SCIENTIFIC COLLECTIVE and ARTIFICIAL INTELLIGENCE
AI and Collective Intelligence
The recent progress in Artificial Intelligence provides both challenges and opportunities for scientific collective intelligence. Among the most notable examples of AI progress are ChatGPT and other Large Language Models, which have shown unexpected capabilities (Wei 2022, Mitchell 2023), and AlphaFold, which can predict the 3D shape of proteins from their genetic sequence with unprecedented accuracy (Jumper 2021). The development of AlphaFold has been recognized by the award of the 2014 Nobel Prize in Chemistry.
Challenges
AI poses specific challenges for science. There are many reports of errors in statements from ChatGPT and from other AI systems. The types of errors and blind spots seem different from those more common in humans.
These AI systems consist of neural networks with billions to trillions of parameters (Mitchell, 2023) and it is therefore not possible to provide a simple explanation for their outputs.
Another concern is that the most powerful AI systems are currently privately owned and not as transparent as they could be. One of the topics emerged from our discussion is the potential benefit for society of a public or nonprofit AI effort with the same scale and level of funding as the current large private efforts. Many comments have pointed out that, if the most advanced science were to be done only in the private sector, the lack of transparency will decrease trust in science, support for academic research will decline and society will not be able to fully benefit from the opportunities provided by AI in science.
Opportunities
There is wide support for the view that human intelligence evolved in response to intellectual challenges, possibly posed by social interactions. AI systems can be an opportunity that will stimulate our collective intelligence.
There are several examples of major challenges that have promoted collective scientific efforts. Among these are:
-
The WW2 effort at Bletchley Park, where Alan Turing played a key role; this effort led to the first purely electronic digital computers, responding to cryptography advances by Germany.
-
The Manhattan Project, prompted by the discovery of nuclear fission by German scientists and leading to advances in nuclear physics, for both military and peaceful applications.
-
The creation of NASA that led to the Moon landing, sparked by the Sputnik launch and the space race with the Soviet Union.
AI can make large scale discussions possible by finding new ways to connect individuals and ideas. This could be an iterative process in which human scientists can decide which avenues to pursue and provide novel contributions.
In both artificial and natural neural networks some capabilities emerge when a certain size is reached. In both cases (AI and brain) other properties like connectivity are likely to synergize with size (Wei 2022, Tattersall 2023). It is an open question what properties might emerge in the case of collective intelligence as it grows further.
According to a survey from the Pew Research Center (Pew 2024), 76% of US adults express a great deal or fair amount of confidence in scientists to act in the public’s best interests. Scientists are held in higher regard than many other prominent groups, including elected officials, journalists and business leaders. Society is therefore likely to take seriously their views about AI, if these are reached after an open debate. An example of the effectiveness of transparent, community-level consultations among scientists in increasing support from the public and from funders is the Snowmass process in particle physics. Other fields are also adopting similar processes, as shown by the Decadal Survey on Astronomy and Astrophysics.
​
A fundamental aspect of scientific collective intelligence is the communication of ideas among scientists.
Renato Dulbecco shared his memories about a more open time in biomedical scientific communication. Multiple historical sources confirm his statements, and show that scientific habits that might seem immutable do change, responding to the changing world (using Dulbecco's words). Dulbecco also pointed out that scientific communication does depends more on the motivations of individuals that on technical means.
The section on Science Incentives outlines a strategy for the scientific community to self-motivate an open discussion about AI in science.
​
Discussion platforms can be biased and platforms for scientific collective intelligence would benefit from mechanisms that ensure trust.
In the case of this discussion, we intend to establish two oversight groups:
-
An oversight group composed of younger scientists, with representatives nominated by associations of postdocs and graduate students. Trainees do not have a long-term link with a particular institution, are more likely to rapidly learn new AI techniques and are often more open to innovative ideas.
-
An oversight group composed of current and former scientific leaders. Some of these are already involved in the current discussion and have participated by sharing ideas and by conducting interviews.
These two groups will serve as a system of checks and balances to ensure that the discussion is run for the benefit of the entire scientific community and of human society. They will determine their own internal structures and procedures.
​
An example of the benefit of involving both these groups of scientists can be found in the history of AI in science. The development of AlphaFold was only possible because of the data contained in the PDB (Protein Data Bank). The PDB started in 1971 as a grassroots proposal from a group of young scientists, supported by scientific leaders like Walter Hamilton and Max Perutz (Berman 2008, Strasser 2019). The history of PDB also provides an example of many potential obstacles for scientific sharing and collaboration (Barinaga 1989), which were in this case eventually overcome.
​​
REFERENCES
- Barinaga, M., 1989. The missing crystallography data. Science, 245(4923), pp.1179-1181.
- Berman, H.M., 2008. The protein data bank: a historical perspective. Acta Crystallographica Section A: Foundations of Crystallography, 64(1), pp.88-95.
- Jumper, J., et al., 2021. Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), pp.583-589.
- Mitchell, M. and Krakauer, D.C., 2023. The debate over understanding in AI’s large language models. Proceedings of the National Academy of Sciences, 120(13), p.e2215907120.
- Pew Research Center, 2024. Public Trust in Scientists and Views on Their Role in Policymaking.
- Strasser, B.J., 2019. Collecting experiments: Making big data biology. University of Chicago Press.
- Tattersall, I., 2023. Endocranial volumes and human evolution. F1000Research, 12.
- Wei J et al. 2022. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682.
​
​