AI safety
How AI is used to deceive, impersonate and collect data, and what you can do about it. Every article cites its sources and ends with steps you can take.
AI makes scams cheaper to run and harder to spot: cloned voices, fake video calls and convincing messages. A voice or face is no longer proof of identity. Agree a family code word, hang up and call back on a number you already trust, and never share private data with a chatbot you would not post publicly.
For a worked example, read our guide to AI voice clone scams.
Scammers use AI-generated images, voices and videos of well-known people to make investment schemes look endorsed. The FBI lists fake promotional videos and celebrity images as tools of investment fraud, and the FTC says celebrity endorsements are easy to fake and prove nothing. Check the investment independently and never trust an ad on its own.
If someone makes a fake nude of you with AI, it is abuse and you are not at fault. Save evidence, then ask each platform to remove it. In the US, the Take It Down Act, enforced by the FTC since 19 May 2026, requires covered platforms to remove reported intimate images, including AI-made ones, and known identical copies within 48 hours.
CISA says generative AI mainly amplifies existing election risks such as false voting information, impersonation and influence operations at lower cost. Voters should confirm voting details on official election sites, treat shocking clips with caution, and know that the EU AI Act requires deployers to label deepfakes, though bad actors will not comply.
Deepfake video call fraud uses AI-generated faces and voices to impersonate executives or relatives in a live call, then asks for an urgent payment. In the Arup case, a Hong Kong employee sent HK$200 million (about US$25.6 million) after a call with fake colleagues. The fix is procedural: verify by a second channel you chose yourself.
Voice cloning software can imitate a person from a short audio clip, such as one posted online. You cannot fully prevent copying, so the best protection is not relying on a voice as proof: agree a family safe word, call back on known numbers, and avoid voiceprint-only security at your bank.
Visual tells such as distorted hands, odd shadows and unnatural movement can reveal some AI-generated media, but they are unreliable and fading. Provenance labels such as C2PA Content Credentials can confirm origin when present, though a missing label proves nothing. Checking the source, context and a second channel is more dependable.
General AI chatbots can be useful for explaining a diagnosis or preparing questions, but they should not be your only source for a health decision. The World Health Organization warns that these systems can produce false, inaccurate, biased or incomplete statements. Treat answers as a starting point and confirm anything that affects your care with a clinician or pharmacist.
Many chatbots now keep memories across conversations to personalise answers. Anthropic says Claude saves topics as you chat, excludes sensitive subjects by default, and lets you view, edit, pause or reset memory. Deleting a chat does not remove memories made from it, so check the memory settings as well as your history.
Most AI chat services have a setting that controls whether your conversations help train their models, but it rarely covers everything. Google says Gemini chats with Keep Activity off are not used to train AI models but are still kept for 72 hours. Anthropic says consumer chats are used only if you opt in to its Model Improvement setting, with exceptions such as safety review.
AI smart glasses put a camera and microphone at eye level, and recordings shared with an AI assistant may be processed or reviewed beyond the device. In March 2026 the UK Information Commissioner's Office said reports about human review of Meta glasses footage were concerning and that it would write to Meta. Wearers should check settings and bystanders should know they have limited control.
All three major assistants let you limit training use of your chats and delete history: ChatGPT via Data Controls and Temporary Chat, Claude via the Model Improvement setting and incognito chats, Gemini via Keep Activity. Deleted chats can still linger for weeks, and settings change, so check each vendor's current help page.
AI chatbots sometimes state false things confidently, including made-up court cases, sources and software packages. The cause is how they work: they predict plausible text, not verified facts. Check any legal, medical or financial claim against the original source, and never install a package name an AI suggested without confirming it exists.
Do not type anything into a public AI chatbot that you would not want a stranger to read. Providers store conversations, some are reviewed by humans, and shared links or breaches can expose them. Keep out ID numbers, passwords, financial details, medical records, other people's data and work secrets.
An AI browser agent can navigate sites, fill forms and click buttons for you, and Anthropic says every web page it visits is a potential attack vector. Hidden instructions on a page can steer it, and no browser agent is immune. Limit what it can reach, stay logged out of sensitive accounts, and approve payments and sends yourself.
AI coding assistants can produce code that looks correct but is not secure, and GitHub's own documentation says users are responsible for reviewing and validating suggestions. OWASP warns that model output must never reach a shell or interpreter unchecked, and that models can invent package names an attacker may register. Review, test and sandbox AI-written code like a stranger's.
The Model Context Protocol (MCP) is a standard way to connect AI assistants to tools and data. Its own security guidance warns about malicious local servers, stolen or over-broad access tokens and hijacked sessions, and the UK NCSC says an AI that processes untrusted content should not hold privileged tool access. Connect only trusted servers, grant the narrowest access, and keep approvals on.
Give an AI agent the least access it needs, keep it away from your main email, saved cards and password manager, and approve every payment or message yourself. Hidden instructions in web pages and emails can steer an agent, and the UK NCSC says this may never be fully fixed.
Prompt injection is when instructions hidden in content an AI reads, such as a web page, email or document, make it act against its user. OWASP lists it first among LLM risks, and the UK NCSC says it may never be fully fixed. The safest approach is to limit what your AI assistant can do and approve risky actions yourself.
A friendly AI chat or a polished phone agent is not proof that you have reached a real company. The FTC warns that scammers pay to place fake customer service numbers in search results and that top search results are an unreliable source for contact details. Find the number on your card, statement or the company's own site, and never pay in gift cards, wire transfers or by sharing passwords.
Hiring fraud now runs in both directions. The FBI warned in 2022 that applicants were using deepfakes and stolen personal information to win remote jobs, and in October 2026 that North Korean IT workers use facilitators to appear on camera and take interviews. Fake employers also use job offers to steal money and identities. Verify identity live and never pay to get a job.
Romance scammers use AI to write messages faster, translate without errors, generate profile photos and sometimes automate chat. The FBI says this lets them reach more people with believable content. The warning signs have not changed: fast declarations of love, refusal to meet, and eventually a request for money or crypto.
On February 8, 2024 the FCC ruled that calls made with AI-generated voices are 'artificial' under the Telephone Consumer Protection Act, so the same consent rules apply. The ruling gives enforcers more tools but does not stop scam calls from criminals who ignore the law. Treat any unexpected call as unverified, hang up and call back on a number you look up yourself.
FinCEN reported in November 2024 that criminals use generative AI to alter or create ID images and combine them with stolen or fake personal data to get past identity checks. Banks respond with re-reviews, live checks and detection tools, and NIST's 2025 guidelines add controls for deepfakes and injection attacks. Your best protection is to share ID images sparingly.
Fake reviews can now be written in bulk by software, so a page full of glowing five-star comments proves little. The US FTC bans AI-generated reviews that pretend to come from real customers, and UK rules treat bot-generated reviews as banned too. Judge a shop by checks you control: payment method, company details, independent reviews and how it treats refunds.
Fake AI apps and extensions copy the name and look of popular assistants, then collect your chats, browsing or contacts. Microsoft and OX Security reported lookalike Chrome extensions that harvested ChatGPT and DeepSeek conversations, with reporting citing about 900,000 installs. Install AI tools only from the vendor's own link, and remove any extension you do not recognise.
Scammers use AI chatbots to write fluent messages, translate them into many languages, build fake websites and generate fake photos. Polished grammar no longer proves a message is real. What still exposes a scam is pressure, an unusual payment request and a link or number you did not look up yourself.
Generative AI tools can help pupils and teachers, but England's Department for Education says evidence on pupils using them directly is still emerging, and that pupils should use them only with safeguards such as supervision, filtering and monitoring. Parents can ask which tools are used, what data they collect, and what happens when content is blocked or a child is upset.
AI toys and connected kids' devices can record a child's voice, send it to the cloud and, in some products, pass conversations to third-party AI services. In the US, the FTC treats a child's voice recording as personal information that generally needs verifiable parental consent. Before buying, read the privacy policy, check deletion options and decide whether the toy needs to be online at all.
AI companions are chatbots built to act like a friend or partner, and a 2025 Common Sense Media survey found 72% of US teens had tried them. Parents can reduce risk by talking openly, checking age settings and parental controls, limiting private data shared, and knowing where to get help if a chat turns distressing.
Data poisoning means manipulating the data an AI learns from or retrieves, so that it behaves badly or hides a backdoor that only a trigger activates. OWASP also warns about tampered models and weak provenance in the AI supply chain. You cannot inspect a model yourself, so choose reputable providers, avoid unknown model files and verify important answers.
Banks watch for signals such as odd IP addresses, rapid transactions, mismatched identity data and funds moved in hard-to-reverse ways, and FinCEN's 2024 alert lists many of them. Detection works best on mule accounts and identity fraud. It is weaker when a real customer is persuaded to authorise a payment, so personal verification before paying still matters.
You cannot reliably spot AI-generated misinformation by how it looks or reads, so check where it came from. Trace the claim to a primary source, look for the same report from independent outlets and be cautious of content that triggers a strong emotion. Technical labelling and provenance tools are being developed by NIST and required in the EU, but they are not yet a dependable filter.
The EU AI Act is the European Union's law regulating AI by risk level. It bans some practices, requires chatbots and deepfakes to be disclosed, and puts strict rules on high-risk systems such as hiring and credit scoring. Most of it applies from 2 August 2026, with some high-risk rules delayed to 2027 and 2028.