’s design. The certificate is built for a global, largely non-credit-bearing audience where these rules do not apply. It is a statement of what a California credit deployment must add: faculty-scored capstones against a faculty-authored rubric, plus either faculty-facilitated discussion on a published cadence or faculty-delivered direct instruction.

// the strongest citation you have

Title 5 §55204(d) declares that “Regular and substantive interaction is an academic and professional matter pursuant to sections 53200 et seq.” That makes the Academic Senate’s authority over this decision regulatory, not merely customary. Note also that §55204(a) applies to any portion of a course conducted at a distance — there is no minimum threshold.

07 / accessibility

Accessibility: strong content, one evidence gap

We tested the courseware directly rather than relying on documentation. The headline is favourable: on every criterion tested so far, the authored content passes. The issues that did surface sit in the integration layer, and they are small, specific, and cheap to fix. The table below separates what has been reviewed from what has not — the second half is genuinely untested, not quietly passing.

CriterionResultEvidence
1.1.1 Non-text ContentPasses 67 images sampled across one module — zero missing alt text.
1.2.2 Captions (Prerecorded)Passes Captions provided and verbatim accurate, correctly punctuated and capitalised — consistent with human authoring rather than raw speech recognition.
2.1.1 KeyboardPasses Checkbox questions operate with Space. The drag-and-drop activity is keyboard operable — Space grabs an item and the drop zones become live targets.
2.4.7 Focus VisiblePasses Clear high-contrast focus ring throughout; the page scrolls to bring the focused element into view.
2.5.7 Dragging MovementsPasses Click-to-select then click-to-place works with a single pointer, no dragging. A “No dropzone” target returns a placed item. Tested and reverted.
4.1.2 Name, Role, ValueNeeds a fix Content inside the frames is correctly named; the embedding <iframe>s carry no title. A one-line embed fix.
Screen reader (simulated)Passes 68 focusable elements, none unnamed. H5P announces “Grabbable 1 of 6” and “Dropzone 1 of 3” with a live region. role="application" still needs a real AT session.
Not yet reviewed — no conclusion should be drawn from the absence of a finding below this line.
3.3.1 / 3.3.3 Error handlingPending Capstone submission form: field labels, error identification and suggestion.
2.2.1 Timing AdjustablePending Whether the capstone or final assessment imposes a time limit.
1.4.1 Use of ColourPending How auto-scored results are conveyed — text, or colour alone.
1.4.10 Reflow · 1.4.12 Text SpacingPending 400% zoom / 320 px reflow and text-spacing passes.
1.4.3 / 1.4.11 ContrastPending Text and non-text contrast measurement.
IBM watsonx Prompt LabPending A separate platform required by the final capstone. IBM publishes 152 watsonx ACRs — those get checked first.
Credly claim flowPending Credential claiming. Credly’s published VPAT is v2.4 (May 2022), WCAG 2.0 only, with 19 self-reported “Does Not Support” findings.
// the drag-and-drop works by keyboard

This is the finding worth highlighting, because a drag-and-drop activity is the most common hard blocker in authored courseware. In the Categorize tasks by suitability for AI activity, Tab reaches each item, Space picks it up, and the category zones become live drop targets. The instruction wording — “Match each task… then select Submit” — avoids both “drag” and “click”, which is deliberate good practice.

It also satisfies WCAG 2.5.7 Dragging Movements: a single click selects an item and a second click places it, with a “No dropzone” target to undo. A keyboard path and a single-pointer path — belt and braces, and the credit belongs to H5P.

Two refinements would still help: Escape does not cancel a grab, and the keyboard interaction is never explained on screen, so a keyboard user has to discover it.

// simulated screen-reader pass — the content is well named

Inspecting the accessibility tree — the names, roles and relationships assistive technology consumes — found 68 focusable elements in the course layer with not one missing an accessible name, five landmarks, six headings with no level skips, a “Skip to lesson” link, and four live regions.

The H5P activity is the standout. Draggable items announce as “Grabbable 1 of 6. Automatically tagging photos…” and the categories as “Dropzone 1 of 3. Suitable for AI”, with a polite live region to announce placements. Those positional counts are exactly what a non-visual learner needs to know where they are in a set. This is deliberate, well-executed assistive-technology support.

Which sharpens the one defect. H5P named everything correctly and gives its activity a meaningful document title — but the <iframe> placing it in the lesson carries no title. A screen reader user arrives at an unnamed region containing a well-built activity they were told nothing about. That is a one-line fix in the embed configuration, in the layer IBM controls, requiring nothing from H5P or the instructional designers.

What this method cannot settle

One thing genuinely needs a real screen reader

The H5P wrapper uses role="application", which tells screen readers to pass keystrokes to the widget rather than use their own browse mode. It is a defensible choice for drag-and-drop and it is what makes Space-to-grab work — but application mode is a known sharp edge, and behaviour differs across NVDA, JAWS, and VoiceOver.

Inspection cannot assess that. It is the clearest example of why the simulated pass is a starting point and a real assistive-technology session remains the authoritative check.

The one gap is published evidence. IBM operates a public accessibility conformance report portal covering roughly 4,100 offerings. Filtered by hand on 27 July 2026:

152ACRs for watsonx
0ACRs for SkillsBuild
0ACRs for Prompt Lab

The watsonx result confirms the filter works, so the zeroes are a genuine searched-and-not-found. This is emphatically not evidence that the platform is inaccessible — our own testing above found the opposite. IBM commits its teams to WCAG 2.2 Level AA under the IBM Accessibility Requirements v7.3, a higher bar than the DOJ rule’s 2.1 AA. The product appears to meet that standard; what is missing is the published report saying so.

That matters only because California procurement asks for the document. The college’s own signoff requires a VPAT, an internal review, or an approved alternative before a platform becomes mandatory. An ACR for the SkillsBuild delivery stack would remove that step entirely — and given watsonx already has 152 published reports, the capability plainly exists inside IBM.

On timing: DOJ’s ADA Title II web rule adopts WCAG 2.1 Level AA, and an Interim Final Rule of 20 April 2026 extended both compliance tiers by a year, to 26 April 2027 and 26 April 2028. Neither has passed, so there is room to do this properly rather than in a rush. Which tier applies to a community college district turns on whether it counts as a special district government — a question for district counsel, not for this analysis.

08 / status board

What is done, and what is still open

Published in the open so nobody mistakes a partial evaluation for a complete one. Anything not marked complete has not been examined, and no conclusion should be drawn from its absence.

  1. Complete

    Certificate structure

    All 13 modules and 4 capstones parsed from IBM’s syllabus.

  2. Complete

    Delivery chain

    Frame tree walked; six hostnames across three organisations identified.

  3. Complete

    Keyboard & focus

    Tab, Space and focus visibility tested on knowledge checks and the drag-and-drop.

  4. Complete

    Captions

    Caption presence and verbatim accuracy checked on one video.

  5. Complete

    Screen-reader pass (simulated)

    Accessibility-tree and ARIA inspection of the course layer and the H5P activity. A simulation — a real NVDA, JAWS or VoiceOver session is still recommended, particularly for the role="application" behaviour.

  6. Complete

    2.5.7 by single pointer

    Confirmed: click-to-select then click-to-place completes the drag-and-drop with no dragging. The activity satisfies both 2.1.1 and 2.5.7.

  7. Queued

    Capstone project

    Submission flow, error states, and auto-scoring feedback.

  8. Queued

    Final assessment

    Timing behaviour (2.2.1) and attempt handling.

  9. Queued

    IBM watsonx Prompt Lab

    A separate platform, required by the final capstone.

  10. Queued

    Credly claim flow

    Credential claiming; the published Credly VPAT is from 2022 and covers WCAG 2.0 only.

  11. Queued

    Reflow & contrast

    400% zoom / 320px reflow, text spacing, and colour contrast passes.

  12. Queued

    Remaining modules

    Modules 3–13. Modules 1 and 2 matched on every structural measure, so these are expected to be consistent.

// on the screen-reader pass

The screen-reader work uses accessibility-tree inspection — reading the same structure that assistive technology consumes. That is a simulation, and it is genuinely useful for catching missing names, roles, and relationships. It is not a substitute for a real NVDA, JAWS, or VoiceOver session, which is the only authoritative check and remains a recommended step before anything is filed as a conformance claim.