1. Messages that sound like someone you know
What it is: Scams that use AI to copy a voice, a face or a writing style, so a request seems to come from someone you trust.
What it looks like for a member: A phone call in a grandchild's voice asking for money today, or an email in the treasurer's style asking for a bank detail to be changed.
Why it matters, and what to do: A voice can now be copied from a few seconds of recorded audio. A 2024 UK survey for Starling Bank found that 28% of adults said they had been the target of an AI voice-cloning scam in the past year, and 46% had never heard of such scams. In New Zealand, scams and fraud were the category most often reported to the National Cyber Security Centre in the second quarter of 2025, and direct losses reported across all incident types reached 12.4 million New Zealand dollars in the third quarter of 2025, about half of it from falsified payment transfers. The oldest check we have, recognising a voice or a face, no longer settles the matter on its own. A careful community agrees a second channel for anything involving money or access: a call back on a known number, a safe phrase within a family, a rule that bank details change only after two people confirm.
2. Leaked passwords and member details
What it is: A security breach at any service your members use, which exposes passwords and personal details that are then tried elsewhere.
What it looks like for a member: A login that works for a stranger because the same password was used on a shopping site that was breached last year.
Why it matters, and what to do: New Zealand's Privacy Commissioner received 1,093 privacy breach notifications in the 2024/25 year, up 27% on the year before. A misdirected email, a lost device or a reused password can each expose member details. A community's record is only as safe as the weakest login into it. A careful community turns on two-step sign-in, uses a different password for each service, keeps member lists to the minimum needed, and chooses platforms that tell members plainly what is held and where.
3. Confident AI mistakes
What it is: An AI assistant states something false with the same fluency as something true. The trade calls it hallucination.
What it looks like for a member: A meeting summary that invents a motion nobody moved, or a quoted rule that does not exist in the constitution.
Why it matters, and what to do: This is measured, not anecdotal. A 2025 study in the Journal of Empirical Legal Studies found that AI legal research tools from two leading legal publishers, marketed as avoiding it, each produced false information more than 17% of the time. A fluent wrong answer is harder to catch than a hesitant one. A careful community treats AI output as a draft, keeps the source document beside the summary, and makes a person responsible for anything that becomes a decision or a record.
4. Pictures you cannot vouch for
What it is: Images, audio and video that are generated or altered well enough that the eye cannot tell.
What it looks like for a member: A photo of a councillor at an event they never attended, passed around before anyone checks.
Why it matters, and what to do: A community's shared memory depends on its pictures and recordings meaning what they appear to mean. The practical answer is not better eyesight but provenance: who published this, when, and can the original be found? A careful community keeps its own copies of its own media, records when they were made, and treats anything arriving without a source as unconfirmed.
5. Unfair automated decisions
What it is: Software that sorts, ranks or screens people learns patterns from old data, and those patterns can disadvantage some groups.
What it looks like for a member: A volunteer-screening tool that keeps rejecting applicants from one suburb, and nobody can say why.
Why it matters, and what to do: NIST, the US standards body, describes bias in AI as endemic across technology processes and able to cause harm regardless of intent. A decision nobody can explain is a decision nobody can appeal. A careful community keeps a person making the call, writes down the rules the software applies so they can be read, and reviews outcomes by group from time to time to see whether the pattern is fair.
6. AI agents that overstep
What it is: An AI agent given tools (email, payments, calendars, files) does more than anyone intended, because it was given too much authority, too many permissions or too little oversight.
What it looks like for a member: An assistant asked to “tidy up the mailing list” removes members, or one asked to “pay the invoice” pays the wrong one.
Why it matters, and what to do: The OWASP security community lists excessive agency among the ten main risks of LLM and generative-AI applications, and names its causes as excessive functionality, excessive permissions and excessive autonomy. Among its remedies: give the agent the least access it needs, require a person to approve high-impact actions, and let the system that does the work check the permission rather than trust the agent's say-so. An agent acts at speed and leaves no natural pause for a second look. A careful community writes down what each agent may do, keeps the risky actions behind a second person, and keeps a record of every decision an agent asked for.
7. Foreign law
What it is: Data is held by companies, and companies answer to the law of the country that owns them, not only the country where the servers stand.
What it looks like for a member: A platform “hosted in New Zealand” whose parent company can be legally required to hand over the data from abroad.
Why it matters, and what to do: The clearest example is the US CLOUD Act of 2018, under which a US provider can be required by US legal process to produce data in its possession, custody or control, wherever it is stored. The point is not that this happens often to community groups. It is that the location of the servers does not settle the question most people think it settles. A careful community asks a provider who owns the company, which laws reach it, and who holds the keys. The seven questions on our page about what “sovereign” has to mean are a short way to ask.
8. Locks that will not hold forever
What it is: Today's encryption and digital signatures rest on mathematics that a large enough quantum computer could undo. Records encrypted today could be stored and read later.
What it looks like for a member: Nothing today. The question is which records must still be private in fifteen years.
Why it matters, and what to do: NIST published its first finished post-quantum standards in August 2024 and noted that some experts expect a capable machine within a decade. The UK's National Cyber Security Centre advises organisations to have a migration plan by 2028 and to finish by 2035, including a record of the data they hold, its expected lifetime and its value to an adversary. This is a horizon, not an alarm. A long-lived record, a whakapapa, a land file, a medical history, should be protected for its own lifetime, not for the lifetime of today's algorithm. A careful community asks providers for their plan, keeps records that must last in formats that can be re-signed, and prefers signatures that can be checked without the provider.