Generative artificial intelligence has come a long way since its early days, where laptops would laboriously churn through code to produce quirky outputs like Halloween costumes made of shredded bourbon and chopped water. Optics researcher Janelle Shane has been documenting this evolution on her blog AI Weirdness, showcasing the progress from endearingly bumbling to surprisingly coherent and sometimes hilariously wrong AI-generated content.
In a recent interview with Scientific American, Shane discusses how generative AI has evolved over the years. She notes that there is now more commercial interest in AI compared to when she first started exploring it. Google Translate was one of the early commercial applications that hinted at the potential of machine-learning AI techniques. However, she emphasizes that despite the advancements, the tendency for people to assign deeper meaning to AI-generated text remains prevalent.
Shane’s blog focuses on the differences between how AI generates text and how humans write, highlighting the unexpected and novel aspects of AI-generated content. She points out the importance of glitches and mistakes in AI-generated text, such as the case of a spotless giraffe that exposed the limitations of image-labeling algorithms.
While Shane acknowledges the advancements in generative AI, she also highlights the persistent errors and glitches that have been present since the beginning. Despite the growth and changes in AI technology, the fundamental flaws in these algorithms remain.
One of the key concerns raised by Shane is the potential dangers of relying on generative AI for accurate information. She warns against using AI for critical tasks where accuracy is crucial, as the algorithms are trained to sound correct rather than retrieve accurate information.
Moreover, Shane discusses the issue of bias in generative AI, noting that while efforts have been made to reduce surface-level biases, the underlying biases in the input data still influence the outputs.
In terms of the hype surrounding generative AI, Shane believes that there are many practical applications of AI techniques that are often overshadowed by the focus on generative AI. She highlights the success of AI in drug-discovery research and other areas where a degree of inaccuracy is acceptable.
Overall, Shane emphasizes the importance of recognizing the limitations of generative AI and focusing on practical applications where AI can provide value. While generative AI has made significant strides, it is essential to approach its use with caution and awareness of its inherent flaws.