Ai

THOUSANDS OF AI AUTHORS ON THE FUTURE OF AI PREPRINT

Katja Grace, Harlan Stewart, Julia Fabienne Sandkühler, Stephen Thomas, Jan Brauner

In the largest survey of its kind, we surveyed 2,778 researchers who had published in top-tier artificial intelligence (AI) venues, asking for their predictions on the pace of AI progress and the nature and impacts of advanced AI systems. The aggregate forecasts give at least a 50% chance of AI systems achieving several milestones by 2028, including autonomously constructing a payment processing site from scratch, creating a song indistinguishable from a new song by a popular musician, and autonomously downloading and fine-tuning a large language model. If science continues undisrupted, the chance of unaided machines outperforming humans in every possible task was estimated at 10% by 2027, and 50% by 2047. The latter estimate is 13 years earlier than that reached in a similar survey we conducted only one year earlier [Grace et al., 2022]. However, the chance of all human occupations becoming fully automatable was forecast to reach 10% by 2037, and 50% as late as 2116 (compared to 2164 in the 2022 survey). Most respondents expressed substantial uncertainty about the long-term value of AI progress: While 68.3% thought good outcomes from superhuman AI are more likely than bad, of these net optimists 48% gave at least a 5% chance of extremely bad outcomes such as human extinction, and 59% of net pessimists gave 5% or more to extremely good outcomes. Between 37.8% and 51.4% of respondents gave at least a 10% chance to advanced AI leading to outcomes as bad as human extinction. More than half suggested that "substantial" or "extreme" concern is warranted about six different AI-related scenarios, including spread of false information, authoritarian population control, and worsened inequality. There was disagreement about whether faster or slower AI progress would be better for the future of humanity. However, there was broad agreement that research aimed at minimizing potential risks from AI systems ought to be prioritized more.

Background and context:

The paper presents the findings of a large-scale survey conducted in 2023 to gather insights from AI researchers on the future progress of AI systems. The survey builds upon previous iterations conducted in 2022 and 2016, allowing for longitudinal analysis of how AI researchers' views have evolved over time.

Research objectives and hypotheses:

The primary objectives of the study were to:

1) Assess AI researchers' predictions on when various AI milestones and capabilities would be achieved.

2) Explore AI researchers' perceptions of potential risks and impacts associated with advanced AI systems.

3) Investigate the factors influencing AI researchers' judgments and the uncertainty surrounding their forecasts.

Methodology:

The survey was distributed to individuals who had recently published at six top-tier AI conferences. Respondents were asked a range of questions on the timeline of AI progress, the feasibility of "Full Automation of Labor" (FAOL), the potential characteristics of future AI systems, and the risks and impacts associated with advanced AI. The survey used various framing approaches, including fixed-years and fixed-probabilities, to probe the robustness of responses. Randomization was used to ensure that each participant only received a subset of the questions.

Results and findings:

The key findings of the survey include:

1) The majority of the 39 AI tasks assessed were predicted to have at least a 50% chance of being feasible within the next 10 years, including valuable tasks such as creating a payment processing site and generating songs indistinguishable from those by popular musicians.

2) There was a substantial gap (around 70 years) between the predicted timelines for achieving human-level machine intelligence (HLMI) and full automation of labor (FAOL), suggesting a potential framing effect.

3) Respondents expressed significant concerns about potential risks associated with advanced AI, with more than 30% considering several scenarios (e.g., spread of false information, authoritarian population control) as worthy of "substantial" or "extreme" concern.

4) The survey found a wide range of views among respondents, with some exhibiting highly optimistic or pessimistic perspectives on the future of AI.

Discussion and interpretation:

The authors acknowledge the inherent challenges in accurately forecasting the future of AI, given the complex and unpredictable nature of technological progress. They note that the AI researchers surveyed, while experts in their field, may not possess the same level of general forecasting skills and experience as domain-general experts. The authors also highlight the potential for framing effects to influence the responses, as seen in the divergent predictions for HLMI and FAOL.

Contributions to the field:

This survey provides a comprehensive and up-to-date snapshot of AI researchers' perspectives on the future of AI, building upon previous work in this area. The findings offer valuable insights into the current state of expert opinion, which can inform ongoing discussions and decision-making around the development and governance of advanced AI systems.

Structure and flow:

The paper is well-structured, with clear delineation of the background, research objectives, methodology, results, and discussion. The various sections flow logically, and the presentation of the findings is comprehensive and easy to follow.

Achievements and significance:

The study represents a significant contribution to the field, as it constitutes the largest survey of AI researchers to date on the topic of AI progress and risks. The longitudinal nature of the research, with comparisons to previous surveys, adds to the richness and depth of the insights provided.

Limitations and future work:

The authors acknowledge several limitations of the study, including the potential for participation bias and the reliance on expert judgments, which may be subject to various cognitive biases. They suggest the need for further research, including incorporating insights from other sources, such as trends in hardware and economic analyses, to develop a more holistic understanding of the future trajectory of AI.