Hallucination

A hallucination in AI refers to a phenomenon where a model ofGenerative AI (such as LLM) produces information that is incorrect, incoherent or completely invented, while presenting it as factual.

These errors stem from the very nature of LLMs, which are designed to predict statistically plausible rather than verified responses.

---

Types of hallucinations in AI

1. Invented information 📖: AI can generate inaccurate facts, fictitious references or non-existent sources.
   👉 For example, in 2023 ChatGPT invented legal cases used in a real court case, causing a scandal.
2. Lack of logical coherence 🔄: some answers may seem plausible but contain errors in reasoning or internal contradictions.
   👉 For example, a chatbot can claim that the James Webb telescope has taken the first image of an exoplanet, while the Very Large Telescope is responsible for this achievement.
3. Apparent conviction 🎭: the AI doesn't always point out its mistakes, making it difficult to distinguish between a correct answer and a hallucination. AIs present their answers confidently, even when they are making things up.
   👉 A Stanford study (2024) shows that models tend to "agree with the user, even if they are wrong.
4. Bias and contextual errors ⚠️: AI can misinterpret a question and generate an irrelevant or incorrect answer.
   👉 For example, if a user mentions a false premise ("Helium is the most abundant element"), the AI will confirm the error rather than correct it.

---

💥 Examples of hallucinations

• Dangerous advice A chatbot suggested adding glue to a pizza to fix the cheese, based on a sarcastic Reddit comment.
• 🔍 False quote Inventing references to articles or studies that do not exist. In 2023, ChatGPT generated citations to non-existent legal articles used by a lawyer in a real case.
• 📅 Historical error To claim that an event took place on the wrong date.
• 🏢 Dummy company Providing information about a company that does not exist.
• 👥 Imaginary character Attributing achievements to someone who has never done them.
• Scientific errors Google Bard incorrectly attributed the first image of an exoplanet to the James Webb Telescope (2023)

---

📊 Statistics

• 30-90% of errors in scientific references generated by chatbots (study 2024) 3.
• 3.5% → 1.8% Rate of hallucinations reduced between GPT-3.5 (2023) and GPT-4 (2025)

---

Causes of AI hallucinations

1. Lack of precise data 📉
   AI generates responses based on statistical models and not on a single answer. real understanding.
2. Data compression
   LLMs compress billions of pieces of data into parameterssometimes losing critical information (e.g. 2% answers are invented)
3. Noise or bias in training data 🔊
   Models can learn from biases or errors in their training data (e.g. sarcastic Reddit posts).
4. Lack of access to real-time sources ⏳
   An AI model may not have access to the latest updates, leading to errors on recent facts.
5. Optimising fluidity and completeness  🏃‍♂️
   Models prefer to provide a complete answer rather than admit uncertainty, encouraged by human reinforcement learning.

---

💉 How can hallucinations be reduced?

✔ Checking sources Always cross-reference the information provided by AI with reliable sources.
✔ Access to external databases Connecting AI to up-to-date knowledge bases.
✔ Improved model training Refine the training data to minimise bias.
✔ Use of filters and post-processing detect and correct inconsistencies before displaying a response.

 

Here are some other ways:

• Augmented Generation by Recovery (RAG)
  Connect LLMs to external databases (e.g. Google's Gemini checks responses via real-time searches). This reduces 30% in document summaries
• Self-checking and introspection
  Forcing models to think in several stages ("chain of thought") improves consistency. Example: OpenAI is testing models capable of expressing uncertainty ("I'm not sure about this").
• Human validation and filters
  96% of AI-generated content requires human proofreading, according to a marketing study (2025). Tools such as Vectara measure hallucinations using indices of vulnerability.
• Training on targeted data
  Specialised models (e.g. legal, medical) reduce errors by basing themselves on verified corpora

---

🔮 Outlook future

Despite progress, hallucinations remain a structural challenge. Jensen Huang (CEO of Nvidia) believes that the problem will persist "for some years to come". However, innovations such as RAG and "metacognitive" AI (self-assessment of credibility) could limit the risks, particularly in critical areas (medical, legal).