Generative AI chatbots are known to make mistakes, including hallucinations where the model makes up information. Researchers are not optimistic about the technology improving soon, according to a panel of AI experts. The reliability and trustworthiness of AI tools need to increase significantly for them to surpass human intelligence. Issues of bias, unintended consequences, and accuracy must be addressed. Despite recent rapid advancements in AI, progress in accuracy and reliability is still needed.
Improving the accuracy of AI systems involves fine-tuning, retrieval-augmented generation, and chain-of-thought methods. Researchers are doubtful that factuality concerns will be solved soon, with 60% expressing uncertainty. Optimism about scaling up existing models to reduce hallucinations may be overly optimistic, with no evidence of highly factual language models on the horizon. AI users need to be aware of the limitations of generative AI models and should verify results independently.
The quest for artificial general intelligence (AGI), which would surpass human-level thought, is a priority in AI development. Most researchers believe that scaling up current AI techniques will not lead to AGI. There are concerns about the ethics and potential downsides of creating systems that can outthink humans. The majority of researchers feel that systems capable of AGI should be publicly owned, and research should continue with safety and control systems in place.
Large language models like ChatGPT are seen as a step toward AGI, but researchers believe that true AGI is still far off. The current technology lacks the ability to handle truly open-ended creative tasks and operate robustly in a human environment. The definition of AGI varies, with some researchers defining it as surpassing human capabilities in most tasks. While researchers are doubtful about the near-future development of AGI, they acknowledge the rapid technological advancements in recent years and the unpredictability of future progress.
