AI Hallucination
The phenomenon of Ai Hallucination is characterized by convincing affirmations, plausible output, but with misleading, false information in artificial intelligence (AI). These models generate text, images and audios that look and sound reliable, and turn out to be invented or incorrect (OpenAI, 2023).
Image from: AI hallucinations and how to avoid them.
What types of Ai Hallucination can occur:
According to Truefound, an enterprise Ai infrastructure company, here are some types of Ai Hallucination :
- Text and language hallucinations: Text with invented facts and false quotations.
- Image and visual hallucinations: Images generated with visual details that do not correspond to reality.
- Code hallucinations: AI coding systems can generate functions that do not exist, resulting in errors in generated technical documentation or code.
- Multimodal hallucinations: It’s when an AI system produces an incorrect interpretation based on multiple types of input, such as text, images, audio, and video.
Ai Hallucination example:
The most famous example of Ai Hallucination happened in 2023, ChatGPT AI hallucination case called “Mata v. Avianca”. In February 2023 a new yorker Lawyer, Roberto Mata, sued the airline Avianca. He claimed that during the international flight, he sustained a serious knee injury after being struck by a metal in-flight service cart pushed by a flight attendant. One of the responsable advocates in the case, called Steven Schwartz was with some difficulties to find out some similar cases that could support the arguments for his client. Therefore, he decided to use a tool that was new at the time: ChatGPT (Wikipedia, the free encyclopedia 2023).
Schwartz asked Ai for some examples of judicial decisions, in which passengers had sued airlines and won it. ChatGPT answers him with various detailed cases, including names, data and even though judges vote. The problem is that ChatGPT invented everything. All the cases did not exist. This case was one of the first cases of Ai Hallucination in chatbots that turned out in the spotlight.
Why can AI “hallucinate”?
Experts analyze Ai hallucinations not as mysterious glitches in the system, but as the natural statistical outcome of how language models are trained and evaluated. When there is no learnable pattern in the data, base models are forced to guess based on how often a prompt appears in the training data (Kalai; Vempala 2024). The root causes of this happen is during the pre-training phase, language models learn the distribution of language. The authors mathematically prove that generating valid text is closely tied to solving a binary classification problem, which is deciding whether a statement is valid or an error. Because models are optimized to minimize cross-entropy loss, natural statistical pressures force them to make guesses, which leads to hallucination, rather than always generating “I don’t know” (Kalai et al.; Venpala 2025).
Based on Google Cloud, Ai Hallucinations occur primarily due to two main factors;
- Imperfect Training Data:
Ai models learn from patterns in datasets. If training data is incomplete or biased, the model adopts these errors and generates hallucinations.
- Lack of Proper Grounding:
Without real world grounding, Ai models generate outputs that seem true but are factually wrong or completely made up.
The CEO at Standard Ai, Jordan Fisher, explain more about what is Ai Hallucination and the causes, also offering advice on how to identify risks on this collaborative Youtube video of Claude and Anthoprific channel: Why do AI models hallucinate?
AI hallucinations through some students perspective
As large language models become more integrated into student’s academic lives, they raise critical concerns about content accuracy and learning integrity. In this context, the study undertaken by Shoufan and Esmaeil (2026) investigates how university students perceive, identify, and conceptualize Ai. The study highlights that AI literacy must go beyond basic prompt engineering. Students need explicit training in critical thinking, fact-checking, and understanding how LLMs actually work so they do not blindly trust confident but incorrect Ai responses. The most common issues students encountered were fake citations/sources and completely fabricated information delivered in a very convincing, overconfident tone.
One of the students stated “when I asked an AI model to give me the biography of a scientist. It gave me a detailed answer with dates, education, and achievements but when I checked online, some of the facts were wrong or completely made up.”
Another comment “doesn’t provide the full answer, especially when we ask a specific question”.
One student stated “it could provide an answer by filling in assumptions that you as the prompter have not included.”
Source: https://arxiv.org/html/2602.17671v1
Source: https://arxiv.org/html/2602.17671v1
Most students on this research rely on their perception and own judgment or use active verification, for example; cross-checking with textbooks or running code to see if it works. Some students mistakenly view Ai like a search engine that pulls from a database and makes things up only when it can’t find the answer. But in reality, AI works by predicting the next most likely word based on statistics, not by searching a database.
How to prevent AI hallucinations
According to Google Cloud exist some tips, things that can be done to help prevent AI hallucinations, containing;
- Limit possible outcomes
- Train your Ai with only relevant and specific sources
- Create a template for your Ai to follow
- Tell your Ai what you want and don’t want
Limiting possible outcomes means that you can apply a technique called regularization when training a model. This penalizes the Ai if it makes predictions that are too extreme, stopping it from overfitting the training data and keeping it’s answer more realistic and constrained. Additionally, you need to train Ai with relevant and specific sources, because the quality of the output depends on the training data. Developers must feed the Ai data that directly relates to specific tasks it needs to perform to prevent bad patterns from forming, for example; using a clean dataset of medical imagen if it1s meant to analyze healthcare scans. It’s necessary to create a template for your Ai to follow, by giving the model a structured format or template to follow helps guide it’s generation process, for example; giving prompts for writing the title, introduction, body and the conclusion of a text. It is important to explain to the Ai model what you want and do not want. This can happen when you give feedback by telling which types of texts you like and do not like (Google Cloud Discover).
Aysan Nazarmohammadi,an AI engineer and consultant highlights that:
“As language models become more sophisticated, the gap between their articulate outputs and their actual grounding in truth remains a fundamental challenge. The techniques outlined here represent the current state-of-the-art in prompt engineering for hallucination reduction, backed by peer-reviewed research and real-world application.”
REFERENCES
KALOTRA, Sachin.; Understanding AI Hallucinations: Implications and Insights for Users. Signity Solution, 10 Abr. 2025. Accessed: 26 Out.2026.
BRODEUR-URBAS, Max.; AI hallucinations and how to avoid them. Gumloop. 19 Dec. 2023. Accessed: 27 Out. 2026.
TAUMAN KALAI, Adam et al.; Why Language Models Hallucinate article. 4 Set. 2025. Accessed: 26 Out.2026.
OPEN.CX.; AI hallucination examples. 13 May. 2025. Acesssed: 28 out. 2026.
WIKIPEDIA.; Large language model. Acesssed: 27 out. 2026.
TRUEFOUNDRY.; What are AI hallucinations. Acesssed: 27 out. 2026.
WIKIPEDIA.; Hallucination (artiAI Hallucinationficial intelligence). Acesssed: 27 out. 2026.
NAZARMOHAMMADI, Aysan.; Reducing AI Hallucinations: 6 Prompt Engineering Techniques That Actually Work . Acesssed: 5 Dec. 2025.
CHINA, Christal R.; What are AI hallucinations?. 01 Sep. 2023. Acesssed: 26 Out. 2026.
GOOGLE. Google Cloud.; What are AI hallucinations?. 2026. Acesssed: 26 Out. 2026
ESMAEIL, Ahmad-Azmi-Abdelhamid.; AI Hallucination from Students’ Perspective: A Thematic Analysis. 11 Jan. 2026. Acesssed: 25 Out. 2026
