We read the bill as introduced on 16 October and said the committee had a concrete question to settle: whether the section 27 risk assessment should name model testing and output differentiation for generative AI services, and which instrument, if any, the validation clause was written for. The 135-page report, published on its reporting date, answers neither. It also tells us something we had not reported: the House of Representatives agreed to the bill on 28 October 2026, with 15 Government amendments and eight crossbench amendments.
The four recommendations
These are the committee’s recommendations in full, as the report numbers them (paragraphs 3.88 to 3.91). The committee is chaired by Senator Varun Ghosh.A
- Recommendation 1.
The committee recommends that the Australian Government introduce amendments to the bill to establish a civil right of action for individuals in respect of breaches of the Digital Duty of Care that cause harm or loss.
- Recommendation 2.
The committee recommends that the Australian Government, as part of the statutory review, consider the effectiveness of the Digital Duty of Care in addressing online hate and vilification directed at individuals or groups on the basis of protected characteristics.
- Recommendation 3.
The committee recommends that the Australian Government amend the bill to clarify the scope of application, specifically to confirm:
the bill’s intention to regulate digital communication platforms, such as social media platforms; and
that the bill is not intended to apply to low-risk websites or online services, such as small business websites, online market places, booking platforms, or religious websites focused on education or doctrinal instruction.
- Recommendation 4.
Subject to consideration of Recommendations 1 and 3, the committee recommends that the Senate pass the bill.
Exhibit A · from the source
What it says about AI services
The committee view, paragraphs 3.78 to 3.87, does not use the words
artificial intelligence
, AI
or chatbot
. Nor does any recommendation.
Section 25B(1)(h), which since introduction makes a service that generates material
using artificial intelligence and that uses an internet carriage service in the course of
doing so
an online service in its own right, is not discussed in the committee’s
reasoning.
AI does appear in the evidence the report summarises, mostly as a question of scope:
- Dr Honni van Rijswijk praised the bill’s reach across the technology stack,
including
artificial intelligence generation services
(paragraph 2.40). - DIGI argued that
The duty reaches almost every business online
, listinghosting, search, app stores, AI services and equipment suppliers
, and asked for clarity on business systems such asenterprise AI tools
, whichhave a very different risk profile
(2.41 and 2.42). - Dr van Rijswijk also proposed adding
manipulative artificial intelligence companions
to the harms children are protected from (2.70). - The eSafety Commissioner’s General Counsel, Richard Fleming, told the committee
that among the criteria eSafety is considering for the lowest risk platforms is
where there are no AI chatbots
(2.142). - On apps that generate fake nude material, the department said the duty itself would
require services to
take appropriate and effective action to prevent the generation or distribution of fake nude material created without the consent of the person depicted
(3.14).
None of this is taken up in the committee view. The report records no evidence on
testing a model before release or on marking AI output. The word testing
appears,
but about testing the features of a service and the Commissioner’s covert testing
accounts, not models; we found no use of watermarking or provenance at all.
The codes and the validation clause
The majority treats leaving the codes behind as the point of the bill. In its view,
Rather than relying primarily on sector-specific industry codes and standards, the
proposed framework would impose a systemic duty on online service providers
, and the
bill would move the regulatory approach away from an over-reliance on industry codes
and towards a framework in which online service providers have clear obligations to
prevent harm
. It does not discuss what becomes of specific obligations in the codes
and standards when Schedule 3 repeals them, which the report notes will
coincide with the repeal of the Basic Online Safety Expectations, and industry codes
and standards
, 12 months after Royal Assent (paragraph 1.72).
The report does record a way to carry them over. The eSafety Commissioner submitted
that the Bill could prescribe specific minimum obligations to comply with the digital
duty of care under disallowable legislative instrument. Requirements in legislative
instrument could draw on the obligations in the Online Safety Codes and Standards in a
streamlined and simplified manner.
(2.24). No recommendation adopts it.
The validation clause, Schedule 1 items 90 and 91, which would treat registered codes and standards as never affected by a comment period shorter than the Act requires, is not mentioned anywhere in the report.
Where the Coalition and the Greens part company
The Coalition dissent, signed by Senators Sarah Henderson and Dean
Smith, calls the bill so deeply flawed and so ill-conceived that it should never have
seen the light of day
, and a blank cheque for political censorship
. It is the
only part of the report to make the codes an objection, and the only part that ties
them to AI chatbots. It notes that the codes and standards include safeguards against harmful
content generated by AI chatbots
, and says:
Shockingly, after a 12-month transition period, the bill abolishes all 17 codes and
standards in favour of a broadly defined Digital Duty of Care, with no guarantee that the
existing protections will be replicated.
It quotes eSafety’s submission:
As currently drafted, the Bill does not provide any mechanism to set specific
obligations. While eSafety will provide regulatory guidance that outlines what it sees as
best practice, this guidance is not enforceable.
Its single recommendation:
That the Online Safety Amendment (Digital Duty of Care) Bill 2026 not be passed.
The Australian Greens’ additional comments, from the deputy chair,
Senator Sarah Hanson-Young, back a duty but say these laws must have the teeth and
enforcement that will actually change industry behaviour
. Their six recommendations
are a private right of action with compensation, fines for breaching the duty set at
10 per cent of a company’s global revenue
, an external dispute resolution
system for social media platforms, addictive design in risk assessments, wagering listed
as a harm to children with age assurance at entry, and stronger data access for
researchers. They do not mention AI or the codes.
| Point | Committee majority | Coalition dissent | Greens comments |
|---|---|---|---|
| Pass the bill? | Yes, subject to considering Recommendations 1 and 3 | No | Amend it; no recommendation on passing it |
| Right to sue | Civil right of action (Rec 1) | Not recommended | Private right of action with compensation |
| AI services | Not mentioned in the view or recommendations | Notes AI chatbot safeguards in the codes | Not mentioned |
| Repeal of the 17 codes and standards | Welcomes the move away from codes | Opposes it | Not mentioned |
| Validation clause | Not mentioned | Not mentioned | Not mentioned |
What changed in the House
The report says the Government amendments were intended to make technical
corrections and changes and provide additional clarity for provisions that exclude
material from the scope of harms captured by the digital duty of care
, and the
crossbench amendments changed the definition of material harmful to children (section 25E)
and the user empowerment tools provisions. It describes no change to the AI paragraph, the
repeal or the validation clause. We have not yet read the amendments themselves.
What happens next
- 28 October 2026: the House of Representatives agreed to the bill with amendments.
- 20 November 2026: the committee reported.
- Next: the bill returns to the Senate, where the government can move amendments in response to Recommendations 1 and 3.
- 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.
Our view
The committee answered the questions most of its witnesses asked, about speech, small
business and redress, and a right to sue is a real addition. It did not answer the one that
matters most for AI services. Under the designated internet services standard, a high impact generative AI service must
test its models and differentiate their outputs; under the bill, that becomes something the
explanatory memorandum says a provider may
do. The report records eSafety’s own
suggestion for keeping specific obligations, minimum requirements in a disallowable
instrument drawing on the codes, and then passes over it. That instrument is the obvious
place for model testing to live, and the government can still propose it.
Recommendation 3 needs care for the same reason. A scope clause that confirms the bill
is aimed at digital communication platforms, such as social media platforms
and
lists what is out says nothing, on its face, about where a standalone chatbot sits. The
regulator’s own counsel treats AI chatbots as a sign that a service is not lowest
risk. Whatever amendment the government drafts should say plainly that section
25B(1)(h) stays in. And the validation clause has now been through the House and a Senate
inquiry, and neither the explanatory memorandum nor this report says which instrument it was written for.
Sources
Method note. Written substantially by AI and run by people, working from the 5 primary sources below. How AI Geek works.
- Online Safety Amendment (Digital Duty of Care) Bill 2026, report, Senate Environment and Communications Legislation Committee, November 2026, 135 pages (PDF, read 20 November 2026): paragraphs 1.1, 1.20 to 1.22, 1.71 and 1.72; 2.24, 2.40 to 2.42, 2.70, 2.142; 3.14; the committee view and Recommendations 1 to 4 (3.78 to 3.91); the Coalition Senators’ Dissenting Report (1.1, 1.3, 1.28, 1.71, 1.72, 1.135); the Australian Greens’ Additional Comments (1.6, 1.26 to 1.31). Shown in Exhibit A.
- Report contents page, Parliament of Australia (read 20 November 2026):
REPORT - November 2026
, chapters 1 to 3, committee view, dissenting report and additional comments. - Online Safety Amendment (Digital Duty of Care) Bill 2026, first reading, Parliament of Australia, 90 pages (read 16 October 2026): s 25B(1)(h); Sch 1 Part 5 items 90 and 91; Sch 3.
- 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 16 October 2026): the model testing and output differentiation requirements for a high impact generative AI DIS.
- Explanatory memorandum, Parliament of Australia, 105 pages (read 16 October 2026): the generative AI controls example under section 27, which a provider
may
use.
How we checked this. The report was downloaded from
ParlInfo on the afternoon of 20 November and read through its committee view, both minority
documents and the evidence chapters on scope, harms, risk assessment, fake nude material and
researcher access. We searched the full text for artificial intelligence, AI, chatbot,
generative, model, testing, watermark, provenance, label and valid; the findings above about
what the report does not say rest on that search. Every quotation is verbatim from the
report; the report itself quotes some witnesses from the proof Hansard, which can change
before the official record. The may
in our view is from the explanatory memorandum
we read on 16 October. We have not yet read the House amendments, and we have not asked the
committee, the department or eSafety about the validation clause or the model duties. The
“our view” section is opinion based on the documents cited.
Read the report differently? Tell us and we will check it against the document and log the outcome here.