Methodology

AI Jobs Index Australia is an independent research project tracking AI's visible footprint on Australian employment. It is published monthly by Building Tech Teams and supported by NTP Talent. The index is designed as a real-time overlay to read alongside ABS Labour Force, ABS Job Vacancies, and the JSA Internet Vacancy Index. It does not replace official statistics. It surfaces signals that move faster than quarterly official releases.

The index publishes three separate series. Each measures a different signal. The series are never summed and never subtracted from one another.

ADI · AI Diffusion Index

ADI measures the share of Australian online technology job advertisements that reference AI, machine learning, generative AI, or named AI tools and skills.

ADI is a diffusion signal. It tells you how widely AI language has spread through Australian hiring copy. It does not tell you how many new jobs AI has created, nor how many existing jobs have been redesigned around AI.

The scan covers Australian technology job advertisements only. It does not cover all Australian job advertisements. An advertisement counts toward ADI if it contains at least one term in the published ADI keyword list. The denominator is the count of all Australian-located active technology advertisements in the Adzuna feed during the measurement window. The window is a trailing 30 days. The figure is refreshed weekly.

ADI is deliberately broad. A single keyword match anywhere in an advertisement title or description counts it toward the index, and the keyword list includes short tokens such as AI and ML. This captures wide, incidental use of AI language, not only roles built around AI. ADI therefore reads higher than measures that require AI to be central to the role, and is not directly comparable to them. Read ADI as the diffusion ceiling and ASDI as the strict-demand floor.

ASDI · AI-Specialist Demand Index

ASDI counts active Australian job postings where AI is the primary work, defined by tight inclusion rules. A posting qualifies if it meets any one of three rules:

1. The title contains an explicit AI-specialist designation (AI Engineer, ML Engineer, AI Research Scientist, Computer Vision Engineer, NLP Engineer, MLOps Engineer, AI Ethicist, Responsible AI Lead, AI Governance Lead, Chief AI Officer, Head of AI, Director of AI, and a small number of related titles).
2. The job requires two or more skills from the Tier A list below.
3. The job requires one Tier A skill plus two or more Tier B skills.

Tier C skills (general data and cloud infrastructure) do not qualify a role for ASDI on their own.

ASDI is a strict subset of ADI. Every ASDI-qualifying ad also references AI, so it also counts toward ADI, but most AI-mentioning ads are not AI-specialist roles. ADI and ASDI are reported as separate figures and are not expected to be equal. A specialist-role count is always smaller than the total count of AI-mentioning ads.

Replacement vs net-new hires

ADI and ASDI count Australian technology job advertisements. Every advertised role is included, whether the role is brand new at the employer or a backfill for a departing employee. The index does not distinguish replacement hiring from net-new role creation.

This is a structural property of advertisement-based measurement, not a missing feature. A vacancy posted to an Australian job board is observable. The intent behind that vacancy is not stated in the ad and is not retrievable from any other public dataset at scale. The vacancy may replace a leaver or expand the team; the ad does not say which. ABS Job Vacancies Survey faces the same constraint.

The practical implication is on year-on-year comparison. As the dataset matures, a portion of the AI-specialist advertisement flow will be re-hiring for roles already counted in earlier periods. ADI and ASDI growth figures should be read as growth in advertised demand, not net headcount change. Where headcount change matters, the right reference is ABS Labour Force, Detailed (cat. 6291.0.55.001), not this index.

ALWR · AI-Linked Workforce Reductions

ALWR tracks publicly cited Australian workforce reductions where AI has been named as a driver or contributing factor. Every event in ALWR has a public source. Each event is classified into one of three attribution classes:

• EXPLICIT. The company has cited AI, automation, or machine learning as a driver in an official statement, press release, regulatory filing, or quoted executive interview.
• MIXED. AI has been cited as one of several contributing factors. Other factors must also be named in the source. At least one tier-A or tier-B source must cite AI specifically.
• BLAMED. A credible tier-A or tier-B media outlet has independently attributed the reduction to AI without company confirmation.

The three attribution classes are ranked by evidence strength. EXPLICIT carries the highest confidence: the company itself has named AI on the record. MIXED sits in the middle: AI is cited alongside other factors. BLAMED is the lowest-confidence tier: a media outlet has attributed the reduction to AI without company confirmation. The dashboard renders the three tiers with distinct visual weight that reflects this ranking, and the BLAMED tier is labelled on the page as “media attribution, not company-confirmed.”

The headline ALWR figure is EXPLICIT-only. EXPLICIT + MIXED and BLAMED are reported separately. The three classes are never summed into a single composite. ALWR counts are floors, not ceilings. Companies that quietly reduce headcount without public attribution to AI do not appear in ALWR, so ALWR systematically understates total AI-linked workforce change.

Primary feed

Adzuna plus major Australian job boards, accessed through the Adzuna developer-tier API. Adzuna aggregates Australian job postings from multiple boards and direct-employer feeds and exposes them through a single documented API. The index does not name individual boards in its public reporting; coverage, refresh cadence, and rate limits are documented in the project repository.

Source: The Adzuna API. Advertised-vacancy and salary figures published by the index are derived from the Adzuna API and are acknowledged as such wherever they appear.

Anchor datasets

The index reads, but does not compute against, the following official datasets:

• ABS Labour Force, Australia (cat. 6202.0). Monthly. Headline employment, unemployment, and participation. Used for national labour market context in the monthly issue.
• ABS Labour Force, Australia, Detailed (cat. 6291.0.55.001). Monthly, with quarterly industry and occupation breakdowns. Tables 13a-15 (employed persons by industry, occupation, and state) provide sector and geographic context for ADI, ASDI, and ALWR readings.
• ABS Job Vacancies, Australia (cat. 6354.0). Quarterly, employer-reported. The closest official reference for the advertised-demand signal ADI and ASDI track.
• JSA Internet Vacancy Index. National pooled-board vacancy index. Monthly. The closest official comparator to ADI in cadence and source type.
• JSA Generative AI and the Australian Labour Market. Periodic. Task augmentation and automation exposure scores by ANZSCO occupation. Read for ALWR sector commentary.

These datasets are anchors, not inputs. The index does not join ABS or JSA data into its calculations. ADI, ASDI, and ALWR are computed entirely from Adzuna ad data and the verified ALWR event set. ABS and JSA series are read alongside to interpret what an index movement means in the broader Australian labour market. The index publishes new figures monthly, the day after the ABS Labour Force release, so editorial commentary can reference current ABS context.

Worked example. If ASDI rises from 480 to 540 active AI-specialist ads over a month, the index reports 540 as the new figure. The interpretation paragraph in the monthly issue reads ABS 6291.0.55.001 Table 13a (employed persons by industry) for the same month: if Professional, Scientific and Technical Services employment was flat in that release, the ASDI rise is a hiring-intent signal running ahead of headcount change. If 6291.0.55.001 also rose, the ASDI lift probably reflects ongoing demand already showing up in headcount. Neither reading enters the ASDI calculation. Both inform the prose.

Displacement event sourcing

ALWR events are sourced from monitoring of Australian and international tier-A and tier-B outlets including ASX announcements, AFR, ABC, SMH, The Australian, Reuters Australia, Bloomberg Australia, Guardian Australia, iTnews, InnovationAus, ITWire, Financial Times, Fortune, Capital Brief, and CNBC. Tier-C outlets (Startup Daily, SmartCompany, ACS Information Age, IT Brief, Channel Life, Computerworld, ZDNet) are accepted with a MIXED classification cap. Tier-D outlets are not accepted. The full domain lists are version-controlled at lib/sources/tiers.ts.

Where an original source URL has gone offline, the index re-points to a Wayback Machine snapshot of the same article. The Wayback URL is stored, but the embedded original domain still drives the tier. A rescued Bloomberg article is treated as Bloomberg, not as archive.org. The unwrapping is implemented in unwrapWaybackUrl() in the tiers module.

Salary signals

Salary distributions are derived from Adzuna's structured salary fields where employers have disclosed compensation. The index does not source compensation data from third-party recruitment salary guides. Where advertised salary data for a role classification is thin (fewer than ten observations in the window), the role-level salary insight is suppressed rather than estimated.

Every displacement source is classified by tier. The source tier gates the maximum attribution class an event can be assigned.

• Tier A. ASX announcements, official company press releases and IR statements, ABS, parliamentary disclosures.
• Tier B. AFR, ABC, SMH, The Australian, Reuters, Bloomberg, Guardian Australia, iTnews, InnovationAus, ITWire, Financial Times, Wall Street Journal, Fortune, Capital Brief, CNBC.
• Tier C. Startup Daily, SmartCompany, FSU updates, ACS Information Age, IT Brief, Channel Life, Computerworld. Caps the event at MIXED.
• Tier D. Content aggregators, niche industry blogs, PR republishers. Not accepted.

Each ASDI-qualifying posting is enriched where the fields are extractable: seniority level, work arrangement, skills (matched against the published taxonomy), industry (mapped to ANZSIC categories), salary (AUD ranges), and visa sponsorship. Enrichment fields populated by Claude are reviewed for accuracy at the aggregate level. Individual classifications may contain errors. Aggregate accuracy is reported quarterly on the validation page.

• Live dashboard: updated weekly.
• Displacement events: media monitored daily; events classified, sourced, and human-reviewed before publication.
• Monthly report: published monthly, the day after the ABS Labour Force release.
• Quarterly benchmark note: published in the second week of each quarter, comparing index movements against ABS and JSA series.

The published weekly series begin in week 15 of 2026, the week commencing 6 April 2026. Earlier weeks are withheld.

Collection began in March 2026. An advertisement posted before then appears in the dataset only if it was still live when collection started, so the further back the record runs, the more of the market is missing from it. That is survivorship bias, and it affects every series equally. Left in place it produces a steady climb from near zero across advertised volume, ADI, and ASDI alike, which reads as rapid growth in AI hiring. It is not. It is the record of when the index started looking.

Rather than publish a curve that describes our own collection history, the index withholds every week before the series start. The underlying rows are retained in the project's records and excluded from published figures and from the weekly snapshots export.

The three series are never summed and never subtracted from one another. Whole-of-market advertised volume is not an AI measure and is not comparable with ADI or ASDI on a shared axis.

The ADI and ASDI classifiers are deterministic regular expressions, not LLM calls. Every advertisement is matched against the word-boundary patterns below. Patterns are case-insensitive. The patterns are rendered directly from the source code in lib/classifiers/adi.ts and lib/classifiers/asdi.ts, so what is published here is what the classifier runs. They are also available in machine-readable form at /methodology/regex.json.

ADI keyword patterns

An advertisement counts toward ADI if its title or description matches any one of the 30patterns below. A trailing “.ai” company-name suffix is stripped before matching so a “Software Engineer at Mistral.ai” posting does not match on the company name alone.

ASDI title designations (rule 1)

An advertisement qualifies for ASDI under rule 1 if its title matches any of the 18explicit AI-specialist designations below. Rules 2 and 3 use the Tier A and Tier B skills listed under “Skills taxonomy” below; the rule wording is in the ASDI subsection of “What this index measures” above.

The patterns are reviewed quarterly. Every change is recorded in the methodology changelog.

The ASDI skills taxonomy has three tiers. It is published in machine-readable form at /methodology/taxonomy.json and reviewed quarterly.

The classifiers are audited quarterly against a manually-labelled sample of 200 ads per series. Precision, recall, and false-positive rate are reported per category at /validation. Inter-rater agreement between two labellers is reported alongside the headline accuracy figures. The displacement classifier is audited continuously: every event passes through a human reviewer before publication.

• The index measures advertised demand and publicly cited displacement. It does not measure hires, internal moves, contract conversions, or quiet workforce changes.
• Adzuna coverage is not complete. The index underweights roles posted only on industry-specific or government-only platforms.
• Pre-November 2022 baselines for GenAI-specific roles are not available. The index does not publish year-on-year growth on GenAI-specific series with less than 36 months of data.
• AI-assisted classifications may contain errors. Aggregate accuracy is reported quarterly.
• The displacement source tier hierarchy is conservative. Events sourced only from tier-D outlets are excluded even when the underlying claim may be true.

The preferred citation format is shown below. Replace[Series name] with the specific series the figure came from (ADI, ASDI, or ALWR) and the date with the access date.

AI Jobs Index Australia (2026). [Series name], as of [date]. Building Tech Teams. https://www.aijobsindex.com.au/[page]

Each series page also includes a copy-to-clipboard citation block. The same block sits below for the methodology itself.

Questions about methodology, source additions, and data corrections: contact@aijobsindex.com.au.