A new study explores how artificial intelligence can not only better predict scientific discoveries but also effectively extend these discoveries. The researchers published their findings in the journal "Nature Human Behaviour," where they developed a model that can predict human inferences and the scientists who make these inferences.
The authors also created a model to generate scientifically plausible "alien" hypotheses that may only be considered in the distant future, if at all. They believe that these two demonstrations - the first allowing for accelerated human discoveries and the second identifying and surpassing blind spots - imply that artificial intelligence with human consciousness will enable advancements beyond the current scientific frontier.
"If you understand what people are doing, you can improve the accuracy of predictions and surpass them to accelerate scientific discoveries," explained co-author James A. Evans, a professor in the Department of Sociology and director of the Knowledge Lab. "You can also figure out what people currently cannot do or what they cannot do for decades or even longer in the future. By providing this complementary intelligence, you enhance their capabilities."
Artificial intelligence models trained on published scientific discoveries have been used to invent valuable materials and targeted therapies, but they often overlook the contributions of the human scientists involved. The researchers took into account the competitive and collaborative nature of human research throughout history, so they wondered if we could better supplement collective human abilities by pursuing and exploring areas that humans have yet to explore, with artificial intelligence programs explicitly aware of human expertise.
Predicting the Future of Discoveries
To test this question, the team first simulated the inference process by establishing a random walk in the research literature. They started with an attribute, such as the COVID vaccine, and then jumped to a paper with the same attribute, followed by another paper by the same author or one that cited the initial paper.
Running millions of these random walks, their model improved predictions of future discoveries by 400% compared to predictions that only focused on the content of the research, especially in cases where relevant literature was scarce. They could also predict the actual individuals who would make these discoveries with over 40% accuracy because the program knew that the predicted individuals were among the few with experiences or relationships related to the discussed attributes and materials.
Evans referred to this model as a "digital surrogate" of the scientific system, which can simulate possible scenarios within the system and experiment with alternative possibilities. He explained that this highlights the close relationship between scientists and the methods, attributes, and personnel they possess.
"It also gives us some information about the system and its limitations. For example, broadly, the research suggests that some aspects of our current scientific system, such as graduate education, have not been adjusted for discovery. They were set up to give people a label to help them find jobs - to fill labor market needs. They have not been optimized to maximize new, technology-related discoveries. To do that, every student could be an experiment - crossing new gaps in disciplinary knowledge," he said.
In the second demonstration of the paper, they asked the artificial intelligence model not to make predictions that were most likely to be discovered by humans but to identify scientifically feasible predictions that were least likely to be discovered by humans.
The researchers considered these as "alien" or "complementary" inferences, referring to predictions or hypotheses that are scientifically feasible but unlikely to be discovered by humans in the near future. These inferences have three characteristics: they are rarely discovered by humans, even if discovered, the scientific system would need many years to reorganize itself to acknowledge them, and these alien inferences are often better than human inferences. They explore entirely new areas, avoiding the connections and configurations of human scientific activity, opening up new possibilities for scientific exploration.
Thoroughly Enhanced Intelligence
Evans explained that viewing artificial intelligence as an attempt to replicate human abilities - based on Alan Turing's imitation game, where humans are the standard for intelligence - does not help scientists accelerate their problem-solving abilities. Instead of artificial replication, we are more likely to benefit from the thorough enhancement of collective intelligence, he said.
"In fields like science, technology, and culture, people strive to stay ahead. When others use your ideas or technology, you survive through influence. And the best way to maximize that influence is to keep up with the mainstream. Our model supplements this bias by creating algorithms that follow signals of scientific credibility but entirely avoid cases that align with mainstream thinking," Evans said.
Rather than reflecting recent human scientific thinking, it is better to use artificial intelligence to go beyond existing methods and collaborations, expanding human capabilities to support better exploration.
Evans said, "This is about changing the framework of artificial intelligence from artificial intelligence to thoroughly enhanced intelligence, which requires more, not less, research on individual and collective cognitive abilities. As we understand more about human understanding, we can explicitly design systems that compensate for its limitations and enable us to collectively gain more knowledge."
Source: https://techxplore.com/news/2023-07-human-aware-ai-scientific-discoveries.html
Alphabet Q1 paid subscribers reached 350M, fueled by YouTube & Google One. YouTube ad revenue slightly missed at $9.88B amid Premium growth. Total Q1 revenue hit $109.9B, beating forecasts.
Anthropic releases Claude connectors for Adobe, Blender, and Ableton, enabling AI workflow automation and app control. It also sponsors the Blender Development Foundation.
Apple introduces an App Store model splitting 12-month plans into monthly installments. On iOS 26.4+, it lowers upfront costs without reducing total price, adding billing tracking.
Google launches Aadhaar digital IDs in Wallet for India, enabling secure QR/NFC verification. A phased partner rollout targets key sectors, while passport-based credentials expand globally.
YouTube tests "Ask YouTube," an AI tool blending text and video for complex queries. Initially for US Premium users, it supports interactive follow-ups, expanding Google's AI search.
Google grants US DoD classified AI access, joining OpenAI and xAI in defense deals amid Anthropic's dispute. Contractual safeguards against surveillance and lethal weapons remain unenforceable.
Mistral AI launches Workflows, a Temporal-powered orchestration platform for secure, Python-driven enterprise AI automation featuring human-in-the-loop approvals.