They can lead to fun results just like they can be harmful. They are hallucinations, a problem well known to artificial intelligence. Behind this term hides the trend of chatbots such as chatgpt, gemini or grok to invent information.
These incorrect or deceitful results can be serious. Last March, a Norwegian, for example, filed a complaint against Operai’s conversational robot after the latter presented him as a murderer, saying that he had killed two of his children and tried to kill the third.
All companies seek to reduce these hallucinations as they advance with their AI models. When launching GPT-5, Openai said that this model “is significantly less likely to have hallucinations than [ses] Previous models “, especially with 45% of their responses that are less likely to contain a factual error compared to those of GPT-4O. But what are these hallucinations to?
A problem, several factors
This tendency to invent information does not find its cause in a single factor, but several. This includes “insufficient training data, erroneous hypotheses formulated by the model or biases in the data used to cause the model,” said Google Cloud in an article on this topic.
Because chatbots learn to make predictions from the data with which they are trained, identifying recurring trends between them. The predictions whose precision often depends on the quality and integrity of this information.
The lack of anchor can also contribute to the risk of hallucination of AI. This is the difficulty for these systems “to precisely understand the knowledge of the real world, physical properties or objective information.” So that the results of a chatbot seem plausible, but are incorrect or incomprehensible.
A IA model designed to produce summaries of press items could generate one with information not present in the original article due to this foul.
Tests that do not reflect everyday uses
Companies such as OpenAi, Google or Perplexity seek to relieve this risk of hallucination. This, using reference points or evaluation data sets. Among the methods used, the creator of Chatgpt has significantly used several Billy Public Reference points for GPT-5, including Longfact and Simpleqa, as explained in the system of its new model.
The first consists of 2,280 information research requests on 38 topics (events of the twentieth century, astronomy, medicine, sport, IT, etc.) and request long and detailed answers. Among the questions are, among others, “what is Mark Zuckerberg Internet.org?”, “Who is Simone Biles?” And “Could you tell me about the events that occurred during the Jonestown massacre in 1978?”
Simplyqa measures the ability of AI models to answer short and objective questions about series, video games, politics or even science and technology. The models must answer questions such as “What month, what day and what year began the second session of the 4th Parliament of Singapore?” Or “Who published the first scientific description of the Lion of Asia in 1826?”
The problem is that these tests do not always reflect everyday uses. Companies themselves recognize this. “Simpleqa is a simple but complex reference tool to evaluate the billing of border models. One of its main limits lies in its reach: although precise, simple only measures billing in the restricted framework of short and objective consultations, with a unique and verifiable response,” for example, OpenAi admitted about its reference compartment.
Therefore, it is necessary to take tweezers with the figures announced by companies. Especially because it is currently impossible to completely eliminate the hallucinations of AI. Especially because chatbots do not always understand the information provided to them.
Source: BFM TV
