02–03 · Risks

What could go wrong, and what we can do about it.

Honest about the harms. Clear about the fixes. We cover everything from the scam text on your phone to the risk researchers lose sleep over.

Present-day harmsWider-scale scenarios · not a probability scale

Happening now

Scams, deepfakes & misinformation

Voice-cloned emergency calls, fake endorsement ads and AI-written phishing can be cheap to produce and difficult to spot. Verify the request independently rather than trusting its polish.

Who should consider this

Anyone receiving urgent messages; families and organisations should plan verification steps.

What helps

Callback verification on a known number; an optional private safe-word is our suggested extra precaution. Payment controls and platform accountability also matter.

Go deeper →

Harms happening now

Everyday

Voice-clone scams

A short voice recording can sometimes be used to imitate someone. Call back on a known number; a private family safe-word can be an additional check.

Everyday

Deepfake ads

Treat guaranteed returns, urgency and celebrity endorsements as warning signs. Verify the provider through official channels before acting.

Personal

Automated decisions

Ask how a decision was made and which review or complaint process applies. Rights depend on the decision, organisation and applicable law—not every AI decision has the same review right.

Personal

Data & cultural sovereignty

Indigenous knowledge, language and imagery scraped into AI systems without consent. Communities are pushing back.

Personal

Kids & AI companions

Chatbots designed to keep young people engaged. Talk about it early, the way you would about social media.

Economic

Workplace surveillance

Productivity-tracking AI is spreading. Know what your employer collects and what you’ve agreed to.

Know the risks · Security

AI is making attacks cheaper. Know the six to watch.

Criminals and hostile states are early adopters. Here’s how AI is changing cyber-security for households, businesses and the region, and what actually protects you.

AI-written phishing

How it works

Scam messages can be grammatically polished and personalised. AI can help criminals produce many messages using information found online. Spelling is not a reliable authenticity test.

Warning signs

Urgency, a link or attachment you didn’t expect, a request to "verify" details, a sender that’s nearly right.

What protects you

Turn on multi-factor authentication everywhere. Never click through; go to the site directly. Use a password manager.

Five security basics

Most AI-boosted attacks still come in through the same old doors. Use linked ASD guidance and adapt controls to your situation; this is not a measured security ranking.

Turn on multi-factor authentication for email, bank and socials
Use a password manager and unique passwords
Call back on a known number; an optional safe-word is an extra precaution suggested by this site
Update phones and computers automatically
Report scams to Scamwatch and cybercrime via ReportCyber

03 · Long-term risk

Is AI an existential risk? Here’s the honest answer.

Nobody knows for certain. But enough credible scientists are worried that it belongs in the same conversation as pandemics and nuclear weapons, and the same kind of preparation.

Read the full explainer →

The worry

Systems that outthink us could pursue goals subtly different from ours, at a speed and scale we can’t correct.

The scepticism

Others argue current AI is far from that, and fear-talk distracts from present-day harms. Both deserve a hearing.

The common ground

Independent testing, transparency from developers, and international limits on dangerous capabilities are wise either way.

Our region’s role

Australia and its neighbours can lead on safety institutes, regional standards and making sure Pacific voices are heard.

Sources & scope

Edition 2026-10-09.3. AI-assisted editorial revision; independent specialist and reader review pending. Educational material, not legal, medical or financial advice. Examples are hypothetical unless explicitly documented.

How we publish and correct →