What happened this week

On 20 November 2026 the Australian Senate's Environment and Communications Legislation Committee tabled its report on the Online Safety Amendment (Digital Duty of Care) Bill 2026. The committee drew on 182 submissions and two public hearings. It recommended the Senate pass the bill "subject to consideration" of its recommendations, according to NewsWire's report. Coalition senators lodged a dissenting report and the Greens added comments of their own. The bill is listed for debate in the Senate's final sitting week of 2026, starting Monday 23 November.

Most coverage has treated the bill as another move against social media platforms. For people who build with AI, the more important detail is that the bill names AI generation services as a category of online service in their own right. A support bot, companion app or image generator that Australians can reach would carry the same duty as a social network.

How the duty works

The bill amends Australia's Online Safety Act 2021. A person responsible for an online service "must take appropriate action to ensure a safe online environment", which includes mitigating the risks found in that person's own risk assessments. Analysis by Webcoda sets out the parts that matter most for builders:

  • Scope: proposed section 25B(1) lists nine kinds of online service. Paragraph (h) covers "a service that generates material using artificial intelligence". The September exposure draft only caught AI services whose output could be shared through another listed service. The House version dropped that condition, so a chatbot on its own can now fall in scope.
  • No size threshold: the bill has no minimum user count, turnover or small-business threshold.
  • Who is responsible: section 25B(2) puts the duty on the provider and also on the person with the greatest day-to-day control of the service. If your bot runs on a vendor's platform, both you and the vendor could be responsible.
  • Proportionality: section 26A says the action taken must be proportionate to the risk, and the provider can't count cost as a reason to do less.
  • Risk assessments: section 27 requires written assessments at least once a year and before any significant change that could add risk. Each assessment rates how likely and how severe each foreseeable harm is, and providers must keep them for six years.

CyberPeace's summary adds that eSafety, the Australian online safety regulator, can ask for an assessment and expect it within 30 days. The harms sit in tiers. The top tier covers everyone and includes content promoting suicide or self-harm. The second tier covers material harmful to children under 18 and applies whoever a service's intended audience is.

During the House stage, independent MP Kate Chaney added paragraph 25E(1)(da). It covers material or conduct that "simulates emotional intimacy" and fosters a child's dependence on the service in place of human relationships. Johnson Winter Slattery read it as aimed at AI companion-style features. The amendment protects children only, and AI companions used by adults fall outside it. Once the duty starts, persona bots, character chat and memory features that build rapport with users all belong in a risk assessment.

Why NZ teams can't treat it as an Australian problem

New Zealand has no equivalent law. Even so, plenty of Auckland products serve Australian customers, and the bill makes no exception for foreign providers. Johnson Winter Slattery reports that the House version removed a ministerial exemption for services with minimal Australian use. The minister keeps a power to exempt services that pose little risk to users. The age assurance obligations in section 27D are written around users "in Australia". On a plain reading, a Wellington SaaS company with an AI help desk and Australian users would have to meet the duty for those users. We haven't seen eSafety guidance on how the regulator will enforce the duty against offshore firms.

Penalties are expressed in penalty units. Several obligations, including the risk assessment rules in section 27B, carry a maximum of 60,000 units, while the minister's media release puts the top penalty for a company at up to A$109.2 million. The final penalty settings may change. The Greens want fines set at 10 per cent of global revenue, and the government needs Greens votes to pass the bill this year.

What the committee changed, and what's still open

According to NewsWire's account, the committee recommended these changes:

  • an amendment so the duty doesn't target "low-risk" websites such as small businesses and religious education sites
  • a civil right of action, which would let individuals sue over breaches of the duty
  • a statutory review of how well the scheme deals with online hate and vilification

The low-risk recommendation is the one AI builders should watch most closely. The report doesn't define "low-risk" in any coverage we've read, and the government hasn't said whether it'll adopt the change. Streamline's analysis suggests assuming a public-facing chatbot stays in scope until the final text says otherwise. Under the same reading, a tool used only by staff is less likely to be caught. A civil right of action would make dated risk assessments far more valuable, because those records would become a firm's main evidence that it took appropriate action.

The Coalition's dissent called the bill "so deeply flawed and so ill-conceived that it should never have seen the light of day". The dissent argues the burden falls on small businesses, charities and other low-risk services. UTS law professor Derek Wilding criticised the bill from another direction. Wilding argues that defining a safe environment by an exhaustive list of harms falls short of a design-based approach.

The bill also affects rules already in force. Schedule 3 would repeal Australia's online safety industry codes and standards 12 months after Royal Assent. Those codes include the ones covering generative AI. Until then the codes stay in force, so AI companion providers serving Australians still have to meet the existing code obligations.

Timing and cost

The duty would start 12 months after Royal Assent. If the Senate passes the bill before Christmas, the earliest start is around the end of 2027. The bill sets no compliance budget. The cost falls mostly on engineering and documentation time: threat modelling for each AI feature, records of guardrail testing, a channel for user complaints and a log of how each complaint was handled. For child-facing or companion products, age assurance is a further cost, and the bill expects that cost to stay proportionate to the risk.

What a builder can do now

None of these steps depends on how the Senate votes, and each one also helps with New Zealand privacy and consumer law:

  1. List every AI feature that people outside your organisation can reach. Note whether Australians, and anyone under 18, can use each one.
  2. For each feature, write a short risk assessment against the bill's harm tiers. Include self-harm prompts, sexual content and dependence-forming persona behaviour. Date it and keep it.
  3. Test your guardrails against those harms, store the results, and run the tests again after every model or prompt change. Section 27 asks for a fresh assessment before significant changes.
  4. Work out with your model or platform vendor who holds day-to-day control. Under section 25B(2) both of you may be responsible.
  5. Check the bill's final text after next week's Senate debate before you commit to an age assurance approach.

If you're building safety evaluation tooling for chatbots, other members would like to hear about it at asai.nz/projects.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member