AI-Generated Armed Police Image Sparks False Panic Claims at Sydney Shopping Centre

The specific AI tool used to generate the Hurstville image has not been identified in verified reporting; News24 (Sky News Australia) described it only as an "AI-generated photo," without naming a generator. ^[1] No photographer credit or media-organization sourcing accompanied the image as it spread.

A genuine police response was already under way because of the alarm test when the image appeared, which plausibly made it easier to believe. ^[1] Widely circulated claims about exactly what percentage of deepfakes are "indistinguishable" from genuine media vary across vendor marketing reports and could not be traced to a reliable primary source that actually supports the number commonly quoted — one specific figure checked for this dossier turned out to be a misattribution of an unrelated statistic in its own source report. That number has been removed rather than repeated.

For context on detection generally: Google's SynthID watermarking embeds invisible, pixel-level markers in images generated by Google's own AI tools. AAP FactCheck used Google's "About this image" tool to confirm that a separate fake image — purporting to show a wounded police officer from the Bondi Beach attack — carried a SynthID watermark identifying it as "Made with Google AI." ^[3] SynthID only applies to Google's own AI outputs, not to images from Midjourney, Stable Diffusion, FLUX or other generators, and because the Hurstville image's generator was never confirmed, no equivalent watermark check was possible for it.

The evidence for the "fabrication" verdict is: NSW Police's own statement that no evacuation or threat occurred [1]; the Facebook poster's own admission the photo was AI-generated [1]; the absence of any NSW Police record of a tactical deployment at Hurstville that day [1]; and the absence of any verifiable photographer or media credit for the image. [1]

Commercially available AI image generators can produce a photorealistic image in well under a minute, a widely documented capability of tools such as Midjourney, FLUX and DALL-E — meaning the cost of fabricating an image like the Hurstville one is trivial next to the cost of verifying and denying it. ^[1]

NSW Police's clarifying post, based on its own timestamp, went up at approximately 1:45pm the same Friday as the alarm — same-day, but still only after the image had already been shared, screenshotted and re-shared. ^{[2] [1]} This is not a failure unique to NSW Police: official verification necessarily takes longer than posting an image, because it requires confirming the scene status and coordinating a communications response before anything goes out under an agency's name. ^[1]

The broader pattern is well documented. A 2018 MIT study found that false news on then-Twitter reached people "up to six times faster" than accurate reporting and was retweeted roughly 70% more often. ^[11] By the time an agency's denial reaches the public, an AI image anchored to a real, developing event has typically already had its widest-seen moment.

Even a prompt, same-day police correction cannot outrun an AI image that launches the moment a real event starts developing — the asymmetry is structural, not a communications failure by any one agency.

The Hurstville incident is one entry in a documented, escalating pattern of AI imagery anchored to real-world public-safety or crisis events.

Pentagon Explosion — May 22, 2023. A fabricated photo depicting smoke near the Pentagon was shared by a since-suspended fake verified Twitter account calling itself "Bloomberg Feed." The S&P 500 dipped roughly 0.26% before quickly recovering once the image was shown to be fake. The Pentagon Force Protection Agency and Arlington Fire & EMS issued a joint on-record denial: "There is NO explosion or incident taking place at or near the Pentagon reservation, and there is no immediate danger or hazards to the public." ^[21]

Bondi Beach Terror Attack — December 14, 2025. Among several fabricated images, the most consequential depicted human rights lawyer Arsen Ostrovsky with fake blood being applied by a makeup artist, framed as "proof" the attack was staged. Ostrovsky, who had genuinely been grazed by a bullet, posted while being prepped for surgery: "I saw these images as I was being prepped to go into surgery today and will not dignify this sick campaign of lies and hate with a response." ^[9] Grok, X's in-house AI chatbot, told users the fake image was "consistent with a real photo"; a consumer AI-detection website separately returned a 0% "AI generated probability" on the same confirmed fake. Full Fact states plainly: "We do not find these types of AI detection websites reliable and do not use them in our fact checking of suspected AI content." ^[9] A separate fake article from the same event, according to Startup Daily, "cited fake quotes from figures like Prime Minister Anthony Albanese," and "incorrectly identified the ex-NSW Police commissioner Karen Webb as still in the role and at a press conference" — Mal Lanyon had in fact succeeded Webb, sworn in on October 1, 2025. ^{[5] [15] [16]}

Boulder County, Colorado — late 2025 to early 2026. AI-powered alert tools CrimeRadar and Nextdoor (via its Samdesk AI feature) sent inaccurate emergency notifications to residents, including a false "commercial blaze" report in Frederick on January 30 and a fabricated apartment-fire report in Longmont, after misreading radio dispatch traffic; a separate Nextdoor alert had previously reported a non-existent active shooter at a federal facility in Boulder. ^[4] This shows AI-driven public-safety misinformation operating through nominally official-seeming notification systems, not just social media.

University of Florida research published in February 2026 found AI programs "up to 97% accurate at detecting pictures of deepfake faces" in still images. But "the algorithms' performance declined sharply when it came to detecting deepfake videos. In those tests, programs performed at chance levels, while humans correctly identified real and fake videos about two-thirds of the time." ^[10]

The structural gap this illustrates: enterprise-grade detection accuracy, measured under controlled conditions, does not reliably translate to real-time consumer-platform performance — exactly the gap exposed by the 0% consumer-detector result and Grok's "consistent with a real photo" judgment during the Bondi attack. ^[9] As of March 2026, YouTube's expanded AI likeness-detection tool — announced March 10, 2026, three days before the Hurstville incident — covered only a pilot group of "government officials, political candidates, and journalists," not general public imagery. ^[14]

The cost of AI-fuelled false alarms is not abstract. NSW Police can charge a prescribed $1,600 fee under section 209 of the Police Act 1990 for responding to a repeat false security-alarm callout. ^[13] In the US, "in 2002, American police responded to 36 million alarm calls, costing $1.8 billion" — a figure specific to burglar-alarm callouts generally, not to AI-driven hoaxes, but illustrative of how quickly false-alarm costs compound at scale. ^[12]

NSW's Crimes Amendment (Intimate Image and Audio Material) Act 2025 came into effect on February 16, 2026 — less than a month before the Hurstville incident. ^[18] It criminalises creating, sharing or threatening to share sexually explicit deepfakes of identifiable real people, and covers AI-generated audio, carrying penalties of up to three years' imprisonment. ^[17]

That law applies only to sexually explicit material — material depicting "a person engaged in a sexual act," "a person with their private parts exposed," or "a person in a state of undress in a context that is sexual." ^[8] It does not address AI-generated images used to manufacture public panic or fabricate emergency scenarios; no NSW law has been specifically amended to cover that harm. ^[8]

Australia's eSafety Commissioner, Julie Inman Grant, was awarded a Public Service Medal in the 2026 Australia Day Honours "for her work to improve online safety after spearheading a world-first social media ban for people under 16." ^[19] Her most prominent 2026 regulatory work has centred on that under-16 social media ban rather than on AI-generated public-safety hoaxes specifically.

Covered (from Feb 16, 2026): AI-generated sexually explicit images/audio of real identifiable people, up to 3 years' imprisonment. Not covered: AI-generated fake police deployments or other imagery designed to trigger public panic. [8] [17] [18]

The EU AI Act requires providers of AI systems, including general-purpose systems, that generate synthetic audio, image, video or text to ensure outputs are marked in a machine-readable format and detectable as AI-generated — a binding obligation that applies from August 2, 2026. ^[25]

In the US, Senators Curtis, Schatz and Warner introduced the bipartisan AI Labeling Act (S.4915) on June 25, 2026, which would require visible and machine-readable disclosure of AI-generated image, video and audio content and push AI developers and major platforms to collaborate on content-provenance standards; its sponsors cite deepfake election imagery as one motivating example, though the bill itself is not limited to political content. ^[26]

Australia has no equivalent general requirement. The Australian AI Safety Institute — announced November 25, 2025 and funded at $29.8 million over four years, rising to an ongoing $7.9 million annually from 2029-30 — "has no powers of enforcement or information gathering," per Crikey, which also reports an active disagreement among experts over whether it should prioritise frontier-AI safety research, present-day harms like misinformation, or working through existing regulators. ^[20]

GNET (Global Network on Extremism and Technology) research published February 11, 2026, warned that "the confusion, uncertainty and manufactured 'crisis of knowing' created by deepfakes may itself propel radicalisation" — a dynamic that, applied to incidents like Hurstville, suggests AI images of tactical police operations could be weaponised for narratives about government overreach, beyond the immediate panic. ^[22]

The Future Policing Institute has been direct about the stakes: "Fake, viral photos or videos depicting horrific police use-of-force...could result in global demonstrations, many of which may result in significant violence." And: "If police leaders want the public to trust and believe them after fake images go viral, they must ensure the public trusts and believes them before the images are released." ^[7]

The image of armed tactical police purportedly deployed at Westfield Hurstville on March 13, 2026 is an AI-generated fabrication. No tactical deployment occurred. ^[1] The underlying alarm was a scheduled test; NSW Police confirmed no evacuations took place and there was no threat to the public, and posted a public clarification the same day. ^{[1] [2]} No law in Australia specifically criminalises generating or sharing an AI image to manufacture public panic, as of March 2026. ^[8]

The Westfield Hurstville AI image incident is one entry in a documented pattern — from the Pentagon in 2023, to Bondi Beach in December 2025, to Hurstville in March 2026 — of AI imagery anchored to real-world events outrunning official correction. ^{[21] [9] [1]}

AI image generation takes seconds; official verification and public communication take longer, because they require confirming facts before an agency puts its name on a statement. ^[1] Consumer-accessible AI detection tools have returned false negatives on confirmed fabrications, and even X's own chatbot misjudged a fake image as "consistent with a real photo" during the Bondi attack. ^[9] University of Florida research shows detection accuracy on video drops to chance levels even as still-image detection reaches 97% under controlled conditions — underscoring that today's detection tools are not yet reliable enough, deployed broadly enough, or fast enough to intercept this kind of hoax in real time. ^[10]

Australia's regulatory response is not yet calibrated to this specific harm. The NSW law that took effect weeks before Hurstville covers only sexually explicit imagery. ^{[8] [18]} The AI Safety Institute has no enforcement powers over AI-generated emergency misinformation. ^[20]

Until AI provenance labeling, faster verified-source amplification, and detection tools that work on real-world video exist at platform scale in Australia, the open question is not whether another AI-generated public-safety hoax will hit an Australian shopping centre, suburb or stadium — it is how quickly the next official denial can arrive, and how much has already spread before it does.