The Online Safety Act never says AI. Here is why it still binds your AI service

The Act regulates service types, not technologies

This is the design decision everything else follows from. The Online Safety Act does not list technologies and then regulate them. It defines a handful of service types, and obligations attach to whichever type you are. Three of them matter here.

A social media service (section 13) is an electronic service where the sole or primary purpose of the service is to enable online social interaction between 2 or more end-users, plus the ability to link to or interact with other end-users and to post material.

A relevant electronic service (section 13A) is a list: email, instant messaging, SMS, MMS, chat and online games. Every limb is phrased the same way, for instance a service that enables end-users to communicate, by means of email, with other end-users.

Notice what those two have in common. Both are built around other end-users. Social interaction between people, or communication with other people. A tool where one person prompts a model and gets material back has no other end-user in the loop at all.

So AI lands in the category defined by subtraction

The third type is the designated internet service, and section 14 defines it almost entirely by what it excludes. It means a service that allows end-users to access material using an internet carriage service, or one that delivers material by that means, but does not include a social media service, a relevant electronic service, an on-demand program service, a service the Minister specifies, or an exempt service, which is one where no material reaches anyone in Australia.

That is a residual category. It is what you are when you are not one of the other things. Nothing in it is about AI, and nothing in it needed to be: a generative AI service allows end-users to access material over the internet, and is none of the exclusions, so it is a designated internet service by default.

This is why the phrase designated internet service appears 136 times in the Act while AI appears zero times, and why eSafety's AI enforcement runs through a standard with that unpromising name.

A standard is what happens when a code is not delivered

Under section 135 the online industry is split into eight sections, one of which is providers of designated internet services. Industry bodies get first go at writing a code for their own section. Section 145 lets the eSafety Commissioner instead determine a standard, but only in specified circumstances: the request to develop a code was not complied with, the draft lacked appropriate community safeguards, the targets were missed, or the Commissioner refused to register the code that was produced.

The Designated Internet Services standard is one of those, made under section 145 and stating its object as to improve online safety for Australians in respect of class 1A material and class 1B material. Class 1 material is the Act’s term for the most seriously harmful content. So the standard is aimed at that material specifically: it is not a general-purpose AI safety law and does not pretend to be.

What the standard then requires, which we have covered separately

Once a service is a designated internet service, the substantive duties come from the standard rather than the Act. We read that instrument in full and set out what it obliges providers to do in Australia already regulates generative AI: the definition of a high impact generative AI DIS, the clause that lets a provider climb out of that category by building effective controls, the duty to differentiate AI outputs, and the obligation that reaches upstream to prompts. That piece is the one to read if you want the obligations. This one is about the prior question of how you end up inside its scope at all.

What non-compliance costs

Section 146 says that a participant in a section of the online industry to which a registered standard applies must comply with it, and attaches a civil penalty of 30,000 penalty units. We have deliberately not converted that to dollars here: the value of a penalty unit is set elsewhere and is indexed, so any figure we printed would quietly go stale. Look up the current unit value and multiply. The Commissioner can also issue a formal warning for a breach under section 147.

What we are not saying

This is not a claim that any AI provider is non-compliant, nor that the Act is badly drafted. Regulating by service type rather than by technology is a deliberate choice with a real advantage: a residual category catches things the drafters never imagined, which is exactly what happened here. Nor is it a claim that the Act is silent on AI harms in substance, only that it is silent on the words. We have not assessed any particular service against any tier, and we are not in a position to. This is general information about two public instruments, not legal advice.