In a world increasingly fascinated by the possibilities of artificial intelligence (AI), a recent study by researchers at the University of California, Berkeley provides a nuanced view of its capabilities and limitations. Details of their findings can be found at Psychological science perspectivesuggests that despite AI’s great ability to generate text and images, it falls short of its ability to innovate, a key area that defines human intelligence.
This research was motivated by rapid advances in AI, particularly large-scale language models like OpenAI’s ChatGPT. These systems have demonstrated incredible capabilities, from creating compelling narratives to creating complex visual art.
However, researchers argue that it can be misleading to view these AI systems as individually intelligent sentient agents. Instead, they suggest thinking of AI as a powerful new form of cultural technology, similar to writing and the internet, that greatly enhances access to and transfer of knowledge.
This research revolves around two main components: an “imitation” component and an “innovation” component, each of which captures the ability to recognize traditional uses of objects and the ability to innovate new uses of them. Tailored to test.
In the “imitation” part of the study, participants were presented with objects and asked to choose the one that best complemented a particular tool based on their understanding of the traditional relationships between the objects. This component is designed to assess your ability to recognize and replicate existing knowledge and associations. We expect AI systems trained on large datasets to excel at this task, as they become adept at identifying patterns and correlations.
However, the “innovation” component posed a more significant challenge. Participants were given a problem-solving scenario in which they had to accomplish a goal without the typical tools at their disposal. Instead, alternative objects were provided, some of which could achieve the objective due to their functional properties despite their superficial differences. This test was important in assessing the ability to go beyond traditional associations and apply creative thinking to utilize objects in new ways to solve problems.
Participants in this study included children and adults ages 3 to 7 and included a wide range of human cognitive abilities. These human participants were contrasted with several state-of-the-art AI models, including, among others, OpenAI’s GPT-4. To ensure a fair comparison, the AI model was shown a textual description of the scenario, mimicking the input given to the human participants.
The findings reveal significant differences in the capabilities of humans and AI. “Even young human children can generate intelligent responses to specific questions. [language learning models] We can’t,” study author Eunice Yu explained. “Rather than viewing these AI systems as intelligent agents like us, we can think of them as new forms of libraries or search engines. They will let us know.”
Both children and adults showed great innovation abilities, including selecting functionally related but superficially dissimilar objects to solve the problem at hand. This demonstrates the ability to not only recognize traditional associations, but to innovate beyond them and apply abstract thinking to recognize and exploit the potential functional properties of objects.
“Children can imagine completely new ways to use things they’ve never seen or heard of before, such as using the bottom of a teapot to draw a circle,” Yui says. “It’s very difficult to generate such a response in large models.”
Conversely, while AI models were good at identifying superficial relationships between objects (the imitation component), they showed notable deficiencies in the innovation component. When tasked with selecting objects for new uses, AI systems often defaulted to traditional associations and lacked the human-like ability to reason about new functional applications for these objects.
This was particularly evident in the inability to select unconventional but functionally appropriate objects for the task, highlighting a fundamental gap in AI’s ability for innovative problem solving.
These results highlight the limitations of current AI systems in mimicking the full spectrum of human cognitive abilities, especially when it comes to innovation. While AI can replicate known patterns and relationships with remarkable efficiency, its ability to forge new paths and imagine uncharted applications of existing knowledge remains challenged.
“AI helps convey information that is already known, but it is not an innovator,” Yiu says. “These models can summarize conventional wisdom, but they cannot extend, create, change, abandon, evaluate, and improve conventional wisdom the way younger humans can.”
The study, “Communication vs. Truth, Imitation vs. Innovation: Children Can Do What Large-Scale Language and Language-Visual Models Can’t (Yet)”, was written by Eunice Uy, Eliza Kosoy, and Alison Gopnik. Ta.
