Part of the AI Regulation News hub.
Australia's news media bargaining law defines a search service partly by whether it uses large language models, and the test decides which side of a charge boundary a service sits on
Three of the words doing the work here are undefined: large language models, primarily, and merely. That is what makes this a classification problem rather than a checkbox.
Bottom line: Binding and in force since 27 August 2026. This is dated back coverage: the Act is twelve days old and predates our current sweep window, and we flag that rather than present it as this week's news.
Who this affects: Competition and media regulatory counsel, and compliance officers, at operators of internet search and search adjacent services with Australian digital advertising revenue exposure.
Issue date: Assented 26 August 2026. In force 27 August 2026, the day after Royal Assent. No transitional deadline in the Act itself.
What changed: The statutory definition of a search service now turns in part on whether the service solely or primarily uses large language models.
Analysis: It is a boundary on one limb, not an exemption. A service failing the search limb may still be caught as a social media service under section 8, liability sits with the parent entity of the service group, and Part 5 gives the Commissioner a conditional power to negate a charge benefit obtained through a scheme.
Primary sources: News Media Bargaining (Administration) Act 2026, as made
- Instrument (EN)
- News Media Bargaining (Administration) Act 2026
- Authority
- Parliament of Australia; administered by the Department of the Treasury
- Jurisdiction
- Australia, Commonwealth
- Status
- In force. Assented 26 August 2026, registered as C2026A00077
- Bindingness
- Binding. It imposes no duty on large language models or their developers; LLMs appear only inside a definitional boundary
- Issue date / next deadline
- In force 27 August 2026. No deadline runs from this Act; the charge itself is imposed by the companion News Media Bargaining Charge Act 2026
- Number
- Act No. 77 of 2026
- Primary source
- https://www.legislation.gov.au/C2026A00077/asmade/2026-08-26/text/original/epub/OEBPS/document_1/document_1.html
What does the Act say about large language models?
Section 9 defines a search service. A service qualifies only if it is an internet search engine service, it enables searches of the internet broadly rather than of a limited database or for price comparison, and, at paragraph (1)(c), "the service neither solely nor primarily uses large language models".
Subsection 9(2) then carves back: "A service is not excluded by paragraph (1)(c) from being a search service if the service merely uses large language models to summarise or interpret search results."
Those two provisions are the only place the phrase appears. Across the whole Act, large language model occurs twice, and artificial intelligence, machine learning, algorithm, automated and generative do not occur at all.
Why a definition carries liability
Search service is not free standing. It feeds the composite term significant social media or search service in section 7, which is subject to a user threshold, and that in turn feeds the liability provision at section 13.
Section 13 applies where, in the language of paragraph (1)(c), "the sum of the totals worked out under subsection (2) for those services exceeds $250 million". Note the construction: it is an aggregate across qualifying services, and can include specified predecessor service and group revenue. It is not a single service's revenue.
The charge itself is imposed by a separate Act, the News Media Bargaining Charge Act 2026, and Part 4 of this Act provides an offset for qualifying commercial or arbitrated expenditure. That offset is conditional; among other things entitlement requires new eligible expenditure across at least eight news business corporate groups.
The reading to avoid
The tempting summary is that a service built on large language models falls outside the definition and therefore escapes the charge. That reading does not survive the rest of the Act.
Failing paragraph 9(1)(c) removes a service from one limb only. A service may still be captured independently as a social media service under section 8. Liability under section 13 sits with the parent entity of the service group, so the question is not only how one service is classified but what else the group operates.
Part 5 addresses avoidance in general terms rather than naming any particular restructuring. Where a scheme has produced a charge benefit and the purpose or effect and timing tests in sections 22 to 24 are satisfied, the Commissioner may make a declaration under Division 2 negating that benefit, and may disregard the scheme in doing so. That is conditional and discretionary, not automatic, and whether it would reach a given change of service design is a question of fact.
What makes this hard to advise on
Three words central to that classification are undefined. The Act does not say what a large language model is, what it means for a service to use one primarily, or where merely summarising results ends.
That leaves classification fact sensitive in exactly the places where a compliance answer is wanted. Section 9(2) expressly addresses merely using large language models to summarise or interpret search results, which is the conventional search engine that adds an AI summary panel. A service where retrieval and generation are no longer separable is not obviously either thing.
For counsel the practical output is not a conclusion but a record: what the service actually does at the retrieval and generation stages, documented well enough to defend a classification later.
What we did not verify
What we opened: the as made text of Act No. 77 of 2026 from the Federal Register of Legislation, 145,271 bytes, read as extracted text, including section 9, the section 13 threshold, the commencement provision and the assent statement on the face of the Act.
What we did not open: the companion News Media Bargaining Charge Act 2026, the News Journalism Payments Act 2026 and the consequential Acts in the same package, the explanatory memorandum, and any registered instrument made under this Act. We therefore describe the charge only as this Act frames it.
What we refuse to claim: we do not say that any service is exempt, because failing the section 9 limb does not by itself remove a service from the Act; section 8, the parent entity liability in section 13, and the conditional power the Commissioner holds under Part 5 are each reasons why not. We do not say Australia has enacted AI regulation, because the Act imposes no duty on large language models or their developers. We do not name any provider as inside or outside the definition, because the terms are undefined and the answer depends on facts we have not examined. We also do not state a commencement date for the charge, because that sits in a separate Act we did not open.
One sourcing note worth recording. The Federal Register page at the /asmade/text path returns a short shell that contains none of the operative text and none of the words quoted above. The text we read is at the epub path linked as the primary source.
Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.
If you advise a search or search adjacent operator with Australian advertising revenue, section 9 is a classification exercise, not a planning opportunity. Failing the large language model limb does not by itself remove a service from the Act: section 8 may still catch it, liability sits with the parent entity of the service group, and Part 5 gives the Commissioner a conditional power over schemes producing a charge benefit. Three words central to the paragraph 9(1)(c) and 9(2) classification are undefined. Document what the service does at retrieval and at generation now, while the record is easy to build.
Source File
Open the as made text and confirm four things: paragraph 9(1)(c), subsection 9(2), the aggregate wording of paragraph 13(1)(c), and the commencement provision at section 2 read against the assent line on the first page.
the service neither solely nor primarily uses large language models. ยท News Media Bargaining (Administration) Act 2026, s 9(1)(c), assented 26 August 2026
FAQ
Does this mean an AI answer engine escapes the charge?
No. Failing paragraph 9(1)(c) removes a service from the search service limb only. It may still be caught independently as a social media service under section 8, liability under section 13 sits with the parent entity of the service group, and Part 5 lets the Commissioner negate a charge benefit obtained through a scheme where the conditions in sections 22 to 24 are met.
Does the Act regulate large language models?
No. It imposes no safety, disclosure, transparency or developer duty. The phrase appears only inside the definition of a search service.
Is the $250 million figure a single service's revenue?
No. Paragraph 13(1)(c) refers to the sum of the totals worked out under subsection (2) for those services, which is an aggregate and can include specified predecessor service and group revenue.
What counts as using large language models primarily?
The Act does not say. Neither large language models, nor primarily, nor merely is defined, so classification turns on the facts of what a given service does.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.