Australia has an AI safety agreement with Canada. It is not binding, has no money, and one deadline lands next month

Australia signs a lot of AI cooperation instruments and almost none of them are readable. This one is, because the Department of Industry, Science and Resources published the operative text rather than a summary of it. That is worth taking advantage of, because “we signed an AI safety agreement” is a sentence that can mean almost anything, and here you can check exactly what it means.

What it commits the two countries to

The participants are the Department of Industry, Science and Resources of Australia and the Department of Industry of Canada, and the memorandum states its purpose as providing a framework for technical cooperation in AI safety. Seven areas of cooperation are listed, and they are genuinely substantive:

• exchanging information on AI technologies, risks, methodologies, tools and evaluations
• coordinating work toward internationally recognised and interoperable guidelines and standards for evaluating AI
• joint studies on best practice for the risks of synthetic content significantly modified or generated by AI
• joint research and development on AI measurement science, including tools, methods and frameworks for testing and evaluation across the full life cycle of AI systems
• joint research on mitigating risks from advanced AI systems, including technical safeguards and oversight methodologies, building on earlier collaboration on deepfakes
• exchanges of talent between the two countries, including between the two AI safety institutes and between non-governmental organisations
• anything else the two sides agree to add

The two dates that make it checkable

Most cooperation instruments are impossible to hold to account because nothing in them has a clock. This one has two.

Within 30 days of coming into effect, each side must designate the entities and officials responsible for overseeing the memorandum and tell the other side in writing. That fell in early April.

Within 180 days, the two sides must jointly develop a workplan to carry out the memorandum's purpose and achieve its intended outcomes. Measured from 5 March, that lands around 1 September 2026. Whether a workplan exists by then, and whether anyone publishes it, is the single most informative thing about whether this agreement is doing anything.

What it is not

The memorandum is admirably direct about its own limits, and a reader should know them before treating it as a safeguard.

Paragraph 12.1: the memorandum “represents the understanding reached between the participants and does not create any legally binding rights or obligations”. Paragraph 10.1: it “does not have any financial implication for either participant” and is not a basis for either funding the other. Paragraph 10.2: each side bears its own costs. Paragraph 10.3: it creates no expectation that either will make additional spending commitments. Paragraph 15.1: either participant may terminate on six months' written notice. Disputes are to be resolved by consultation and “will not be referred to any national or international court, tribunal or third party”.

Two further limits shape what can actually be shared. The memorandum does not authorise sharing classified, sensitive or protected information carrying a security classification. And it is not intended to cover the sharing of information about an identified or reasonably identifiable individual, with each side to comply with its own privacy law if any activity turns out to involve personal information.

A timing detail worth noticing

The departmental announcement says the two governments “will collaborate through their AI safety institutes”, and then says: “Australia will establish our AI safety institute in the first half of 2026.” So an institute-to-institute memorandum was signed while one of the two institutes was still to be stood up.

That is not improper and it is not unusual in international cooperation, where a framework often precedes the body that will use it. It is worth noting because it is the second time we have found Australia's AI Safety Institute credited with activity that predates it: when we checked the Institute's claims we found a multi-agent risk report attributed to its work had been published four months before the Institute was announced. Neither is a scandal. Together they suggest the Institute's public history should be read with the dates checked.

One observation we cannot fully explain. Australia's AI Safety Institute page carries a short news list, and that list includes a May 2026 memorandum with the United Kingdom on safe and secure AI. It does not mention Canada. A news list is not an index and nothing turns on it, but if you were looking for this agreement from the Institute's own page you would not find it.

What we will be watching

Our view, labelled as such. A non-binding, unfunded memorandum is not a weak instrument by definition. Technical cooperation between measurement scientists rarely needs a treaty, and the areas listed here, particularly the deepfake and evaluation work, are exactly what these institutes should be doing together. What makes it real or not is the workplan.

So the test is simple and dated: is there a joint workplan by about 1 September 2026, and can the public see it? The memorandum requires the participants to coordinate in advance on public releases about their activities, so a published workplan is permitted. If nothing appears, the honest description of this agreement will be a statement of intent between two departments that cost nothing and obliged no one.

Spotted an error? Tell us and we will check it against the sources and log the outcome here.