Switch off zoom and transition animations for a static presentation.
The SURF 2026 poster
Tap or click any part of the sheet to open it.
Whole poster 100%
Poster map
Tap or click any part of the sheet to fly into it. Everything the printed sheet could not fit is here: full thresholds, denominators, exclusions and limits.
Identity
The registered project, its team and its supervisors.
Valuation of AI Skills in Online Marketplaces — a job-anchored map of textual alignment between sampled public AI-skill pages and business-school courses.
Project supervision
Supervisor — Jia Hao Bryan Chai.
Co-supervisor — Joshua Lee Sze Ray.
What the title does and does not say
This is the title the project was registered under.
It is not a finding. What this study measures is textual alignment between sampled documents, not price, capability, demand or realized value.
Why the title and the measure are stated separately
A project title is registered before the work is done, so it describes an intended area of study rather than a result.
The poster therefore prints the registered title as identity and states the measure it actually used immediately beside it, so neither is read as the other.
This study compares sampled AI-skill pages from one marketplace with business-school courses using occupation- and task-derived topics.
How the abstract bounds itself
Here, relative value means textual alignment, not monetary value.
The map identifies shared topics for review; it does not establish capability, adoption, automation, replacement, causation, demand, price, or monetary value.
Reading an abstract that defines its own terms
Words such as value and alignment carry everyday meanings much broader than the one measured here, so the abstract redefines them before using them.
When it says relative value, it is naming a comparison between the wording of two sets of documents, and nothing else.
How closely do sampled AI-skill pages align with business-school courses?
The measure used here
Job-anchored textual alignment between sampled documents, not price, capability, or realized value.
One shared vocabulary makes unlike documents comparable.
What job-anchored means
Both sets of documents are described using the same fixed vocabulary of work topics, derived from occupation and task text rather than from either set being compared.
Anchoring to that outside vocabulary is what makes a course description and a marketplace page comparable at all, since neither is written in the other's language.
AI-skill value depends on context and the outcome being measured.
Three findings from prior research set up the narrower question this study asks.
Skill change is contextual
Digital technologies reshape skills through mechanisms that are not fully captured by simple automation-versus-augmentation binaries.
Grimshaw & Miozzo, 2026
AI can complement or substitute
Effects depend on tasks, technical capabilities, workflows, and organizational change.
Tschang & Almirall, 2021
Value is measured by outcomes
Research on AI-skill value uses job-posting demand and wage premiums; sector studies examine quality, fees, and employment outcomes.
Alekseeva et al., 2021; Fedyk et al., 2022; Friedman et al., 2026
The gap this study addresses
This study isolates one narrower construct: textual alignment between sampled AI-skill and curriculum documents in a job-anchored topic space.
Why the literature matters to a wording study
The cited work measures outcomes in the world — what employers pay for, what gets hired, what changes in a workflow. This study measures none of those.
Setting the two side by side is what keeps a document comparison from being mistaken for an outcome, which is the confusion the poster is built to prevent.
Occupation and task text define the fixed topic map. Course documents and public AI-skill pages are the two sets compared on it.
How to read the counts
First: rows collected. Second: documents placed on the shared map. Different stages, not interchangeable.
Where the AI-skill documents come from
The skill corpus is built from unique public AI-skill pages on skillsmp.com, deduplicated by page URL before preparation.
One page is not one tested capability.
Why the two counts differ
Not every collected row enters later analysis, so the drop between the two stages is not a claim about quality or loss.
Documents may share repositories or content.
Why two different counts appear for each corpus
The first count is how many rows were collected. The second is how many documents were placed onto the shared topic map, which is a later and separate stage.
They are different units describing different moments in the pipeline, so swapping one for the other would misstate the base every later result rests on. The poster prints both, side by side, for exactly that reason.
One job-topic map makes two document sets comparable.
Fit a fixed set of job topics from occupation and task text, project both document sets onto it, screen the topics down to the interpretable ones, then compare.
The screening rule, in words
Keep factors with one dominant pattern and a top word outside the filler list; the second signal must be no more than 75% of the first.
The full collection and preparation route
Course and programme text came from official university sources. Public skillsmp.com pages were collected by a separate part of the team.
Course records kept their year and their source, and were checked for quality and completeness. Duplicate skill pages were removed before preparation.
Where the method comes from
The general textual-factor method is Cong, Liang, Zhang and Zhu (2025). The labour-market adaptation is this project's.
The topic vocabulary is derived from the O*NET occupation and task database.
What a textual factor is
A textual factor is a recurring pattern of words that tends to appear together across a body of text. Fitting them produces a fixed list of topics, each recognisable by the words that load onto it.
Once that list exists, any document can be described by how strongly it relates to each topic. That description is what makes two very differently written documents comparable.
Why the topics are fitted on neither corpus
The topics are derived from occupation and task text, not from the courses and not from the marketplace pages.
If the vocabulary had been learned from either set, that set would have defined the terms of its own comparison. Fitting it on outside text keeps the map neutral between them.
Why a screen is applied at all
Not every fitted topic is interpretable. Some spread their weight across many unrelated patterns, and some are dominated by filler words that carry no subject matter.
The screen keeps the topics a reader could name and describe, and the poster states its rule in words so it can be argued with rather than taken on trust.
Breadth counts how many topics appear in both document sets. Depth describes where course alignment is stronger.
Breadth, across corpora
How many topics appear in both document sets?
Count a factor as present when at least 10 distinct documents in a corpus link to it at an absolute loading of 0.05 or more. Shared means both corpora meet that rule.
Depth, within courses
Where is course alignment stronger?
Average skill-document hit rates across course topics, weighted by stronger course links.
The exact depth definition
Textual alignment depth is a course-level squared-loading-weighted mean of factor-specific skill-document hit rates over screened factors; it is undefined when a course hits none.
Two questions that sound alike but are not
Breadth asks whether a topic shows up in both sets at all. It produces a yes or no for each topic, then counts the yeses.
Depth asks, for a course, how consistently the marketplace pages engage the topics that course leans on most. It produces a score rather than a count.
A topic can be shared under the breadth rule while the depth score for a discipline stays low, because the two are measuring different things.
Why there is a threshold at all
Any two large bodies of text will brush against each other somewhere. A rule that counted a single faint mention would report near-total overlap and mean nothing.
Requiring a minimum number of distinct documents, each above a minimum strength, is what makes shared a statement about a pattern rather than about a coincidence.
How many screened job-topic factors meet both document thresholds?
The sampled documents share many job-anchored topics.
The threshold behind the word shared
Shared = a factor linked by at least 10 distinct documents in each corpus, at an absolute loading of 0.05 or more.
How to read the proportion
The map identifies shared thematic ground for closer review.
Thresholded textual overlap is not demonstrated capability or replacement.
What the result does not mean
It does not mean AI can perform 69.5% of a course or job.
What the bar is a proportion of
The bar's whole length is the set of screened topics — the ones that survived the interpretability screen. The filled part is those that cleared the threshold in both corpora.
It is a share of topics, not a share of a course, a job, a task, or anyone's time. Every one of those would be a different measurement, and none of them was made here.
Depth is a weighted skill-document hit-rate score, not curriculum share.
The base these values are computed on
Depth base: 30,213 classified course documents with at least one hit; 3,848 classified no-hit documents excluded and 4,703 unclassified omitted. Skill hit rates use 73,749 projected documents.
How far the ranking can be pushed
Descriptive only; ranks are not tested as statistically different.
Unequal samples, no uncertainty intervals and no significance test.
Why some bars are outlined
Outlined bars mark disciplines with fewer than 500 included course documents; read those ranks with extra caution.
Why an outline rather than a colour
The small-sample warning is carried by the bar's outline and by the document count printed beside it, never by hue alone.
That way it survives greyscale, colour-vision differences and a photocopy — the same reason the chart labels its values directly instead of relying on a separate legend.
Why the ranking should be read loosely
The disciplines are ordered by score, but no test was run to establish that neighbouring positions differ at all, and the number of documents behind each varies widely.
The order is a description of the sample in hand. It is a prompt for a closer look, not a league table.
The map identifies shared thematic ground for closer review, and each audience has a different next step.
For universities and students
Use the map to identify topics for curriculum review and independent tool testing.
For skill builders and marketplaces
Use shared topics identified by the map to prioritize capability and demand validation before making value claims.
How firm these two statements are
Both are interpretation, not measured outcomes.
What a screening tool is, and is not
A shared-topic map narrows where to look. It says which subjects appear on both sides of the comparison, so attention can go there first rather than everywhere at once.
It cannot say whether the tools work, whether anyone is using them, or what any of it is worth. Those are separate questions needing separate evidence, and this map is a way of choosing which ones to ask.
Scope: sampled public AI-skill pages from one marketplace and business-school course documents.
Not established here
Task performance, adoption, or skill quality
Automation, worker or course replacement
Causal impact, demand, or monetary value
On the rankings
Ranks are descriptive; no significance test.
Why a poster states what it did not find
A shared-topic map is easy to over-read. The step from these documents discuss the same topics to this work can be done by a machine is a large one, and nothing here supports it.
Listing the conclusions the evidence cannot carry is how the poster keeps that step visible rather than leaving it to the reader to notice.
