Units
1. What is Artificial Intelligence, Machine Learning and LLM? 2. How Does Generative AI Work? 3. What Can AI Do and What It Can't Do? 4. Fundamentals of Effective Prompt Writing 5. Hallucination and the Limits of Credibility 6. Prejudice and Justice 7. Data Privacy and KVKK: Where Can I Write Company Data? 8. Everyday AI Use Scenarios in the Workplace 9. End-to-End AI Workflow: Draft, Validate, Decide 10. Copyright, Ownership and Transparency 11. Responsible Use and Human Control 12. Personal AI Roadmap and Continuous Learning
Unit 5 / 12

Hallucination and the Limits of Credibility

Gains:

  • Can explain what a hallucination is and why it occurs
  • Can determine when AI output needs to be validated
  • Can apply practical techniques that increase reliability

Perhaps the most critical topic of AI literacy is in this unit: Model making up, that is, hallucination. Understanding this term is the most basic safety precaution for anyone using AI. Because a user who does not recognize the hallucination may mistake false information as true and transfer it to a business decision, customer communication or an official document. In this unit, you will learn what a hallucination is, why it is inevitable, and how to protect against it.

What is a hallucination?

A hallucination is when a language model produces information that does not actually exist in a fluent and confident language, just as if it were true. The model does not intend to lie; In producing only the "most probable continuation", it creates a text that seems probable but is actually false.

Example: If you tell a model “Recommend an academic source on topic Everything appears in the correct format and convincingly; But the source is completely imaginary. This is a classic example of hallucination.

Why Does It Happen? (Remember Unit Two)

The source of the hallucination lies in the working principle of the model. The model is not an "encyclopedia of facts"; It is a "probability machine". Its job is to produce the smoothest possible continuation of your question; It is not about bringing the most accurate information. Model:

  • He doesn't know what he doesn't know; When faced with a blank, he tends to fill in a possible answer rather than saying "I don't know."
  • It has no mechanism to control reality; It does not check whether the text it produces matches the world.
  • Fluency and accuracy are two different things for him; Just because it's fluent doesn't mean it's correct.
Caution: Hallucination is not a "malfunction"; It is an inherent feature of technology. New models reduce this but cannot eliminate it completely. Therefore, instead of waiting for a "non-hallucination model", it is necessary to learn to live with hallucination.

Where is the Risk of Hallucinations High?

Not every outcome is equally risky. The risk increases when:

High Risk

Low Risk

Exact figures, dates, statistics

Generating general ideas, brainstorming

Requesting sources, quotes, references

Change the tone of a text

Niche, specialized topics

Common, general topics

Current events and people

Summarize your text

Inside information specific to your company

Grammar correction

Three Mini Cases

Case 1 — A fake figure entered the presentation. An employee asked AI for "2023 e-commerce growth rate and source in Türkiye" for the industry report. The model returned a convincing percentage (48%) and an institution name. The employee put it into the presentation without verifying it; When it was revealed at the meeting that the figure was fabricated, both the presentation and the company's reputation were damaged. If the same employee had received the figure from the official source, there would be no problem.

Case 2 — Non-existent piece of legislation. An assistant asked the AI ​​for the "relevant article of law" for a petition. The model produced a realistic-looking item number and text; but that item was not there. Fortunately, the legal department noticed it before delivery. Lesson: legal citations are always verified from the primary source.

Case 3 — Correct method, safe result. Another employee on the same team simply had the AI ​​draft the report, entering all the numbers himself from official data and instructing him to "add no figures that are not in this text." The result was both fast and accurate. Difference: AI wrote draft, human fed facts.

Weak Prompt / Strong Prompt

Weak prompt: Give the 3 most important academic sources on this subject with the author and year.

Result: The model can safely fit non-existent sources.

Powerful prompt: Write a summary based on the text I pasted below. Use only the information in this text. Do not add any names, numbers or sources that are not in the text. If you are not sure, write "not specified in the text".<text>[paste here]</text>

Result: Since the model gets the source from you, the fitting space becomes narrower.

Practical Techniques That Increase Reliability

Mark each verifiable claim in your answer (number, date, name, source) and put them on a separate "must check" list.

Answer this question. If you're not sure, don't guess; Say "I'm not sure" and explain that you don't know what. Subject: [subject]

Answer the same question in two different ways and mark the contradictions between the two answers. So let me see which points are questionable.Question: [question]

I will verify the text below. List the risky claims I should check in order of importance and suggest where I can verify each one.Text: [paste here]

Additionally, adopt these habits:

  1. You provide the source. Paste the relevant document into the text and say "answer based on this alone."
  2. Always check verifiable facts. Verify numbers, dates, names, quotes and sources from an independent, reliable source.
  3. Ask him to indicate that he is not sure about the model. This reduces the risk but does not eliminate it completely.
  4. For high-risk jobs, have an expert inspect the output. Legal, financial and health-related outputs must pass through the filter of human experts.
  5. Use AI as a “starting point,” not as a “final word.” The draft is from him; The verification and decision is yours.
Caution: "The model said it twice, then it is true" logic is wrong. Can repeat the same mistake confidently. Repetition is not evidence of accuracy.

Common Mistakes

Common mistakes

  • Using sources without verifying them. It can create model books, articles, articles of law and links; Every reference must be confirmed.
  • Mistaking the confident tone for accuracy. The most dangerous thing about a hallucination is precisely that it seems confident.
  • Just having the model say "you verify" is enough. Self-validation of the model is not independent validation.
  • Using high-risk output without showing it to an expert. Human control is essential in money, law and health.

In summary

  • Hallucination is when the model produces non-existent information in a fluent and confident manner as if it were true.
  • The cause is not a malfunction, but the probability-predicting nature of the model; cannot be completely eliminated.
  • The risk is especially high when it comes to precise figures, sources, up-to-date information and niche issues.
  • To protect yourself: give the source, verify the facts, say "if you're not sure, tell me", have high-risk output checked by an expert.
  • Use AI as a starting point; The ultimate verification and responsibility always remains with you.

Application Task

Tell your AI tool to "suggest 3 academic sources" on a subject that requires expertise and try to confirm the sources it suggests on the internet; Note how many are real. Then repeat the same task with the "you give the source" method and observe the difference.

Checklist

  • [ ] I can explain what a hallucination is and why it occurs.
  • [ ] I can distinguish between high and low hallucination risk tasks.
  • [ ] I can reduce the risk of fabrication by giving the source to the model.
  • [ ] I independently verify the numbers, dates, names and sources.
  • [ ] I do not consider the confident tone as evidence of accuracy.
Previous Next