Hallucinations without psychedelics: When AI shows its weaknesses

From healthcare to education and corporate security, many sectors now use artificial intelligence. Although responsible for major advancements in many areas, AI is not always very reliable: it can invent information. This has significant consequences...

What is an hallucination?

This term refers to a response generated by AI that contains information that might seem plausible but is actually incorrect or unrelated to the context.

These hallucinations often result from AI's ability to extrapolate by applying probabilities based on the data it was trained on. Additionally, it is important to remember that artificial intelligence cannot distinguish between what is possible and what is not. This is simply because it lacks personal experience and awareness of its environment or reality, which sometimes leads it to generate false or misleading information.

These hallucinations are generally related to the validity and quantity of the data provided during training, and to how the AI interprets and generates new information. For example, Google’s AI integrated data from the Reddit social network, suggesting that users should use glue to keep food on a pizza.

During training, it is necessary to ask the AI many questions and indicate what is right or wrong, covering a sufficiently varied range of topics so it can improve. As Greg Brockman (co-founder of OpenAI) noted in a TED talk in 2023, ChatGPT had not been trained to indicate to a human when they were mistaken.

Excerpt from a conversation between Greg and ChatGPT > Q4 2022 - 5 + 7 = 15, right? 5 + 7 = 15, right? Answer: Yes, that is correct. 5 + 7=12 | Answer Q1 2023 > No, that's incorrect. The correct sum of 5 and 7 is 12. — Excerpt from the TED talk: The inside story of ChatGPT's astonishing potential

Health, Education, Security: What Could Be the Consequences of an AI Hallucination?

According to the Ministry of Economy, the introduction of AI in the healthcare sector aims to "significantly improve the diagnosis of many pathologies." It is therefore an "aid to diagnostic or therapeutic decision-making" intended to "help enrich the reasoning of professionals."

In this context, without careful human supervision, an AI hallucination could have serious consequences: misdiagnosis, inappropriate treatment prescriptions, delays or omissions in necessary care, or even unnecessary or dangerous medical interventions. This could lead to a loss of trust from patients and healthcare professionals towards artificial intelligence, hindering the adoption of this technology in the medical field.

Moreover, an AI hallucination in healthcare inevitably raises questions of responsibility. In the case of a medical error, who should be held responsible? The doctors, the artificial intelligence itself, or the people who designed it?

The education sector is also experiencing disruptions due to the widespread use of AI. As students can easily use it to answer course questions or write essays, AI hallucinations pose a risk of misinformation. Since it is now used by a very large audience, incorrect information could spread quickly, potentially compromising students' learning.

On the other hand, AI hallucinations can pose a serious security problem. For instance, in the context of facial recognition use, a hallucination could lead to the arrest of innocent people, privacy breaches, and human rights violations. It is already established that these systems have biases that disproportionately affect certain populations, such as ethnic minorities, exacerbating issues of discrimination and social justice.

Some Examples of Hallucinations

In writing this article, I chose to assume the role of a student wanting to write a report on the first round of the 2020 French legislative elections using ChatGPT. However, no legislative elections took place that year: I intentionally included an error in my request. We can imagine that the student I am embodying did not realize her mistake when writing her prompt (the English term for sending a written command to an AI).

Will ChatGPT point out my mistake, inform me that this information is not available, or write an incorrect report?

Experience:

Prompt : Write me a report on the first round of results of the 2020 French legislative elections.

Report on the First Round of Results of the 2020 French Legislative Elections

Introduction

The 2020 French legislative elections took place in a unique political and social context, marked by social movements and economic challenges. The first round of these elections was held on March 15, 2020. This report provides a detailed analysis of the results of the first round, highlighting the performance of major political parties, turnout rates, and key issues of this election.

Political Context

The French political landscape in 2020 was dominated by several major issues:

Yellow Vest Crisis: A significant social movement that has marked the political scene since late 2018.
Economic Reforms: Notably those concerning pensions and the social security system.
Environment and Climate: An increasing concern among voters, influencing party programs.

Turnout Rate

The turnout for the first round of the 2020 legislative elections was relatively low, around 40%, partly due to concerns related to the COVID-19 pandemic which was beginning to spread in France at that time. This low turnout had a significant impact on the results, favoring parties with well-established and motivated support bases.

Results by Political Parties

The Republic on the Move (LREM)

The presidential party, La République En Marche (LREM), managed to maintain a significant presence, although its support showed signs of fragility compared to the 2017 legislative elections. LREM obtained about 22% of the votes, reflecting a slight decrease in popularity.

The Republicans (LR)

The Republicans experienced a slight rebound with about 18% of the votes, capitalizing on dissatisfaction with some of the government’s policies. Their campaign focused on economic and security proposals, which found a favorable echo among some voters.

The National Rally (RN)

The National Rally, led by Marine Le Pen, consolidated its position by obtaining around 23% of the votes, becoming one of the main beneficiaries of protest voting. Their focus on immigration and security continued to mobilize a significant portion of the electorate.

France Insoumise (LFI)

France Insoumise, led by Jean-Luc Mélenchon, recorded about 11% of the votes. Although this represents a slight decrease compared to previous elections, they maintained a solid core of support among radical left voters.

Europe Ecology - The Greens (EELV)

Europe Ecology - The Greens experienced a notable rise with about 14% of the votes, reflecting the increasing prominence of environmental concerns among French voters. This result positions EELV as a key player in any left-leaning coalition.

Socialist Party (PS)

The Socialist Party, once dominant, continued to lose ground with about 8% of the votes, illustrating the difficulty of this party in redefining itself in the current political landscape.

Issues and Perspectives

The results of the first round of the 2020 legislative elections highlighted several key issues for the second round:

Coalitions and alliances: Parties will need to consider alliances to maximize their chances in the second round.
Voter mobilization: The main challenge will be to mobilize abstainers, particularly in the context of a pandemic.
Strategy adaptation: Each party will need to adjust its strategy based on the results of the first round and local dynamics.

Conclusion

The first round of the 2020 French legislative elections revealed a continued fragmentation of the political landscape, with a rise in environmental concerns and a sustained focus on security and immigration issues. The low turnout remains a major challenge for the legitimacy of future elected officials. The results of the second round will be decisive for the political direction of France in the coming years.

ChatGPT therefore responded to my request by writing a report on elections that never took place and providing incorrect information.

To go further, I wanted to check if this information matched the results of other legislative elections (2022, 2017, etc.), or if it was entirely fabricated by the AI. I searched for the results of previous legislative elections but did not find the same data. I also restarted the conversation with ChatGPT several times, asking the same question. Each time, ChatGPT provided slightly different figures, for example, stating that LR obtained 18% of the votes, before announcing in another response a score of 18.7%.

This exercise constitutes a striking example of AI hallucinations and the misinformation they can produce.

Erroneous Statements from Bard

Bard, the generative AI created by Google and deployed in France in 2023, is also capable of generating text through prompts... and hallucinating. Google has published examples of responses generated by Bard on its website and social media. Among these is a screenshot of an exchange regarding the James Webb Space Telescope.

Prompt : “What are the discoveries of the James Webb Space Telescope that I can talk about with my nine-year-old child?”

In its response, Bard states that “JWST (James Webb Space Telescope) has taken the very first photos of a planet outside our own solar system”. This is an incorrect statement as the first telescope to photograph an exoplanet was the Hubble Space Telescope, launched in 1990.

Excerpt from a response by Google Bard: Your 9-year-old might like these recent discoveries made by The James Webb Space Telescope (JWST): • In 2023, The JWST spotted a number of galaxies nicknamed — Google's social media post

How to Detect and Minimize Hallucinations?

It is important for users to maintain a critical perspective on the responses provided by AI and cross-check sources to verify if the generated content is not a hallucination.
To reduce hallucinations, users can provide clear and precise context to AIs like ChatGPT, so that they better understand the situation and provide a coherent response. It is preferable to avoid vague or confusing prompts.
Bilingual users can also give instructions to the AI in English. Indeed, there is more data available in English as it is one of the most spoken languages in the world. As the AI is trained more on English data, using this language could reduce the risk of hallucinations.
It is also possible to copy articles, links, or sources directly into ChatGPT and ask it to generate a response based on these sources. This technique helps ensure the accuracy of the information provided by the AI.
Users can employ a method called "multi-prompting", which involves creating different prompts to address the same topic from different angles. This method allows for verifying the coherence of AI responses by cross-checking them.
Finally, it is our responsibility to stay informed about the limitations of AI to reduce misunderstandings and combat misinformation.

AI hallucinations represent a major challenge. Indeed, decision-making by AI can be dangerous due to these hallucinations, as well as the lack of nuance and understanding of context. Thus, AI responses can be inappropriate and endanger human health, learning, or safety. That is why it is essential not to rely solely on machines and to maintain human intervention to supervise, correct, or make decisions.