Combatting AI Hallucinations and Falsified Information

AI hallucinations are a phenomenon where large language models, like generative AI and chatbots, produce patterns or objects that are incorrect, nonexistent, or entirely fictional. You may have encountered these anomalies in AI-generated images that are meant to, but don't quite represent reality, or in fabricated news stories. While some AI hallucinations may seem minor or even amusing, others can have significant and damaging consequences, hindering the technology’s overall effectiveness. Over time, the persistent spread of misinformation can jeopardize public resources, damage reputations, and undermine our collective knowledge base worldwide. 

What are AI Hallucinations? 

AI hallucinations are Large Language Models (LLMs) that gather extensive text data from websites, articles, and other sources. This data is cleaned and normalized, ensuring the data is ready to train the LLM algorithm. Finally, the model is fed the data, where it uses neural networks to learn patterns and structures within the data. When prompted, the model responds by calling on this learned knowledge to provide a response, thus having accurate data is critical to ensuring the LLM is properly trained and can provide informed outputs. 

AI hallucinations may occur for a variety of reasons. One reason is that, during training, LLMs can become too specialized, which leads to produced responses that don’t align with the intended prompt. Also, if an AI model is trained on biased or unrepresentative data, it may “hallucinate” patterns or features reflecting those biases. And since many models are incredibly complex, they may struggle to provide more general responses if prompted to do so.  

Real- World Examples of AI Hallucinations   

In its first public demonstration, Google’s ChatGPT-type tool Bard stated that the James Webb Space Telescope recently “took the very first pictures of a planet outside of our own solar system,” even though that photo was taken 16 years prior. This error was noted and reported by astronautical professionals, leading to a $100 billion drop in market value for Alphabet, Google’s parent company. 

When two lawyers used ChatGPT to prepare a legal brief, the model invented several court cases that were intended to serve as legal precedents. When the judge could not verify the cited cases, the law firm was fined $5,000, and it undermined confidence in the legal team’s defense. 

Shortly after its launch, Microsoft’s Bing AI tool threatened to harm users and claimed to be spying on the company’s employees through webcams. There are also several reported instances of this Bing model fabricating information related to world events, celebrities, and stats and scores of professional sports.  

The Chronicle of Higher Education reported that a university librarian was asked to produce articles from a list of references a professor provided. When she concluded the articles did not exist, the professor revealed that ChatGPT had provided them. In academia, researchers are finding that generative AI understands the form of what a good reference should look like, but that doesn’t mean that the articles actually exist. ChatGPT can create convincing references with coherent titles attached to authors who are prominent in the field of interest, and studies have found that up to 47% of ChatGPT references are inaccurate. 

Consequences of AI Hallucinations

This type of hallucinated information can spread quickly and lead to public confusion, especially if shared on social media. When this false information is generated, spread, and believed, it can create financial loss and reputational damage for companies, governments, and individuals. It also creates a culture of “fake news” by misleading the public with incorrect facts which could have lasting, long-term effects. 

Hallucinations in medical AI systems, for example, can erode public trust in AI-driven healthcare solutions, slowing down technological adoption. In legal cases, it can undermine trust in the judicial process and lead to potential legal liabilities. Investors relying on AI-generated financial analysis might make poor investment decisions based on incorrect data. Companies that create or use hallucinated outputs may face financial loss and unintentionally violate regulations. Relying on AI for operational decisions may waste resources, produce inaccurate analyses, and affect efficiency. Educators and students that cite hallucinated information may perpetuate flawed research and face serious consequences like job termination or expulsion. These hallucinations can also create misunderstandings and conflicts in personal or professional relationships, as well as reduce trust in AI systems, creating barriers for the improved evolution of this technology. 

Combatting AI Hallucinations 

The frequency and severity of harmful AI hallucinations underscores the importance of careful design, rigorous testing, and ongoing monitoring of AI systems. As the tech industry strives to balance innovation with reliability, there is a focus on reducing hallucinations by using high-quality, current, and diverse training data. Scientists are exploring a new algorithm that can detect AI hallucinations 79% of the time, roughly 10% more than leading models, using semantic entropy – a measure of uncertainty or unpredictability in the meaning of data. 

Frequent and thorough testing and evaluation can help uncover performance issues and biases in AI-generated data outputs. Interpretable models – those that attempt to explain what happens in AI’s “black box” or core, hidden algorithm system – may provide explanations for these outputs. Partnering with vendors committed to ethical development and practices in AI can help create more transparent models that can be updated and adjusted effectively. And strengthening prompts with a process called retrieval augmented generation can improve outputs by limiting the scope of information accessed to a more manageable, reliable frame. 

Humans play an integral role in this quality assurance process. Incorporating human oversight by qualified AI specialists and professionals during model development and deployment can help in validating outputs, providing corrective feedback, and catching AI hallucinations. Open lines of communication with users can also contribute to model improvement by welcoming their insights as live users or beta users.  

Perhaps most importantly, it is crucial that users recognize the limitations of AI models. While these models may produce plausible-sounding responses, it is the responsibility of the user to fact-check and exercise judgment when evaluating or disseminating these outputs. Logic and critical thinking are needed when interacting with AI models and information should be verified even from reliable sources by checking for coherence and consistency in responses.

Exploring AI at Capitol Tech 

Capitol Technology University is the first university in Maryland to offer a B.S. in Artificial Intelligence degree, equipping students with the skills needed to build large language models and understand their limitations and social implications. Our AI Center of Excellence offers further opportunities for students to join like-minded peers and industry experts in their mission to advance the frontiers of artificial intelligence.  

Request information or contact our Admissions team to learn more.