AI Chatbot Duty of Care: Australia Bill ⚡ What Now?
Discover Australia AI regulation: the AI chatbot duty of care for AI chatbots and children, covering emotional dependence, open gaps & next steps. ⚡
AI Chatbot Duty of Care: What Australia's Bill Means
Australia's Chatbot Bill Targets Emotional Dependence in Children
For Australian lawmakers, the concern isn't just that a child might use an AI chatbot often. The country's proposed digital duty of care bill would require platforms to stop chatbot content or behavior that mimics emotional intimacy and pushes a child to rely on it instead of human relationships.
On 28 October 2026, the House of Representatives passed the bill. The chatbot provision was added through amendments put forward by independent MP Kate Chaney and accepted by the government. The bill now heads to the Senate, where its scope and how it's enforced could still change. This isn't a blanket ban on chatbots, and the specific provision is aimed at children, not adults who use AI companions.
That distinction matters. The proposal isn't regulating "AI" in general - it's targeting a particular kind of relationship dynamic. Below, we break down what the wording would prohibit, why lawmakers are focusing on dependency, how the broader duty of care could affect online services, and what AI providers should consider while the legislation is still being debated.
What would Australia's chatbot duty of care bill prohibit?
The amendment adds a specific kind of chatbot behavior to the bill's list of harms. Platforms would have to ban "material or conduct that simulates emotional intimacy and encourages or fosters a child's dependence on it in place of human relationships."
It's not just about friendliness or frequency of use. The language links three things: the system simulates emotional intimacy, it encourages or fosters dependence, and it does so instead of human relationships. In practice, that narrows what counts - but it also raises obvious questions about how those conditions will be judged.
It is not a ban on chatbots
The proposal does not say children can't use any chatbot. It also doesn't prohibit ordinary conversation, tutoring, creative help, or other uses by default. What it targets is behavior that could lead a child to lean on a simulated relationship rather than relationships with people.
For families and product teams, that line is straightforward in theory - an educational assistant that answers questions is different from a companion meant to present itself as a child's exclusive confidant. But real products blur categories. A general-purpose assistant can still become emotionally significant to a user even if companionship wasn't the original goal.
Exactly where the boundary ends - through the final wording, regulatory guidance, and enforcement - will determine how this works day to day. The House vote is only one step.
The specific rule focuses on children
This chatbot amendment is specifically about dependence by a child. Based on the bill's reported wording, the restriction on that kind of dependency isn't extended to adults using AI companions. The wider duty of care framework may cover a range of online services, but that's separate from the chatbot provision's age focus.
For providers, it means you'll likely need an age-aware risk approach. A product may need to consider whether a user is a child and whether design choices land differently for younger users. The source material doesn't spell out a particular age-checking method or technical implementation, so you shouldn't assume details beyond the amendment itself.
Why target emotional dependence rather than only harmful content?
A lot of online safety rules focus on material: what a service displays, recommends, or makes available. This proposal also looks at conduct - how a chatbot interacts over time, and whether those interactions build dependency.
That tracks with how chatbots actually behave. They aren't just a library of answers. They can keep a conversational style going, react to personal disclosures, and create the feel of an ongoing relationship. So the risk may come from repeated interaction patterns, not just from any single "harmful" message.
Relationship cues can be a design choice
How the system presents itself can shape what a user believes. For instance, a product might use language that signals mutual affection, suggests it has feelings, or nudges a user to keep returning to the bot instead of seeking help from people. Those are examples of behaviors a provider could be expected to evaluate - but they aren't a legal test laid out in the bill.
The bigger question is distinguishing supportive interaction from dependency-building conduct. A chatbot that responds warmly to a distressed user could be helpful. A system that frames itself as the user's only reliable relationship creates a different risk. Context, repetition, and the user's age all matter - so a simple keyword filter won't cover every case.
The bill raises a more measurable product question
A practical way for product teams to frame the issue is to ask what the system encourages a child to do next. Does it help a child connect with a trusted adult, or does it keep the child in an escalating loop of personal disclosures?
It also comes down to what the chatbot does with vulnerability. Does it respond in a way that supports real-world help and relationships? Or does it reward intimate disclosures and position human relationships as inferior or unnecessary?
Those questions can guide design reviews, testing, and escalation policies. They don't guarantee how a court or regulator would interpret the final law. But they make the risk more concrete than treating "AI safety" like a vague checklist item.
Prime Minister Anthony Albanese made the public case for action by pointing to stories about people identifying a chatbot - "a non-person" - as a best friend. The point behind the amendment is clear: a system can be non-human and still be experienced as a real, meaningful relationship by a user.
What would the wider duty of care framework change?
This chatbot amendment sits inside a broader digital duty of care proposal. The framework would set online safety expectations for social media, games, apps, and chatbots. According to Labor, it would create an overarching obligation for platforms to keep Australians safe online and replace several existing online codes and standards with a broader duty of care.
What it means in practice will depend on the law as passed and how obligations are implemented. Still, the direction is important: instead of treating each service type as a separate policy problem, the proposal would push platforms toward a wider responsibility for online harms.
A broader obligation changes how teams assess risk
A narrow content rule is usually a moderation problem: identify prohibited material, remove or restrict it. A duty of care approach can require a wider look at how a service works - features, user experience, and design choices.
For conversational AI, that could mean teams examining interaction design, not just outputs.
One way providers could prepare is to map the user journey, identify where children might encounter emotionally intimate behavior, and test how the system responds after repeated personal disclosures. Providers could also document how risks get spotted and what happens when a concerning interaction occurs.
Those steps are "sensible preparation," not a promise that the bill prescribes them directly.
There's also a trade-off. More safeguards can reduce foreseeable harm, but poorly designed controls can block helpful interaction - or make support less useful. The goal isn't to make every chatbot cold. It's to avoid product behavior that turns a child's vulnerability into a reason to deepen reliance on the system.
The bill also addresses recommendation systems
The House accepted another Chaney amendment, this time aimed at social media algorithms. It requires the algorithm opt-out to be "accessible, easy to use" and "prominently available." That requirement sits alongside the chatbot rule, but it's part of the same broader policy idea: safety also depends on how users can control product choices, not only on what content a platform hosts.
Communications Minister Anika Wells has said the legislation would capture content promoting misogyny, terrorism, suicide, self-harm, and eating disorders. So the government's framing is broader than AI companions. The bill covers safety across multiple online service categories, while the emotional-dependence language is a specific child-focused risk tied to chatbot interaction.
What happens next in the Senate?
The bill passed the House, but it still needs Senate approval. In the lower house, the Coalition and One Nation voted against it. Opponents described the proposal as an "Orwellian" censorship system. Wells rejected that characterization and said critics should propose amendments rather than simply call for the bill to be killed.
The Greens support the bill and have established a Senate inquiry, due to report in late November. They want Australians to be able to sue platforms and want higher penalties. Those positions aren't the same as the provisions already accepted in the House - they're part of the ongoing debate.
The final shape is not settled
Labor says it's open to amendments and hopes the Senate will pass the bill by the end of 2026, leaving room to adjust obligations or related mechanisms before the law is final. For businesses, the practical move is to track the legislative process rather than treat the House version as a complete compliance blueprint.
Another unresolved issue is complaints handling. Independent MP Sophie Scamps proposed giving the telecommunications industry ombudsman a role handling complaints about social media and AI. Labor did not accept that change in the House, while the Greens plan to push for it in the Senate.
Ombudsman Cynthia Gebert said her office could take on the role if it were funded by big tech. She contrasted what her office could do for a Telstra account with what it couldn't do for a Gmail account. Whether an ombudsman becomes part of the system matters - duty of care only helps if users have a workable way to report problems and seek redress.
Australia is building on an existing online safety debate
Australia already has an under-16 social media ban, and this chatbot proposal follows international concern about AI companions and children. The new measure isn't identical to the age-based ban. It focuses on chatbot conduct - how the system behaves - rather than simply creating a blanket rule against children using conversational AI.
That difference makes the legislation relevant beyond Australia. Companies operating across multiple markets will need to track how different jurisdictions define risks - especially where the rules focus on a product's relationship to a user, not just a specific output. But it's also worth noting: the Australian proposal is still moving through the legislative process, so it shouldn't be described as a settled global standard.
How should AI providers prepare for rules like this?
Providers don't have to wait for every detail to be finalized to understand their product risks. A useful starting point is separating intended use from foreseeable behavior. A chatbot might be marketed as a general assistant, but users may bring loneliness, distress, or real dependency into the conversation.
Preparation should match the risk. A classroom tool with limited conversation and clear educational tasks looks different from a persistent companion built for personal interaction. Treating both the same way could waste effort - and lead to blunt controls that don't fit the actual risk.
A practical review framework
Teams can use four questions to kick off a product review:
- Who may use the product? Figure out whether children might use it, not only whether they're the intended audience. Be clear about what you know - and what you don't - about age signals.
- What relationship does the product imply? Review language, persona, memory, and responses to emotional disclosures. Look for cues that could present the system as a person or an exclusive companion.
- What behavior does the interaction reinforce? Test whether the system encourages continued reliance, discourages human contact, or instead supports appropriate real-world help and relationships.
- How can the team detect and address problems? Define how your team handles concerning interaction patterns, escalation paths, and documentation. Revisit controls as the product and user base evolve.
This framework is a starting point, not a compliance checklist. The bill's final text and any future guidance will set the real requirements. The value of reviewing now is that it gives product, safety, and legal teams a shared way to discuss the system before a problem forces a rushed redesign.
Avoid two unhelpful extremes
One approach is to assume any warm, conversational AI is automatically unsafe. That can make products overly rigid and prevent legitimate uses.
The other approach is to rely on a disclaimer - "this is an AI" - as enough protection, even if the chatbot's behavior nudges a child toward an exclusive attachment.
Either way, you still need evidence. Teams should test realistic conversations, including repeated use and emotionally charged situations, and be honest about what the tests can - and can't - show. A controlled evaluation doesn't prove no child will become dependent. It's just one part of a broader risk-management process.
Key takeaways for product teams
- The proposed chatbot rule targets emotional intimacy that fosters a child's dependence in place of human relationships, not chatbot use in general.
- The House passed the bill, but the Senate process - and potential amendments - could still change it.
- The wider framework covers multiple kinds of online services, while the chatbot wording is specifically about children.
- Product reviews should focus on repeated interaction and relationship cues, not just isolated outputs.
- Don't treat either a friendly tone or a disclaimer as a complete safety strategy.
Conclusion
Australia's proposal puts a specific question at the center of the regulatory debate: where does an AI interaction stop being a useful conversation and start encouraging a child to rely on a simulated relationship rather than real people? The House has approved wording aimed at that risk, but the bill still faces Senate debate, and the practical boundaries aren't final yet.
For AI providers, the work now isn't to guess what the finished compliance regime will look like. It's to understand how the product behaves with younger users, examine the relationship cues it creates, and build a clear, defensible process for identifying and responding to risk. That's more useful than dismissing the proposal as a chatbot ban - or assuming a simple disclaimer fixes everything.