Bike Riding Robot
← All postsMarkdown →

Australia's Education AI Budget

Four states have each independently developed generative AI chatbots for their school systems, all within roughly the same two-year window, all built on Microsoft's Azure OpenAI platform, all solving near-identical problems, with no shared code, no joint procurement, and limited coordination. New South Wales has spent approximately $3.7 million developing NSWEduChat, with a 2024-25 budget of around $5 million.[1] In May 2026, Western Australia committed a further $4.6 million to extend its ClassmAIte pilot.[2] South Australia's EdChat[3] and Queensland's Corella[4] have followed parallel paths. The total public investment across these parallel builds, before accounting for ongoing infrastructure, maintenance, staff time, and technical debt, is substantial and growing.

None of that spending is indefensible on its own terms. Each state faced real pressure to act, real community concern about students using AI without guardrails, and real political incentive to show something tangible. What is harder to defend is the pattern: four separate governments building four near-identical tools on the same underlying technology, at roughly the same time, without a shared approach.

The Commonwealth is not a bystander to this pattern; it is paying for it. The Australian Government's Workload Reduction Fund has put federal money directly into the parallel builds a coordinated approach would avoid: $1.5 million toward expanding NSWEduChat, $2.3 million into a separate Western Australian cross-sectoral pilot, and $0.5 million into a third generative AI pilot run by the Association of Independent Schools of NSW.[5] Most strikingly, the fund has also put $700,000 into Cechat, a second and entirely different chatbot built by Catholic Schools NSW — operating in the same state as NSWEduChat, solving the same problem, funded by the same federal program, a second time. This is not a failure to prevent duplication. It is duplication with a federal grant attached.

NSWEduChat's own public communications trace a version of this drift in miniature. It was announced in February 2024 as a direct response to the state's 2023 ban on ChatGPT in schools, explicitly framed as a safe, curriculum-aligned alternative for students to use in place of the banned tool.[6] In practice, staff and teachers were given access first, in October 2024, with system-wide student access following a full year later, in October 2025.[7] By April 2026, the department's own promotional material for NSWEduChat was organised entirely around staff productivity, built around prompt libraries for "explicit teaching, assessment, communication and administration," with no mention of student use at all.[8] A tool announced as a safe alternative for students has, in its own department's public communications, become a teacher efficiency product.

The Body We Already Have

Australia already has a mechanism designed for exactly this kind of coordinated national effort in education technology. Education Services Australia (ESA) is a national not-for-profit company owned by all state, territory and Commonwealth education ministers. Its explicit mandate is to provide technology-based shared services for education nationally, connecting policy, technology and practice to advance initiatives that benefit every school system simultaneously. ESA built and operates the national curriculum and assessment infrastructure the school system already depends on.[9]

When all education ministers endorsed the Australian Framework for Generative AI in Schools in October 2023, they committed $1 million to ESA to develop product expectations for AI technology and update privacy principles.[10] A million dollars to set standards for what would turn out to be a technology each state then spent many multiples of that building independently.

ESA was not empowered to build a shared national platform. Whether that was a deliberate decision or an absence of ambition is not clear from the public record. What is clear is that the natural vehicle for a coordinated national approach was given a standards brief while the building happened separately, in parallel, across four states at once.

That may be changing, though not necessarily in the direction this argument would prefer. In October 2025, federal and state Education Ministers agreed in-principle to fold ESA into a new Teaching and Learning Commission, alongside the national curriculum authority (ACARA), the national teaching-standards body (AITSL), and the national education research agency (AERO).[11] The stated purpose is coordinating curriculum, teaching, assessment, research and reporting; on the public record, a mandate to build or procure shared technology is not part of it. The pattern of reluctance is consistent elsewhere too: when a House of Representatives inquiry into generative AI in education reported, the Government's formal response, tabled in April 2026, supported only the recommendation to fully fund schools and merely "noted" every recommendation bearing on national coordination of AI tools.[5] A body with the standing to lead a national platform is being created. Whether it will be asked to is a separate, and still open, question.

This is worth naming not to assign blame, but because it points directly to the opportunity. ESA exists. It has the ministerial ownership structure. It has the relationships across every education system in the country. A shared national platform, procured once, maintained centrally and iterated continuously, is not a novel idea requiring new institutional machinery. It requires a decision to use what we have.

What Other Countries Did

The comparison with comparable education systems is instructive, not because any single model is directly transferable, but because the direction of travel is consistent.

The United Kingdom invested approximately £3 million through the Department for Science, Innovation and Technology to build a national curriculum content store, a library of structured educational data made available to private developers.[12] A separate £994,000 through the Department for Education's Contracts for Innovation funded multiple companies to build AI tools using that foundation.[13] An independent EdTech Evidence Board, run by the Chartered College of Teaching, evaluates and publishes efficacy findings so schools can make procurement decisions with verified evidence.[14] Government as infrastructure provider and standard-setter, not product builder.

Over the month of July 2026, a related version of that same principle has emerged in the United States, built without any government money at all. Learning Commons, launched in 2025 as the education arm of the Chan Zuckerberg Initiative's philanthropic infrastructure work, maintains a Knowledge Graph: a machine-readable dataset mapping curricula, learning-science research and academic standards across all fifty US states, built to be, in Learning Commons' own words, "an open resource available to all AI-powered edtech and LLMs."[27] When Anthropic launched Claude for Teachers on 14 July 2026, giving verified US K-12 educators free access to Claude[21], it connected to that Knowledge Graph rather than building a proprietary standards database of its own, and the lesson-planning and lesson-differentiation skills the two organisations co-developed were published to Learning Commons' public GitHub under an open licence, available to any developer building education AI, not only Anthropic.[27] No department, ministry or state built or funded the shared layer here; a philanthropic initiative built it once, a private company built a product on top of it, and both released the connecting infrastructure rather than keeping it proprietary. It is the same principle behind the UK's content-store model, applied with private and philanthropic capital instead of public money — and the principle ESA is already structurally positioned to apply to Australian schools, if it is asked to.

Singapore's Student Learning Space, built in 2018 and continuously upgraded by GovTech, is frequently cited as justification for sovereign builds in Australia.[15] The comparison does not hold. Singapore has one Ministry of Education, one national curriculum, one examination system, and GovTech as a world-class central engineering agency with more than two decades of continuous platform investment. No Australian state has any of these prerequisites. The lesson Singapore offers is not "build it yourself" but "first invest, over decades, in the institutional capacity to build well."

Globally, the trend is away from custom builds. The Open Contracting Partnership's November 2025 analysis, drawing on interviews with more than fifty public-sector practitioners and experts across the United States, Europe and beyond, found that governments are primarily purchasing off-the-shelf licences through existing cloud and productivity platforms rather than commissioning bespoke systems.[16] The states' parallel build approach is running against the dominant direction in public-sector AI.

A Product Without a Research Base

There is a prior question the parallel-build story tends to skip. A generative AI chatbot deployed to every student in a school system is, in substance, an experimental education intervention at system scale. Normally an intervention like that would go through some version of a pilot, a control group and a published efficacy result before a government committed seven or eight figures to a permanent build. There is no public evidence any of the four states did this. Each chatbot was scoped, procured and shipped inside an election-cycle timeframe, on the strength of general enthusiasm about generative AI rather than evidence specific to how a chat interface changes how a teenager learns.

Khan Academy's Khanmigo is the useful cautionary case here, not because it is comparable in scale, but because it is close to a best case. Khan Academy is a long-established, single-purpose education nonprofit with a larger pedagogical dataset and a longer-running AI-tutoring research effort than any Australian state education department has assembled. If any organisation was positioned to get this right on the first attempt, it was Khan Academy. Its founder, Sal Khan, nonetheless acknowledged in July 2026 that the first version of Khanmigo "did not change student learning as much as many of us hoped," and that adoption was weak enough that "too many students who had it available did not even try it."[17] Khan Academy's own reporting on the tool identifies engagement, not model capability, as the recurring obstacle to it producing better outcomes.[18] If the organisation best resourced to get a student AI tutor right needed three years and a public admission of shortfall to find a version that works, the question of what research base four separate state departments were relying on when they committed tens of millions of dollars to their own versions, built faster and with far less specialised research capacity, answers itself.

Competing With a Market Australians Already Use

The market context for these builds is not hypothetical, and it is distinctly Australian. Anthropic's Economic Index found Australia using Claude more intensively per person than any other country it tracks, at roughly six times the rate its population would predict, with homework — help with assignments, study material and tutoring — the single largest use case, at around one in ten conversations.[19] Australian students are not a future market for direct-to-consumer AI. They are already the heaviest per-capita users of it anywhere in the world, and they are already using it overwhelmingly for the exact task, homework, that state chatbots are built to intercept.

That the frontier labs have not extended a free education tier to Australia sharpens this point rather than undercutting it. OpenAI's ChatGPT for Teachers and Anthropic's Claude for Teachers, both free for verified educators, are US-only programs with no Australian equivalent.[20][21] Australian teachers get no comparable subsidy from either company; Australian students, meanwhile, are already reaching for the unrestricted consumer product at the highest rate measured anywhere, regardless of what a state department builds and irrespective of any US-specific offer. The chatbot each state is spending millions on is not competing with a free product Australian teachers can access. It is competing with a habit that is already dominant among Australian students, on a platform no state department controls.

What Teachers Actually Need

Every sovereign build applies the same content filters and departmental guardrails to a qualified teacher with a four-year degree and years of classroom experience as it does to a Year 7 student. There is no differentiated access model. Compare this to AI use in other professions: clinicians use tools their patients cannot access; legal practitioners use full-capability systems not constrained to client-facing portals. The sovereign build model substitutes a government product for professional trust.

The UNESCO Teacher Task Force's September 2025 position paper on teacher agency in the age of AI is direct on this point: technology must amplify professional judgment, not constrain it.[22] Effective AI in education requires teachers who are empowered to experiment, evaluate and adapt, rather than issued a departmentally approved tool built to the risk appetite of a regulatory instrument.

Learning First's May 2026 survey of more than four thousand NSW teachers and school leaders found educators felt largely unsupported in navigating AI.[23] The gap was not a missing chatbot. It was the absence of genuine professional agency: time, trust, and the freedom to work out what actually helps students learn. A filtered departmental tool addresses none of that.

The opportunity here is significant. Rather than building products for teachers, governments could invest in the professional infrastructure that lets teachers make good decisions about the products that already exist, products that will keep improving faster than any sovereign build can track.

The Crisis That Has Not Been Addressed

Every state AI chatbot built in the past two years is focused on content generation. None addresses the actual crisis that AI has created for Australian schooling: assessment integrity.

Learning First's 2026 survey found that 80 per cent of lower-secondary teachers and 73 per cent of senior-secondary teachers who report their students use AI say those students use it to complete assessment tasks.[23] The HSC, VCE, QCE, WACE and SACE all rely heavily on written tasks, including essays, short-answer responses and extended research assignments, that any current frontier AI model can complete to a high standard. The take-home assignment has, in practice, become an unverified AI output for a significant proportion of students.

The response from universities has been swift and structural. The University of Sydney[24] and the University of Melbourne[25] have moved substantial assessment back to supervised, in-person examination, recognising that the problem is architectural rather than technological. School systems have not made the equivalent move.

That gap raises a further question: what do state departments actually think the future of assessment looks like once AI capability keeps compounding, and whose thinking are they drawing on to work it out? Universities have already made structural changes to assessment and explained their reasoning in public. Employers and professional bodies are grappling with the same collapse in the reliability of unsupervised written work as it flows into hiring and credentialing. There is little public evidence that the departments spending tens of millions of dollars on chatbots are engaging with either group, or treating assessment redesign with anything like the urgency and resourcing given to the tools that sit alongside the problem rather than inside it.

Assessment reform is structural, profession-wide, expensive, and requires admitting that decades of standard practice have been overtaken by a tool every student carries in their pocket. Building a departmental chatbot is comparatively straightforward: visible, technically legible, and easy to announce. The Productivity Commission's August 2025 interim report called for national coordination of both AI tools and curriculum materials, recognising that fragmented state-by-state responses leave the structural questions unanswered.[26] The Government's own April 2026 response to the House of Representatives inquiry into generative AI in education runs to fifteen pages without addressing assessment integrity at the school level once.[5] Assessment integrity is the most urgent of those questions.

The Opportunity

None of this is an argument against AI in Australian schools. It is an argument for doing it in a way that produces durable value.

The opportunity, available now using existing institutions, is a national shared platform procured once against published safety, privacy and curriculum standards, maintained and iterated centrally, and made available across every state and territory system. ESA is the obvious candidate to lead or coordinate that work. The Productivity Commission has pointed in this direction. The evidence from comparable countries supports it. The economics of avoiding further parallel duplication are compelling.

The second opportunity is to direct the investment saved by not building four more state-specific tools toward the problem that no chatbot can solve: an assessment architecture that was designed before AI existed and has not been substantively reformed since. That is the work that would genuinely change outcomes for students.

Australia has the institutions, the evidence, and the spending already underway to make better decisions from here. The question is whether the next round of investment follows the pattern of the last two years, or takes a different path.

References

  1. NSW Parliament, Legislative Council, Question on Notice No. 3627, "AI Tool EduChat", and Question on Notice No. 3724, "Budget for EduChat", May 2025. See also NSW Department of Education, NSWEduChat. ↩
  2. Government of Western Australia, "$4.6 million for AI platform designed to reduce teachers' workloads", media statement, 9 May 2026. ↩
  3. South Australian Department for Education, EdChat, and "AI tool rolled out to support teachers". ↩
  4. Queensland Department of Education, "Generative AI in Schools" (Corella). ↩
  5. Australian Government response to the House of Representatives Standing Committee on Employment, Education and Training, Study Buddy or Influencer: Inquiry into the Use of Generative Artificial Intelligence in the Australian Education System, tabled April 2026. ↩
  6. EducationHQ, "NSW schools to trial AI alternative following last year's ChatGPT ban", 7 February 2024. ↩
  7. iTnews, "NSW Education AI tool set to launch for students from October", 23 September 2025. ↩
  8. NSW Department of Education, "Unlock the game-changing capabilities of NSWEduChat", 30 April 2026. ↩
  9. Education Services Australia, "About ESA", and ESA Strategic Plan 2024–26. ↩
  10. Australian Government Department of Education, "Australian Framework for Generative Artificial Intelligence (AI) in Schools", 5 October 2023. ↩
  11. Australian Government Department of Education, 2026–27 Budget factsheet — Teaching and Learning Commission, 12 May 2026. ↩
  12. UK Department for Science, Innovation and Technology, AI content store contract awarded to Faculty AI, August 2024, reported in Schools Week and Computing; see also UK Department for Education, "Artificial intelligence in schools: everything you need to know", Education Hub blog, June 2025. ↩
  13. UK Research and Innovation / Innovate UK, "Contracts for Innovation: AI Tools for Education", funding competition, September–October 2024. ↩
  14. Chartered College of Teaching, EdTech Evidence Board. ↩
  15. Singapore Ministry of Education, Student Learning Space; GovTech Singapore, Student Learning Space; Singapore Ministry of Education, EdTech Masterplan 2030. ↩
  16. Open Contracting Partnership, "The surprising shifts in how the public sector is buying AI — and what policymakers can do about it", 10 November 2025. ↩
  17. Sal Khan, LinkedIn post, 15 July 2026, reported in EdTech Innovation Hub, "Sal Khan says early Khanmigo fell short as Khan Academy rebuilds AI tutor", 17 July 2026. ↩
  18. Kristen Eignor DiCerbo, Khan Academy, quoted in "Can an AI-Powered Tutor Produce Meaningful Results?", Education Week, July 2025. ↩
  19. Anthropic Economic Index, "Cadences", June 2026 report, findings on Australia reported in Forbes Australia, "Australia is the world's biggest Claude user. Now Anthropic wants more", 8 July 2026. ↩
  20. OpenAI, "A free version of ChatGPT built for teachers", 19 November 2025 (available to verified US K-12 educators only). ↩
  21. Anthropic, "Introducing Claude for Teachers", 14 July 2026 (available to verified US K-12 educators only). ↩
  22. International Task Force on Teachers for Education 2030 (hosted by UNESCO), "Promoting and Protecting Teacher Agency in the Age of Artificial Intelligence", September 2025. ↩
  23. Learning First, "AI Use in Schools: Taking Action Now", May 2026 (survey of approximately 3,400 teachers and 750 school leaders across NSW government, Catholic and independent schools). ↩
  24. University of Sydney, Educational Innovation, "Assessment in 2023". ↩
  25. Higher Education Policy Institute, "How the University of Melbourne moved to digital in-place exams at scale", October 2024. ↩
  26. Productivity Commission, Building a Skilled and Adaptable Workforce, interim report, 11 August 2025. ↩
  27. Learning Commons, "Learning Commons provides key educational infrastructure for Anthropic's Claude for Teachers", 14 July 2026. ↩