AI Hallucinations: From ‘You Should Eat Rocks’ to Serious Legal Pitfalls

ControlAI

May 30, 2024

See All Posts

AI technology is advancing rapidly, with major announcements from leading tech companies showcasing their latest developments. Google recently announced plans to launch an AI assistant capable of seeing and hearing. Microsoft is unveiling Copilot, an AI system set to provide a powerful AI platform for developers to accelerate local AI development. Apple, too, is expected to focus heavily on AI in its upcoming Worldwide Developers Conference. Advanced AI is no longer a futuristic hypotesis but a present reality, but this new reality comes with significant challenges, particularly the phenomenon known as AI ‘hallucinations’.

AI hallucinations occur when generative AI models produce responses that do not match reality. In artificial intelligence, a hallucination refers to a response generated by an AI that contains false or misleading information presented as fact. This occurs because AI systems, especially large language models (LLMs) like ChatGPT, predict text based on their training data, which can lead to plausible-sounding but incorrect outputs. 

While they excel at identifying patterns and generating information, they can also create incorrect or misleading outputs. This issue is well-known in the AI community, with some experts relatively optimistic about future (though not present) solutions. Others believe these errors - that’s what we’re talking about, errors - are an inevitable part of AI. Sundar Pichai, Google’s CEO himself, admitted that hallucinations are an "inherent feature" of  LLMs, the AI technology that drives most AI systems nowadays, and that it still is “an unsolved problem”.

The implications of AI hallucinations are far-reaching, affecting both mundane and critical areas. For instance, Google’s AI Overviews tool, introduced in the U.S. this month, faced significant backlash for spreading bizarre misinformation. One notable incident involved a satirical query asking, ‘How many rocks should I eat?’ The AI tool responded by recommending ‘eating at least one small rock per day’, falsely citing UC Berkeley geologists. On X, internet analyst Jeremiah Johnson shared a thread featuring even more bizarre responses, which include being able to train eight days a week or that 1919 was only 20 years ago.

While the examples of AI hallucinations may initially seem amusing, the implications can be gravely serious, especially in fields that demand high precision, such as law and medicine. In the legal domain, nearly three-quarters of lawyers are considering using generative AI for tasks like document review, contract drafting, and legal memoranda writing. A notable case involved a New York lawyer facing sanctions for citing fictitious cases generated by ChatGPT. These false responses are not only inaccurate but potentially very dangerous, particularly when we consider the latest Stanford report, which reveals that general-purpose chatbots hallucinate 58% to 82% of the time on legal queries. In the medical field, the stakes are even higher. AI has the potential to revolutionise healthcare by expediting diagnostic processes, personalising treatments, and easing bureaucratic tasks for overburdened staff. However, AI's struggle to differentiate true from false information poses significant risks. AI hallucinations can mislead even seasoned doctors if they mistakenly use tools like ChatGPT as search engines, leading to the dissemination of fabricated plausible facts and fake bibliographic references. 

The need for transparency and rigorous benchmarking of AI tools is evident. Many tools lack systematic access, detailed model information and evaluation results, making it challenging for lawyers - or anyone - to assess their reliability. While AI advancements are impressive and herald a new era of technological innovation, the persistent issue of accuracy and reliability cannot be ignored. If we accept that current state-of-the-art AI systems cannot be entirely accurate, we would have to reconsider how we deploy them. They are excellent idea generators but not independent problem solvers. We shouldn’t allow ourselves to forget that.

If you want to know more about the challenges of AI governance, the regulation of synthetic media, and the global security implications of AI advancements, join us on Discord at https://discord.gg/2fR2eZAQ4a. Here, we can collaborate, share insights, and contribute to shaping the future of AI in a manner that safeguards our security and democratic values and fosters responsible innovation.

AI technology is advancing rapidly, with major announcements from leading tech companies showcasing their latest developments. Google recently announced plans to launch an AI assistant capable of seeing and hearing. Microsoft is unveiling Copilot, an AI system set to provide a powerful AI platform for developers to accelerate local AI development. Apple, too, is expected to focus heavily on AI in its upcoming Worldwide Developers Conference. Advanced AI is no longer a futuristic hypotesis but a present reality, but this new reality comes with significant challenges, particularly the phenomenon known as AI ‘hallucinations’.

AI hallucinations occur when generative AI models produce responses that do not match reality. In artificial intelligence, a hallucination refers to a response generated by an AI that contains false or misleading information presented as fact. This occurs because AI systems, especially large language models (LLMs) like ChatGPT, predict text based on their training data, which can lead to plausible-sounding but incorrect outputs. 

While they excel at identifying patterns and generating information, they can also create incorrect or misleading outputs. This issue is well-known in the AI community, with some experts relatively optimistic about future (though not present) solutions. Others believe these errors - that’s what we’re talking about, errors - are an inevitable part of AI. Sundar Pichai, Google’s CEO himself, admitted that hallucinations are an "inherent feature" of  LLMs, the AI technology that drives most AI systems nowadays, and that it still is “an unsolved problem”.

The implications of AI hallucinations are far-reaching, affecting both mundane and critical areas. For instance, Google’s AI Overviews tool, introduced in the U.S. this month, faced significant backlash for spreading bizarre misinformation. One notable incident involved a satirical query asking, ‘How many rocks should I eat?’ The AI tool responded by recommending ‘eating at least one small rock per day’, falsely citing UC Berkeley geologists. On X, internet analyst Jeremiah Johnson shared a thread featuring even more bizarre responses, which include being able to train eight days a week or that 1919 was only 20 years ago.

While the examples of AI hallucinations may initially seem amusing, the implications can be gravely serious, especially in fields that demand high precision, such as law and medicine. In the legal domain, nearly three-quarters of lawyers are considering using generative AI for tasks like document review, contract drafting, and legal memoranda writing. A notable case involved a New York lawyer facing sanctions for citing fictitious cases generated by ChatGPT. These false responses are not only inaccurate but potentially very dangerous, particularly when we consider the latest Stanford report, which reveals that general-purpose chatbots hallucinate 58% to 82% of the time on legal queries. In the medical field, the stakes are even higher. AI has the potential to revolutionise healthcare by expediting diagnostic processes, personalising treatments, and easing bureaucratic tasks for overburdened staff. However, AI's struggle to differentiate true from false information poses significant risks. AI hallucinations can mislead even seasoned doctors if they mistakenly use tools like ChatGPT as search engines, leading to the dissemination of fabricated plausible facts and fake bibliographic references. 

The need for transparency and rigorous benchmarking of AI tools is evident. Many tools lack systematic access, detailed model information and evaluation results, making it challenging for lawyers - or anyone - to assess their reliability. While AI advancements are impressive and herald a new era of technological innovation, the persistent issue of accuracy and reliability cannot be ignored. If we accept that current state-of-the-art AI systems cannot be entirely accurate, we would have to reconsider how we deploy them. They are excellent idea generators but not independent problem solvers. We shouldn’t allow ourselves to forget that.

If you want to know more about the challenges of AI governance, the regulation of synthetic media, and the global security implications of AI advancements, join us on Discord at https://discord.gg/2fR2eZAQ4a. Here, we can collaborate, share insights, and contribute to shaping the future of AI in a manner that safeguards our security and democratic values and fosters responsible innovation.

AI technology is advancing rapidly, with major announcements from leading tech companies showcasing their latest developments. Google recently announced plans to launch an AI assistant capable of seeing and hearing. Microsoft is unveiling Copilot, an AI system set to provide a powerful AI platform for developers to accelerate local AI development. Apple, too, is expected to focus heavily on AI in its upcoming Worldwide Developers Conference. Advanced AI is no longer a futuristic hypotesis but a present reality, but this new reality comes with significant challenges, particularly the phenomenon known as AI ‘hallucinations’.

AI hallucinations occur when generative AI models produce responses that do not match reality. In artificial intelligence, a hallucination refers to a response generated by an AI that contains false or misleading information presented as fact. This occurs because AI systems, especially large language models (LLMs) like ChatGPT, predict text based on their training data, which can lead to plausible-sounding but incorrect outputs. 

While they excel at identifying patterns and generating information, they can also create incorrect or misleading outputs. This issue is well-known in the AI community, with some experts relatively optimistic about future (though not present) solutions. Others believe these errors - that’s what we’re talking about, errors - are an inevitable part of AI. Sundar Pichai, Google’s CEO himself, admitted that hallucinations are an "inherent feature" of  LLMs, the AI technology that drives most AI systems nowadays, and that it still is “an unsolved problem”.

The implications of AI hallucinations are far-reaching, affecting both mundane and critical areas. For instance, Google’s AI Overviews tool, introduced in the U.S. this month, faced significant backlash for spreading bizarre misinformation. One notable incident involved a satirical query asking, ‘How many rocks should I eat?’ The AI tool responded by recommending ‘eating at least one small rock per day’, falsely citing UC Berkeley geologists. On X, internet analyst Jeremiah Johnson shared a thread featuring even more bizarre responses, which include being able to train eight days a week or that 1919 was only 20 years ago.

While the examples of AI hallucinations may initially seem amusing, the implications can be gravely serious, especially in fields that demand high precision, such as law and medicine. In the legal domain, nearly three-quarters of lawyers are considering using generative AI for tasks like document review, contract drafting, and legal memoranda writing. A notable case involved a New York lawyer facing sanctions for citing fictitious cases generated by ChatGPT. These false responses are not only inaccurate but potentially very dangerous, particularly when we consider the latest Stanford report, which reveals that general-purpose chatbots hallucinate 58% to 82% of the time on legal queries. In the medical field, the stakes are even higher. AI has the potential to revolutionise healthcare by expediting diagnostic processes, personalising treatments, and easing bureaucratic tasks for overburdened staff. However, AI's struggle to differentiate true from false information poses significant risks. AI hallucinations can mislead even seasoned doctors if they mistakenly use tools like ChatGPT as search engines, leading to the dissemination of fabricated plausible facts and fake bibliographic references. 

The need for transparency and rigorous benchmarking of AI tools is evident. Many tools lack systematic access, detailed model information and evaluation results, making it challenging for lawyers - or anyone - to assess their reliability. While AI advancements are impressive and herald a new era of technological innovation, the persistent issue of accuracy and reliability cannot be ignored. If we accept that current state-of-the-art AI systems cannot be entirely accurate, we would have to reconsider how we deploy them. They are excellent idea generators but not independent problem solvers. We shouldn’t allow ourselves to forget that.

If you want to know more about the challenges of AI governance, the regulation of synthetic media, and the global security implications of AI advancements, join us on Discord at https://discord.gg/2fR2eZAQ4a. Here, we can collaborate, share insights, and contribute to shaping the future of AI in a manner that safeguards our security and democratic values and fosters responsible innovation.

Safeguard our future

Sign up to keep updates on our current and future campaigns.