Future of AI: Predictions and Concerns from Experts

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Jan 11 2024

In a recent submission to the ArXiV* server, researchers surveyed 2,778 artificial intelligence (AI) experts about the future of AI. Predictions suggest a 50% chance of AI achieving milestones like independently creating payment processing sites and composing songs by famous artists by 2028.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.

There's a shift in estimates, with a 10% chance of machines surpassing humans in all tasks by 2027 and a 50% chance by 2047, 13 years earlier than previously thought. Uncertainty looms, with varying opinions on positive or catastrophic AI outcomes. Despite differing views on the pace of AI progress, there's unanimous agreement on prioritizing research to minimize potential risks AI systems pose.

Related Work

The transformative potential of artificial intelligence is prompting attention from decision-makers across sectors—private, academic, and governmental levels. Despite the impact AI is poised to make, experts have no consensus on its future. Predicting this landscape involves amalgamating various methods, from trend extrapolation and analyzing historical events to understanding current AI capabilities and applying economic models.

While AI researchers provide valuable insights due to their familiarity with the field, their predictions, gathered in the expansive survey of 2,778 experts from top AI forums, offer just one facet of the complex puzzle. Conducted in 2023, after notable AI advancements and increased societal awareness, this survey delved into AI progress dynamics and the societal implications of advanced AI.

AI Progress and Predictions

The survey revealed that experts anticipated most of the 39 AI-related tasks to be achievable within a decade, foreseeing economically valuable endeavors like creating payment processing sites and composing songs indistinguishable from those by well-known artists within the next 10 years. However, a few complex tasks, such as deducing differential equations from virtual world experiences or solving longstanding mathematical problems, were estimated to take up to 27 years. Notably, comparisons with the 2022 survey indicated a shift towards earlier predictions for 21 out of 32 tasks, highlighting a trend towards faster expectations in AI progress.

Regarding the advent of high-level machine intelligence (HLMI), experts predicted a 50% chance by 2047, marking a drastic shift of 13 years from the 2022 survey. Similarly, experts revised their estimation for a 10% chance of achieving HLMI by 2027, lowering it by two years from the previous forecast. In contrast, experts anticipated a substantial decrease from the 2022 prediction of 2164 for complete automation of labor (FAOL), with a 50% chance expected by 2116, marking a 48-year reduction. Despite their conceptual similarities, the disparity between forecasts for HLMI and FAOL puzzled researchers, raising questions about framing effects and interpretations of these milestones.

Demographics played a role in predictions, with respondents from different regions foreseeing variations in the timeline for HLMI. Moreover, participants generally believed the pace of progress in AI increased in the latter half of their careers. The survey also revealed that experts considered declines in computing costs crucial. Yet, they acknowledged that various factors, such as researcher effort and progress in AI algorithms, significantly contributed to advancements in AI. Furthermore, respondents held mixed opinions about the likelihood of an intelligence explosion post-HLMI, exhibiting considerable uncertainty despite maintaining consistent views across surveys since 2016.

AI-Related Scenario Analysis

The survey delved into various AI-related scenarios over the next thirty years, revealing notable concerns among respondents. More than 30% expressed substantial worry about scenarios like AI-enabled misinformation and the use of AI by authoritarian figures. Deepfake-based false information and AI-driven control by rulers particularly stood out, with over 70% considering them highly concerning.

Respondents showed diverse perspectives regarding high-level machine intelligence (HLMI) and its long-term impact. While 68.3% leaned towards positive outcomes, 57.8% saw the possibility of extremely negative consequences, like human extinction. Even optimists acknowledged a 5% or higher likelihood of dire outcomes, mirroring the tendencies among pessimists for positive outcomes. The median prediction for such extreme scenarios settled at 5%, indicating the broad spectrum of expert opinions.

Concerns about AI causing human extinction were consistent across responses, with median probabilities of 5% or 10%. Between 41.2% and 51.4%, estimated probabilities higher than 10% for scenarios involving human extinction or severe disempowerment align with previous assessments. Despite variations in question presentations, it underscored a substantial segment viewing extreme outcomes as non-negligible possibilities.

Experts universally acknowledged the significance and challenge of the alignment problem in AI compared to other issues. However, while a majority recognized its gravity, there needed to be unanimous agreement on its immediate prioritization, underscoring the field's diverse viewpoints and priorities.

AI Future Survey

The survey, conducted in October 2023, gathered perspectives on AI's future from individuals recently published at top-tier AI venues. Questions covered AI progress timing, future impact, and safety. It followed ethical approval from the University of Bonn and underwent preregistration. Most questions mirrored those from previous surveys in 2016 and 2022, with modifications and additional queries based on AI researcher interviews.

The survey's flow, randomization, and recruitment targeted authors from key AI conferences and journals, reaching over 20,000 individuals. The survey ran on the Qualtrics platform, employing incentives to boost response rates. Researchers obtained a 15% response rate, garnering 2,778 responses from more than 18,000 functioning email addresses. Data cleaning involved addressing errors like non-numerical answers, leading to an analysis using various software tools including R, statistical package for the social sciences (SPSS), Google sheets, and creately.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as definitive, used to guide development decisions, or treated as established information in the field of artificial intelligence research.