- Foundation
- About 16 min
- 5 practice questions
AI limitations and ethics: errors, hallucination, bias, privacy, deepfakes and human oversight
Explain why AI makes errors and hallucinates, describe the risks of bias, privacy loss, deepfakes and unexplained decisions, and say why a person checks AI output and stays responsible for the decision.
If something looks wrong, please tell us.
By the end of this lesson you can
- Explain why AI makes errors, including hallucination, false positives and false negatives.
- Describe the main ethical risks - bias, privacy, deepfakes and lack of explainability - with an example of each.
- Explain why a person must check AI output and stays responsible for decisions made with it.
AI can be a powerful helper for a student or an office, but only if you know its limits. A chatbot can explain a topic in seconds and, in the same answer, state a wrong date with full confidence. This lesson covers where AI goes wrong, the ethical risks it brings, and the safeguards that keep a person in charge.
Errors and hallucination
Every AI system makes mistakes. A fraud filter may block a genuine payment (a false positive: it said 'yes, fraud' when the truth was 'no') or let a real fraud through (a false negative). A chatbot may hallucinate: give fluent, confident, but false or made-up information, such as an invented book title, quotation, date or law section. It happens because a chatbot built on a language model produces the words most likely to follow, based on patterns in its training text. It is designed to sound right, not to look facts up in a checked record. Confidence of tone is therefore no evidence of accuracy.
How AI goes wrong
-
False positive
Says 'yes' when the truth is 'no'.
- A genuine payment blocked as fraud
-
False negative
Says 'no' when the truth is 'yes'.
- A real fraud let through
-
Hallucination
Fluent, confident, but false or made-up information.
- An invented book title, quotation, date or law section
The main risks and their safeguards
| Risk | What it means | Example | Safeguard |
|---|---|---|---|
| Bias | Unfair results because the training data were unbalanced or reflected past unfairness | A face system trained mostly on one group of faces works worse for others | Balanced, representative data; testing results across groups; human review |
| Privacy | Personal data collected, stored or exposed without proper care | Pasting a customer's details into an online AI tool, which may store what is typed | Never paste passwords, OTPs, Aadhaar numbers, bank details or confidential office files; remove names and numbers first |
| Deepfakes | AI-made or altered video, audio or images showing people saying or doing what they never did | A cloned voice of a relative urgently asking for money | Pause; verify through another channel, such as calling back on a known number; do not forward |
| Lack of explainability | Many learning models are a 'black box': even their makers cannot fully say why an answer came out | A loan rejected with no reason anyone can give | Use explainable methods for decisions about people; let people ask for a human review |
Human oversight: a person stays responsible
AI cannot take moral responsibility or be held accountable; people and organisations are. So a person stays in the loop for anything that matters - a medical decision, a loan, an exam answer, an official letter. The human catches errors, adds context the AI lacks and owns the final decision. If an office sends a letter with a wrong figure drafted by AI, the office that sent it is responsible, not the tool. For facts such as dates, fees, rules or figures, check against the original or official source - a textbook or the official notice - before you use them.
Worked example 1: checking a chatbot's answer
- You ask a chatbot who coined the term 'artificial intelligence'.
- It replies 'Alan Turing, in 1950' in a confident tone.
- You check your notes or a standard textbook: the term was coined by John McCarthy for the 1956 Dartmouth workshop; Turing's 1950 paper proposed the imitation game.
- Conclusion: the chatbot mixed up two real facts - a typical hallucination. Write the checked answer, not the chatbot's.
Worked example 2: false positives and negatives
Suppose a fraud filter checks 1,000 payments (illustrative figures). It flags 30. Of those 30, 24 really were fraud and 6 were genuine (6 false positives). It also missed 4 frauds among the payments it passed (4 false negatives). Working: total real frauds = 24 + 4 = 28. Wrongly blocked genuine payments = 6. Missed frauds = 4. A stricter filter would cut missed frauds but block more genuine customers - a human team decides the right balance and reviews flagged cases.
Worked example 3: a suspicious video
A video forwarded in a group shows a well-known person announcing that an exam has been postponed. Steps: do not forward it; look for the announcement on the official website or official notice of the body that conducts the exam; if it is not there, treat the video as unverified, however real it looks. Deepfakes can copy faces and voices closely, so the source matters more than the appearance.
Checking a suspicious video
- Do not forward it
- Check the official source The official website or notice of the body that conducts the exam.
- Not there? Unverified Treat it as unverified, however real it looks.
Use AI as an assistant, not an authority. Fluent is not the same as true: check every fact that matters against the official source, never paste personal or confidential data, and remember that a person stays responsible.
Key takeaways: AI makes errors - false positives, false negatives and hallucinations. Bias comes from data, privacy needs care with what you type, deepfakes need checking at the source, and black-box decisions about people need explanation. A person checks the output and takes responsibility for the decision.
Check your understanding
-
1. In AI, 'hallucination' means:
- The AI system switching itself off
- The AI giving confident but false or made-up information
- A camera seeing images in the dark
- A virus attacking the AI
Show answer
Answer: B. Hallucination is when an AI system, especially a chatbot, produces fluent, confident output that is false or invented, such as a fake reference or a wrong date.
-
2. You receive a voice message that sounds exactly like your uncle, urgently asking you to send money by UPI. What is the safest response?
- Send the money at once because the voice matches
- Forward the message to family groups to warn them
- Reply asking him to send a photo as proof
- Call your uncle back on the number you already know before doing anything
Show answer
Answer: D. Voices can be cloned with AI, so a matching voice proves nothing. Verifying through a separate, known channel is the safeguard; a photo could also be faked, and forwarding spreads the scam.
-
3. A clerk wants an AI tool to improve the wording of a letter to a citizen. Which is the right practice?
- Remove the citizen's name, Aadhaar number and other personal details before pasting the text
- Paste the full letter with all details so the tool has context
- Paste the office password as well so the tool can send the letter
- Paste the letter and send the AI's version without reading it
Show answer
Answer: A. Personal data and passwords must never go into AI tools, which may store what is typed. Removing personal details protects privacy, and the clerk still reads and checks the result before sending.
-
4. A recruitment model trained on a company's past hiring records keeps ranking applicants from one background lower. The most likely cause is:
- Hallucination in a chatbot
- Too many hidden layers
- Bias in the training data, which reflected past unfair decisions
- A false negative in the fraud filter
Show answer
Answer: C. A model learns the patterns in its data. If past decisions were unfair, the model repeats that unfairness - this is bias. Checking results across groups and human review are needed.
-
5. A fraud filter flags 60 payments: 51 are real fraud and 9 are genuine. It also lets 6 real frauds through. How many real frauds were there in total?
- 51
- 57
- 60
- 66
Show answer
Answer: B. Real frauds = frauds caught + frauds missed = 51 + 6 = 57. The 9 genuine payments flagged are false positives and are not counted as fraud.
Sign in to keep track of the lessons you finish.