The Digital Duty of Care bill now names AI services. It still repeals the standard that makes some of them test models and differentiate outputs

We read the exposure draft on 6 October and said the bill should state whether model testing and output differentiation are meant to survive the repeal of the codes. The introduced bill, 90 pages with a 105-page explanatory memorandum, answers half of that, and not in the Act.

What changed between the draft and the bill

AI services are in by name

The draft’s AI paragraph had two limbs. It caught a service that allows users to generate material by means of artificial intelligence and share it through one of the listed services. On a natural reading, a chatbot whose output stayed between the user and the model was not caught by that paragraph, and would have been reached, if at all, as a designated internet service, the route the current Act uses.

The bill replaces it with one limb. Section 25B(1)(h) makes an online service of a service that generates material using artificial intelligence and that uses an internet carriage service in the course of doing so. The memorandum says the paragraph is there For the avoidance of doubt. The duty in section 26, with its 60,000 penalty units in section 26C, and the risk assessment in section 27 therefore reach a standalone AI service with no question about sharing. The bill still uses the words artificial intelligence twice and chatbot and generative not at all; the second is still the Commissioner’s sock puppet power in section 205K.

Model testing reaches the memorandum, not the Act

Section 27 requires a written risk assessment at least annually and before any significant change that could add risk, and lets the Commissioner set standards, benchmarks or other metrics for it. It does not mention models. The memorandum does. A provider of a public-facing generative AI service may need to assess the risk that the service will generate harmful material, and its controls may include testing before deployment, testing against reasonably foreseeable misuse, safeguards applying to user prompts and generated outputs, along with reporting processes, monitoring after deployment and review of whether safeguards still work.

That is close to what the designated internet services standard requires of a high impact generative AI service today: regular review and testing of models for misuse to generate child sexual exploitation or pro-terror material. The difference is the word may, and where it sits. An explanatory memorandum helps a court read the Act; it does not impose a duty. The standard’s other AI requirement, to implement systems, processes and technologies that differentiate AI outputs generated by the model, has no counterpart. We searched the bill and the memorandum for marking, labelling, watermarking, provenance and AI-generated content and found none in this sense; the only related hit is synthetic child abuse material, listed as a harm.

The codes go, but first they are shored up

Schedule 3 is unchanged in substance. Its title is still Repeal of online content scheme industry codes and standards, its item 5 still repeals Division 7 of Part 9, and the memorandum’s note on that item reads, in full, This item repeals the Division. The transitional items in Schedule 2 continue authorisations, agreements and guidelines made under the Commissioner’s current functions section, and the office of the Commissioner. None continues a code, a standard or the basic online safety expectations, which Schedule 2 item 13 repeals.

Two things are new, and both take effect the day after Royal Assent, a year before the repeal. Items 84 and 85 lift the penalty for breaching a code direction under section 143(2) and a standard under section 146(1) from 30,000 to 60,000 penalty units. In the draft the same rise started on the day the provisions were repealed, so it would never have applied; now it covers the codes’ last year. And items 90 and 91 add a validation clause to section 140, which registers codes, and section 145, under which the Commissioner determines standards. A code or standard already registered is taken never to have been affected in its validity only because a period allowed for the purposes of this Subdivision was shorter than required. The Act sets those periods: at least 30 days for public and industry comment on a draft code (section 140(3)) and on a draft standard (section 148(2)), among others. The memorandum restates the clause and does not say which code or standard, if any, it was written for.

The Minister for Communications, Anika Wells, put the trade in her second reading speech. The Act has 17 industry-drafted codes and standards, and services complying with them and with the expectations will be well placed to comply with their digital duty of care obligations. That is a statement about readiness, not continuity: being well placed to meet the duty is not the same as being required to keep testing a model or differentiating its outputs.

What happens next

• 30 October 2026: submissions close to the Senate Environment and Communications Legislation Committee.
• 20 November 2026: the committee reports.
• The day after Royal Assent: Schedule 1 starts, including the higher code penalties, the validation clause and the take-down notices for fake nude material.
• 12 months after Royal Assent: the duty starts, and the codes, standards and expectations are repealed. The minister called it a 12-month transitional period for rules and guidance.

Our view

Naming AI services outright is the right fix. The draft’s two-limb test left the chatbot whose output stays with the person using it to an older and less certain route, and the bill closes that. The new age assurance duty in section 27D reaches those services too.

The model duties are the gap that remains. The government plainly knows what good practice for a generative AI provider looks like, because its own memorandum lists it. Writing it there rather than in the Act means that, from the repeal date, nothing in the Online Safety Act will expressly require an AI service provider to test a model before release or to mark what it makes; any such requirement would have to come from the general duty, a Commissioner direction or risk assessment requirements the Commissioner has yet to set. If that is the plan, the committee should hear it said, and anyone making a submission by 30 October has a concrete question to put: should the section 27 risk assessment name model testing and output differentiation for generative AI services, or should the Commissioner be required to? The validation clause invites one more question, since the memorandum does not answer it: which instrument, if any, was it written for?

Sources

Method note. Written substantially by AI and run by people, working from the 8 primary sources below. How AI Geek works.

1. Online Safety Amendment (Digital Duty of Care) Bill 2026, first reading, Parliament of Australia, 90 pages (read 16 October 2026): s 25B(1)(h); ss 26C, 27, 27D; Sch 1 Part 4 items 84 and 85 and Part 5 items 90 and 91; Sch 2 item 13 and transitional items 54 to 57; Sch 3; commencement table.
2. Explanatory memorandum, Parliament of Australia, 105 pages (read 16 October 2026): s 25B(1)(h) For the avoidance of doubt; the generative AI controls example under section 27; items 90 and 91; Schedule 3 item notes.
3. Bill homepage, r7553, Parliament of Australia (read 16 October 2026): government bill, House of Representatives, introduced and read a first time 15 October 2026, second reading moved 15 October 2026, referred to committee 15 October 2026.
4. Inquiry page, Senate Environment and Communications Legislation Committee, Parliament of Australia (read 16 October 2026): referred 15 October 2026; submissions close 30 October 2026; reporting date 20 November 2026.
5. Second reading speech, Minister for Communications, House Hansard, 15 October 2026, proof (read 16 October 2026): 17 codes and standards; well placed; 12-month transitional period.
6. Exposure Draft, Online Safety Amendment (Digital Duty of Care) Bill 2026, Department of Infrastructure, Transport, Regional Development, Communications, Sport and the Arts, 8 September 2026, 75 pages (read in full 6 October 2026, re-read 16 October 2026): s 25A(1)(j); the penalty items in Schedule 2; no age assurance provision.
7. Online Safety Act 2021, Federal Register of Legislation, compilation C2026C00396, compilation date 12 September 2026 (read 6 and 16 October 2026): ss 140(3) and 148(2), 30-day comment periods; ss 143(2) and 146(1), 30,000 penalty units.
8. Online Safety (Designated Internet Services, Class 1A and Class 1B Material) Industry Standard 2024, eSafety Commissioner (PDF, read in full 4 August 2026, section 22(3) re-read 6 and 16 October 2026): the minimum requirements for a high impact generative AI DIS in section 22(3), including model testing and output differentiation.

How we checked this. The bill and memorandum were downloaded from ParlInfo and compared line by line with the exposure draft we read on 6 October; every quotation is verbatim. The designated internet services standard is quoted from the standard itself. The search for output marking terms covered both full texts. The second reading speech is quoted from the proof Hansard, which can change before the official record. We have not asked the department or eSafety what the validation clause was written for or whether the model duties are meant to return through Commissioner requirements. The “our view” section is opinion based on the documents cited.

Making a submission, or read the bill differently? Tell us and we will check it against the document and log the outcome here.