AI & Future
AI tools are powerful and easy to misuse by accident. Here are calm, practical habits for protecting your privacy, your facts, and the people around you.
The single most useful rule for using AI safely is this: treat every generated answer as an unverified draft written by a confident stranger, and never paste anything into a chatbot that you would not email to a company you do not fully trust. Almost every real-world AI mishap traces back to breaking one of those two rules, and everything below is just how to keep them.
Large language models (LLMs) like ChatGPT, Claude, and Google Gemini are trained to produce fluent, confident text, not to be correct. They "hallucinate" — state false facts, invent statistics, and fabricate sources — in exactly the same authoritative tone they use when they are right, so the writing itself gives you no signal about which is which.
This is not a hypothetical risk. In Mata v. Avianca (2023), New York lawyer Steven Schwartz filed a brief in which ChatGPT had invented six court decisions, complete with fake quotes and citation numbers; Judge P. Kevin Castel sanctioned the firm $5,000. In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline liable after its support chatbot gave a grieving passenger wrong information about bereavement fares — the "the bot said it, not us" defense failed outright.
The practical habit is to match your verification effort to the stakes. For brainstorming, rough drafts, or explaining an unfamiliar concept, a small error costs you almost nothing. But before you act on anything touching health, money, law, or safety, confirm it against a primary source: the actual statute, the manufacturer's spec sheet, a paper you can open yourself.
AI is least reliable on precise specifics — exact dates, dollar figures, named studies, statute numbers, URLs, and direct quotations. These are also the details that look most authoritative on the page. So when an answer hinges on a crisp claim you plan to repeat, such as "Section 12(b) requires..." or "a 2021 study found 43%...", treat that as your cue to check. Ask the model to name its sources and then actually open them; fabricated citations collapse the instant you search for them. Tools that cite as they answer, like Perplexity or Gemini with grounding, make this faster but do not remove the need to click through.
When you type into a consumer AI tool, that text usually leaves your device for the company's servers, where retention and reuse rules differ from one product to the next. In 2023, Samsung engineers pasted confidential source code and internal meeting notes into ChatGPT to debug and summarize them; the data left the company's control, and Samsung banned public generative-AI tools internally shortly after.
Keep the following out of any chatbot: passwords, full account or card numbers, government ID numbers, medical records, and other people's personal details that they never agreed to share. For confidential work, check whether your employer offers an enterprise tier — Microsoft Copilot with commercial data protection, ChatGPT Enterprise, or Claude for Work — because those contractually exclude your inputs from model training in a way the free consumer apps do not.
On the major consumer tools, the opt-outs take under a minute:
Be especially wary of free, unknown apps and browser extensions that merely wrap a model. A "free" tool still has to cover its costs, and your data is sometimes the payment. Spend the minute reading the privacy page before trusting it with anything sensitive, and turn on two-factor authentication (2FA) on your AI accounts — a saved chat history is a rich target if an account is ever breached.
Passing off AI-generated work as entirely your own can mislead the people relying on it, and in school, at work, and in publishing it can carry real consequences. The fix is transparency where it matters. You do not need a disclaimer on a grocery list, but you do on anything that could be mistaken for a human's firsthand work — or for reality itself.
This matters most with synthetic media. An AI image that looks like a photograph, or AI text presented as eyewitness reporting, can deceive people even when you intended no harm. Regulators are catching up: the EU AI Act (Article 50) requires clear labeling of AI-generated deepfakes and of interactions with a chatbot. Industry tooling is arriving too — the C2PA "Content Credentials" standard embeds provenance metadata into a file, and Google's SynthID watermarks AI-generated images. A single honest line that AI was involved does the same job for free.
Because LLMs learn from existing human text, they reproduce its biases. A résumé screener or a loan-summary tool built on one can quietly disadvantage groups that were underrepresented or stereotyped in the training data. When AI helps shape a decision about people, keep a human reviewing the actual outcomes and ask directly whether the result is fair, rather than assuming a neutral-sounding tool produces neutral results.
The thread running through every habit above is staying in charge. AI is at its best handling volume and speed while you supply the judgment, the context, and the final call. Two practices keep you in that seat:
The most common mistake is trusting a model more as it becomes more fluent, but fluency and accuracy are unrelated in an LLM. Newer models hallucinate less, not never, and they do it more convincingly. Meet a polished answer with the same healthy skepticism you would give a clumsy one, and you keep all the speed without inheriting the errors.
Used this way, AI is not something to fear or hold at arm's length. Verify what matters, guard your data, disclose where it counts, and keep the final judgment human — build those four habits and the technology becomes exactly what it should be: a fast, capable assistant with you firmly in control.
Only if training is turned off and the documents contain nothing confidential, or you are on an enterprise plan that excludes your inputs from training. On a personal free or Plus account, assume anything you paste could be retained and, unless you have opted out, used to improve the model. When in doubt, strip out names and numbers first.
You usually cannot tell from the writing alone — that is precisely the danger. Check anything specific and consequential: ask for sources and open them, and cross-reference dates, figures, and quotes against a primary source. If a cited study, case, or URL does not appear in a normal web search, treat it as invented.
Not for everything, but yes wherever it could mislead — schoolwork, journalism, professional deliverables, and any image or text that could be taken as real or firsthand. A brief note is usually enough, and some contexts now require it, including many schools and publishers and, for deepfakes, the EU AI Act.
Paid consumer plans are not automatically more private, but enterprise and business tiers usually add contractual protections that keep your data out of training. The larger risk with free tools is the unknown third-party apps and extensions that wrap a model — vet who is actually receiving your text before you trust them with anything sensitive.
Priya translates the fast-moving world of AI and the internet into things you can actually use and understand. She's curious but skeptical, quick to separate genuine progress from hype, and keen to help readers use new tools wisely rather than fearfully.
Keep reading
AI news moves fast and most of it is noise. Here is a calm, jargon-free system for staying informed about what matters without burning out on every headline.
Priya Nadar
Recommendation systems shape what you watch, buy, and read every day. Here is a clear, jargon-free look at how they work and how to stay in control.
Priya Nadar