AI keeps improving, but it can still be wrong in a dangerous way: it sounds just as sure when it is wrong as when it is right. Knowing its five common failures lets you catch most mistakes.
1. Hallucinations: confident and wrong
A hallucination is when the AI states something false as fact. It can invent a quote, a statistic, a law, a book title, or a web link.
This happens because a model predicts what a good answer sounds like. Usually that matches the truth. When it does not know, it may still produce something that sounds right.
Where it happens most: exact numbers, dates, quotes, citations, niche topics, and anything recent.
How to catch it:
- Ask for sources and open them. A real link that says something different is a red flag.
- Ask "How confident are you, and what would change your answer?"
- Ask the same question in a new chat, or in a different AI. Different answers mean you need to check.
- Turn on web search for facts, so it reads before it answers.
Your best defense is knowing the topic. When you know a subject well, mistakes jump out at you. You can question an answer when the logic seems a bit off: "That doesn't match what I've seen. Walk me through how you got there." When you don't know the topic, double-check the important parts in a second AI tool, or with a person who does. This is another reason to practice on subjects you know (see Part 1).
2. Limited context and memory
An AI can only "see" a certain amount of text at once. This is called the context window. Today's top models can hold about 1 million tokens, roughly 750,000 words, or several long books. That sounds endless, but it has two catches.
- Attention fades. In a very long chat, the AI pays less attention to things said near the start. It may forget an instruction you gave an hour ago.
- New chat, clean slate. Unless a tool has a memory feature turned on, a new chat knows nothing about earlier ones.
What to do:
- Start a new chat for each new topic.
- Repeat key instructions when a chat gets long.
- If a tool has memory, check what it has saved about you in its settings. You can usually view, edit, or delete it.
- When a long chat starts to drift, restart on purpose (see below).
Why AI summarizes long chats: it's about cost and speed. Every time you send a message, the AI re-reads the whole conversation before it answers. Remember that AI is billed by the token (Part 0). So:
- Storing your conversation costs money.
- Reading it back costs money, every single message.
- Reading it back fast, so your answer arrives in seconds with everything you said already factored in, costs even more.
A chat that has run for a week can mean re-reading a small book before every reply. To keep the service affordable and fast, many AI tools quietly summarize the older part of a long chat once it nears the limit. The AI then works from that summary instead of your exact words. It's a common trade-off across the industry: a slightly fuzzier memory in exchange for reasonable price and speed.
What that means for you:
- Expect some drift. In a long chat, the AI may lose small details, soften an instruction you gave early on, or slowly wander from the original goal. That is the summary at work, not the AI "getting dumber."
- Hallucinations get more likely. When a detail was summarized away, the AI may fill the gap with a guess. Double-check names, numbers, and decisions from early in a long chat.
- Restart on purpose. When a chat gets long, ask: "Summarize everything important from this conversation in one page." Check that summary yourself, fix what's wrong, then paste it into a new chat along with your "about me" file (Part 2).
3–5. Outdated knowledge, bias, and flattery
| Failure | What it looks like | What to do |
|---|---|---|
| Outdated knowledge | A model learns up to a cutoff date. Without search it may describe old prices, old leaders, or products that no longer exist. | Turn on web search. Ask "What date is your information from?" |
| Bias | Models learn from human writing, including its blind spots. Answers can lean toward one culture, one viewpoint, or common stereotypes. | Ask "What perspectives are missing?" or "How would someone from [group] see this?" |
| Flattery | Models tend to agree with you and praise your ideas. They may change a correct answer if you push back. | Ask for criticism directly: "What is wrong with this plan?" Do not treat agreement as proof. |
Flattery is the sneakiest of the five. If the AI always agrees with you, you are not getting a second opinion. You are getting an echo.
Privacy and safety basics
- Never paste passwords, bank or card numbers, Social Security or ID numbers, or other people's private details.
- Check the data setting. Most chat apps have a setting about whether your chats can be used to train future models. Look it up in Settings and choose what you are comfortable with.
- Work rules come first. Many employers limit which AI tools you can use for work files. Ask before uploading company documents.
- Beware of scams. AI makes fake voices, photos, and emails easy. If a call or message asks for money or a code urgently, hang up and contact the person another way.
- Health, money, and legal questions: AI is a good way to prepare questions for a professional, not a replacement for one.
Found this helpful? Raya.ai is built on the ideas in this series. Learn more at raya.ai (opens in a new tab), or download the app on iOS (opens in a new tab) or Android (opens in a new tab).