A couple of weeks back, one of the members of the research team received more than 35 exam papers from their students. In these papers, the students had been asked to reflect on potentials and potential pitfalls of implementing generative AI in a university context. Most of the students identified the ability of generative AI to ‘hallucinate’ (the ability of many generative AIs to make entirely fabricated or flat-out wrong responses) as the biggest problem when implementing or working with AI. This problem has also been identified by Elon Musk, Steve Wozniak, Yuval Noah Harari, and the co-signatories of ‘Pause Giant AI Experiments: An Open Letter’.
We decided to pose the following questions to ChatGPT: Is the problem of hallucinations in AI real? If so, what should we as individuals and the societies we live in do about it? And what about the companies that develop the systems?
It provided the following answer:
“The phenomenon of AI models “hallucinating” or generating incorrect or nonsensical outputs is indeed a significant concern in the field of generative AI. While these models have demonstrated remarkable capabilities, they can still produce responses that are factually incorrect, incoherent, or irrelevant to the input.
Addressing this problem requires a multi-faceted approach:
Improved Training Data: Enhancing the quality and diversity of training data can help mitigate hallucination issues. Models trained on more comprehensive and accurate datasets, which are carefully curated and annotated, are likely to generate more reliable outputs. Additionally, incorporating data from trusted sources and expert human reviewers can improve the accuracy and reliability of AI-generated content.
Reinforcement Learning from Human Feedback: By incorporating human feedback, AI models can learn from their mistakes and improve over time. Techniques such as reinforcement learning from human trainers or using human reviewers to rate and evaluate generated outputs can help AI systems better understand and generate appropriate responses. This iterative feedback loop allows the models to refine their understanding and reduce hallucinations.
Contextual Understanding and Reasoning: Enhancing AI models’ ability to comprehend context and reason logically can help minimize hallucinations. Efforts are being made to develop models that can not only generate text but also understand the context of the conversation or prompt, interpret nuanced meanings, and apply logical reasoning to produce more accurate and coherent responses.
Ethical Guidelines and Auditing: Establishing ethical guidelines and standards for AI development and deployment is crucial to address the issue of hallucinations. Organizations must ensure transparency, accountability, and responsible usage of AI models. Regular auditing and evaluation of AI systems can help identify and rectify any potential biases, inaccuracies, or hallucinations in their outputs.
While Google’s Bard, Jasper.ai, and ChatGPT are advanced generative AI models, it is important to note that all AI models, to varying degrees, have the potential to generate incorrect or nonsensical outputs, which can be considered a form of “hallucination.” The prevalence and severity of hallucinations can depend on multiple factors, including the training data, model architecture, and specific use cases. However, it is challenging to definitively determine which model hallucinates the most, as it can vary depending on the context and evaluation criteria.
As for the future, it is reasonable to expect improvements in addressing hallucinations as AI research progresses. Ongoing advancements in training methodologies, data quality, reinforcement learning techniques, and context understanding will likely contribute to reducing the prevalence of hallucinations. However, completely eliminating hallucinations may be a challenging task, as it requires comprehensive understanding of language nuances, world knowledge, and commonsense reasoning, which are still active areas of research in the field of AI. Continuous research, development, and responsible deployment will be essential to minimize and mitigate hallucination issues in AI systems.
As an individual, there are several steps you can take to counter hallucinations from AI. Enhancing your critical thinking skills can help you evaluate information and discern between reliable and unreliable sources. Be aware that AI-generated content may not always be accurate, and critically analyze the information provided. You should also cross-reference AI-generated content with trusted and reputable sources to ensure accuracy and fact-check claims made by AI systems. Furthermore, you must educate yourself about the limitations and potential biases of AI models. Understand that AI models may sometimes generate inaccurate or exaggerated responses and be cautious when accepting information without independent verification.”
What do you think of the problem of ‘hallucinations’ and the answer provided by ChatGPT? Leave your thoughts in the comments section below.