Five mistakes I see when I help people learn AI

Mistake 1: what you put into an AI app may not stay private

Human relationships can be real and valuable. We trust our friends and feel they understand us, so we tell them private things from our lives. A chat with an AI can feel the same, but your words are also saved as text in an app that a company runs.

A chat can also leak from the app itself. In October 2026, a Czech online supermarket said an outsider had gained access to part of its customer support communication. That included records of conversations with its virtual assistant. Some customers had written a password or card details in those conversations and the company contacted them. Reports don't say how the outsider got in. In February 2026, a security researcher said a popular AI chat app had left its database open to anyone. The researcher claimed 300 million messages from over 25 million users were exposed. The app's maker reportedly fixed the problem within hours.

Sometimes we may not even notice that a conversation has become public. In 2025, shared ChatGPT chats could show up in Google search results. People had chosen to make them shareable in a pop-up window. OpenAI's security chief said the feature gave people too many chances to share things by accident and OpenAI removed it. Later that month, Forbes reported that Google had indexed more than 370,000 shared Grok conversations. Pressing share created a public web address that search engines could list and users got no warning.

Apps have rules of their own too. OpenAI's help page says it may use content from its Free, Plus and Pro services to train its models. You can opt out in a setting called "Improve the model for everyone". Anthropic's privacy page lists cases when chats can be used to improve Claude, including a setting called Model Improvement. Google's Gemini page asks people not to enter confidential information that they wouldn't want a reviewer to see.

Advice: don't type anything you wouldn't want a person to see. Think before you press share. Then check the app's settings for how it may use your chats.

Mistake 2: an app built with AI may be open to everyone

You've just got a new shop that looks finished. The shelves are full and the lights work. But nobody has checked who has a key to the back door.

With an AI app builder such as Lovable, you describe an app in plain words and get a working one quickly. It's called vibe coding. The app can work well and still leave its data open to anyone. Checking who may read and change that data is a separate job and it is yours.

Two examples. In 2025, a researcher scanned 1,645 apps listed on Lovable's showcase and found 170 with missing or weak database access rules. That is his own count.

In early 2026, the security firm Wiz reported that a social site for AI agents had left its database open. Its founder said he had not written any of its code. Wiz found 1.5 million API tokens, plus thousands of email addresses and private messages. The missing access rules were the cause and the team fixed the problem within hours.

The makers say the same. Lovable's security page says misconfigured access rules are a common cause of data leaks. It also says you are responsible for making sure your app is secure enough for what it does. Lovable runs a quick scan when you publish, but says scans do not replace a thorough security review. Supabase's guide tells builders to turn these rules on for every table the app exposes.

Advice: before you share an app an AI built, find out who can read and change its data. Run the builder's security scan, and if the app holds personal data, use professional tools or get expert advice on keeping it secure.

Mistake 3: the AI telling you it's right is not a check

Humans are social creatures. We are more likely to believe something when we think it is popular, as a 2022 study of more than 600 US adults found. That can work well with a close friend or a group you can rely on. With AI it is different. An AI knows what it learned in training, or what it can find in outside sources such as the internet. It can also make things up, and it may not know how it reached an answer. And even when an answer is true, it may not suit your needs, so you still have to judge it yourself.

Two examples. In 2025, Anthropic's researchers looked inside one of their models and asked how it had added 36 and 59. It described the school method of carrying the 1, but its real method was different. So an AI's account of its own steps may not match what it did. In courts, a public database lists over 2,000 decisions in which a court found or implied that AI-made-up material was relied on.

Advice: ask for a source or the calculation behind an answer, then check it somewhere other than the AI. The higher the stakes, the more this matters.

Mistake 4: the first answer shows what the AI guessed

Imagine a new colleague joins your team and you ask them to write a short update. They are bright and keen to help. But they don't know who will read it or how long it should be. So they fill the gaps with guesses and you see those guesses in the first draft.

An AI handles a request like this in the same way. It can write an answer that sounds great. But if you don't say what you need, it may be more polish than real help. The first answer shows you what it guessed.

Anthropic's own prompting guide tells its readers to think of Claude as a brilliant but new employee who lacks context on their norms and workflows. A 2026 study in Nature Medicine shows a related problem. It used ten medical scenarios and 1,298 people. Tested alone, the AIs named the right condition in 94.9% of cases. People using the same AIs named a relevant condition in fewer than 34.5%, no better than people who used a source of their choice. The authors report people giving incomplete information and the AI misreading some questions. It is a medical study, so treat it as one example, not a rule for every task.

Advice: tell it what's off, for example "it's for my manager" or "keep it to five lines", then ask for a new version.

Mistake 5: long chats use up your limit faster

Imagine you meet a friend. You remember details about them, such as the stories you share and how your last meeting went. You listen and adapt your answers to what they say. That makes the talk engaging and builds the relationship, because they can tell you're listening. Talking to an AI works in a similar way. For a deep and useful conversation, you need to build a shared story: the context.

There is one difference: with AI, you pay for the conversation. The price comes with the context, which is everything said earlier in the chat. The bigger the context, the more you pay. In a chat app the cost is your usage limit, and Claude's help pages list conversation length and file attachment size among the factors that affect it. In a pay-as-you-go tool the cost is money: Anthropic's Claude Code guide says costs grow with the amount of context the tool processes. Other apps count usage differently, so check yours.

An AI can reason well from a short summary. So it is cheaper to start a new chat from a summary than to keep adding text to the old one. Ask the AI to summarise the long chat, or write a handover note: a short summary you paste into the new chat. Long documents take up the same space as the chat.

Advice: start a new chat when the task changes or the chat gets long and attach only the pages you need.