Introduction
Many domains of knowledge have grappled with the arrival of hypercompetent artificial intelligence. We might expect an efflorescence of AI-produced ideas. On the other hand, human-trained AI might simply reproduce or reinforce existing orthodoxies. The strategic impact of AlphaGo post-2016 on the game of Go serves as a data-rich case study.
<br>Methods
Using a large database of top-level human and AI play (62,740 opening sequences, 1956-2020), I study the diversity of opening sequences using multilevel regression models of recent social and individual covariates pre & post-AlphaGo.<br>
Results
The post-AI era has indeed seen an explosion of move diversity (Shannon entropy on opening sequences), reversing decades of decline. Since the 1990s, social trends have come to strongly predict individual first-move openings (b=3.77, psign<0.001), and, while pre-internet Go was characterized by a low correlation between strategy popularity and efficacy (R2=0.11), the post-AI era now sees a near-perfect correlation (R2=0.87).
<br>Conclusions
Human Go strategies are more diverse, more social, and apparently more efficacious under the influence of hypercompetent AI advisors.

Strategic Social Leaning in the Era of Hypercompetent Artificial Intelligence: A Large-Scale, Longitudinal Case Study using the Game of Go

Low rates of data availability and results reproducibility have been observed across the biological and behavioural sciences. Here, we attempt to quantify the rate of reproducibility in the study of social learning research. We surveyed 560 empirical publications on this topic, published between 1955 and 2018 in animal behaviour, behavioural ecology, cultural evolution, and evolutionary psychology.

Data was recoverable online or through direct data requests for 30% of this sample. Data recovery declines exponentially with time since publication, halving every 6 years, and up to every 9 years for human experimental data. Essentially no data is available for any publication before the year 2000. However, when data for a publication is available, we estimate by hierarchical modeling a high probability of subsequent data usability (87%), analytical clarity (97%), and agreement of published results with reproduced findings (96%), giving an overall reproducibility rate of any given result in the literature of 23% [0.89 HPDI: 15-28%].

Our results suggest that the primary threat to the reproducibility of published results in cultural evolution is the lack of long-term data archiving, rather than mistakes in analyses or results. Although recent innovations in research transparency, such as the FAIR data guidelines and use of version-controlled analysis code, address these failure points, individual adoption of these novel research practices, as well as incentives promoting data reuse and openness at the funding and publication level, must become more widespread to prevent similar decay of social learning results in the future.

Bret Beheim

2

3

SHORT BIO

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES