From London: As Australians slept this week, the western world's consensus on artificial intelligence crystallised around a single uncomfortable truth: the technology cannot yet be safely deployed at scale without rigorous human oversight. Washington courts blocked government overreach. Wikipedia volunteers rejected AI content. OpenAI's own advisors warned of mental health risks. Yet in the early hours of 27 March, the Australian Public Service continued rolling out Microsoft Copilot across Treasury, defence, and policy departments with minimal safety architecture in place.
The disconnect is striking. On 26 March, a federal judge in San Francisco temporarily blocked the Pentagon's attempt to blacklist the AI company Anthropic, with Judge Rita Lin ruling the agency's demand for unfettered access to the company's models amounted to "classic First Amendment retaliation." Anthropic had resisted precisely because it wanted assurance its technology would not be used for autonomous weapons or domestic mass surveillance. The court sided with the company's right to refuse.
Two days earlier, on 24 March, Wikipedia's volunteer community voted 44 to 2 to ban AI-generated article content entirely. The reason was not abstract: large language models hallucinate, and when inaccurate text enters Wikipedia, it gets scraped by AI companies and re-enters future training data, poisoning models with false information. The burden of cleaning up AI-generated garbage has become unsustainable for volunteer editors.
On 7 March, OpenAI indefinitely shelved its planned adult mode for ChatGPT after its entire wellness advisory council unanimously opposed it. The advisors warned that sexually explicit interactions could foster unhealthy emotional attachments with serious mental health consequences. Internal testing revealed a 10 per cent error rate in age verification. One advisor described the risk as turning ChatGPT into a "sexy suicide coach."
The same month, Forrester researchers found that despite massive corporate investment in AI tools, workplace adoption has stalled. The culprit was not technology failure but human resistance: 43 per cent of employees fear job loss to automation, and only half of companies offer AI training. Employees are being handed powerful tools without understanding, context, or support.
Now consider what this means for government deployment. Federal researchers have identified a concrete risk: AI hallucinations in government data streams could skew policy analysis and trigger poor decisions affecting national security, public health, and economic stability. Yet Australian government agencies are deploying Copilot to draft policy content and summarise sensitive information.
The timing is not coincidental. Australia's government AI deployment occurs during the exact moment when democratic systems elsewhere are expressing doubt about AI's readiness for high-stakes decisions. The European Union is fining companies up to €35 million for AI governance failures. Colorado's legislature delayed AI impact assessments until June, acknowledging the need for rigour. American courts are defending companies' rights to refuse government demands for unfettered access.
Australia's approach differs. Government transparency requirements for AI begin 15 June 2026, with full obligations in December. This means the Australian Public Service is deploying AI for policy work now, months before any safety framework takes effect. By contrast, private sector transparency requirements begin the same month. The government gets a six-month head start with no equivalent safety review.
The risk is not hypothetical. When Wikipedia volunteers discovered that AI-generated nonsense was entering the encyclopaedia at scale, the platform chose to reject the technology entirely rather than manage the cleanup burden. When OpenAI's advisors found mental health risks, the company chose to shelve the product. When American courts saw government overreach on AI, judges chose to side with the company resisting unfettered access.
Australia, by contrast, is accelerating. The governance gap is real and widening. For policymakers in Canberra, the question is whether the recent lessons from Wikipedia, OpenAI, courts, and workplaces amount to warnings worth heeding, or whether Australia intends to chart a different course.