Thousands of AI Authors on the Future of AI
AI researchers’ forecasts about when “human-level” and superhuman AI might arrive, and how risky it could be, range from near-term optimism to deep skepticism, with some giving nontrivial odds to human extinction and others arguing current models and reasoning are far too weak for such confidence. Participants debate whether AI risk should be treated like climate change, how to define and measure “AGI,” and whether embodied intelligence or brain-scale simulations are necessary milestones. Several comments also question the survey’s methodology and biases, emphasizing that expertise in building AI does not automatically translate into reliable long-term forecasting about its societal and existential impacts.
Comparison to Climate Risk and Scientific Models
- Some see a parallel between AI-extinction worries and climate-change consensus; others argue climate models are quantitatively grounded and heavily validated, while AI risk arguments are largely qualitative and speculative.
- Commenters note climate change has clear physics and detailed models, whereas AI timelines and x-risk rely on weak or untested models and philosophical reasoning.
- One view: broad acceptance of AI existential risk will require statistical models with empirical backing—possibly arriving too late to be useful.
Survey and Expert Opinion Skepticism
- Several argue the survey mostly measures hype and prior beliefs, not reliable forecasts.
- Concerns: low response rate, skew toward junior or incentivized respondents, lack of domain expertise on specific jobs (e.g., surgery, translation, wiring), and crude task definitions.
- Others point out the survey tried to stratify by citation count and affiliation, but critics dismiss citations and affiliations as weak proxies for expertise or seniority.
Timelines, Capabilities, and Embodiment
- Some think a 50% chance of “machines outperforming humans at all tasks” by ~2047 is conservative given rapid recent progress; others suspect researchers are overestimating closeness due to Dunning–Kruger-like effects.
- Debate over whether AGI requires a body or unified embodied intelligence; some say AI and robotics must converge, others see brain simulation or abstract models as the key.
- Rough computational estimates for brain-scale simulation (neurons, synapses, ion channels) suggest human-brain-level hardware might arrive around late 2040s–2050s, but commenters highlight major unknowns in neuroscience, chemistry, and bandwidth.
Definitions: AGI vs. ASI and Task Automation
- Competing definitions:
- Very strong: a single system that can do anything any human has ever or will ever do (often labeled more like “superintelligence”).
- Moderate: human-level performance across most fields, with self-learning, robust common sense, and long-term memory.
- Some argue current systems already surpass humans in narrow domains, but fall far short of being unsupervised “remote workers” or passing long, adversarial Turing tests.
Extinction Mechanisms and Societal Outcomes
- Speculated AI-driven extinction paths include environmental devastation from unconstrained resource extraction, engineered pandemics, and pervasive manipulation of humans via autonomous agents.
- Others envision “soft” extinction through gradual integration: humans become increasingly AI-augmented until unmodified humans effectively disappear.
- Best-case future scenarios (post-scarcity, all needs met) are seen by some as utopian and by others as dystopian due to potential loss of purpose and meaning.