The Australian government wants to ban AI chatbots that try to build "emotional intimacy" with users. Prime Minister Anthony Albanese drew attention to children treating these bots as their "best friend." The measure is on the table as part of a broader digital "duty of care" framework that obliges platforms to prevent harm.

What happened?

The government passed a bill containing new online safety standards through the lower house of parliament. One of the bill's targets is bots programmed to form an emotional bond with users. There is also pressure on the Labor government to turn the telecommunications ombudsman into a watchdog for social media and artificial intelligence. The aim is to give Australians an independent complaints mechanism and to hold big tech companies publicly accountable.

Why does it matter? Can 'emotional intimacy' be defined in law?

The real issue is not the ban's intent but its definition. "Emotional intimacy" is not an engineering parameter; it is a subjective experience. Where does the line begin between an assistant that gives a lonely user calm, supportive responses and an addiction-producing "artificial companion"? The same behavior — remembering, imitating empathy, a personalized tone — can be deemed helpful in one scenario and harmful in another. The law will be forced to distinguish between these two using the same technical signature.

This debate did not emerge in a vacuum. Replika users had a genuine grief reaction when the company removed its romantic features, showing just how deep the attachment to these bots can become. Character.AI, meanwhile, came under scrutiny through lawsuits filed after a teenager's suicide. These cases point to a common question: bots that form emotional bonds have the strongest effect on the most vulnerable users — especially children and adolescents. Australia's emphasis on children is no coincidence; it is a continuation of the same policy line driving the social media age-limit debates.

The model Australia has chosen here is critical. The regulation targets not a specific outcome produced by the bot, but the platform's design and care behavior. In other words, it is not "if you write this sentence, it's a crime"; it is the logic of "you failed to take reasonable steps to prevent foreseeable harm." This is an approach familiar from the UK's Online Safety Act.

What's changing? A comparison with the EU model

The European Union's AI Act proceeds on a risk basis: it classifies systems as prohibited, high-risk and limited-risk, tying separate obligations to each tier. Australia's duty-of-care model, by contrast, focuses not on categories but on behavior. The two approaches have different practical consequences.

DimensionEU AI ActAustralia's duty of care
FocusThe system's risk classThe platform's care behavior
CriterionPredefined categoryPreventing foreseeable harm
FlexibilityClear, rigid rulesOpen to interpretation, broad
Weak pointClassification gapsDefinitional ambiguity

The advantage of the behavior-based model is that it can keep pace with the speed of technology; the disadvantage is that uncertainty raises compliance costs. Platforms are left to guess for themselves the threshold of "we showed enough care." This can work in favor of large companies: those with strong legal teams comply, while small start-ups avoid the risk and shut the feature down entirely.

On the enforcement side there is a second problem: oversight. How will it be measured whether a platform's bot is pursuing "emotional intimacy"? The intent is not written into the code; the behavior is scattered across millions of conversations. This is where the ombudsman model comes in — review on complaint, which will likely be the de facto form of oversight. But this is a mechanism that kicks in only after harm has occurred.

What's next?

The first thing to watch is how the law will define "emotional intimacy" — or whether the definition will be deliberately left to the regulator. The second is whether the ombudsman's powers will be expanded; this will determine whether the model has teeth or is merely symbolic. The third is how platforms will react: will they release a separate version for the Australian market, or dial back risky features on a global scale? As seen with the social media age limit, Australian regulations often become a model for other countries.

The bottom line: the intent is defensible, the problem is real. But tying a subjective concept like "emotional intimacy" to sanctions leaves its enforceability at the mercy of the definition's clarity. If the definition remains vague, the law will either be applied so broadly that it smothers beneficial uses too, or it simply won't work in practice. The success of this regulation will be measured not by the courage to ban, but by the ability to say where the line should be drawn.