PaperlayerView exact revision on arXiv

Source-linked AI summary

Why AI is Harder Than We Think

Melanie Mitchell

arXiv:2104.12871v2cs.AI

TL;DR

The paper examines why AI progress and predictions about human-level intelligence remain difficult to assess. It describes four fallacies in common assumptions about intelligence and AI, arguing that these help explain why humanlike intelligence in machines is harder than expected. It concludes that better ways to describe AI and a stronger scientific understanding of intelligence remain open needs.

  • Problem

    Limited understanding of intelligence and recurring fallacies leave open how to assess progress toward human-level AI and compare task difficulty across humans and machines.

    Source: Springs and Winters · p. 3, Conclusions · p. 8

  • Method

    The paper describes four fallacies in assumptions about AI and intelligence and considers questions they raise for researchers.

    Source: Introduction · p. 2, Conclusions · p. 8

  • Results

    The author argues that these fallacies are at least partly why capturing humanlike intelligence in machines proves harder than expected.

    Source: Conclusions · p. 8

  • Takeaways & Limitations

    More effective AI assessment requires better vocabulary for machine abilities and stronger scientific understanding of intelligence across natural systems.

    Source: Conclusions · p. 8

    The sources of deep networks’ brittleness and vulnerability remain incompletely understood, and whether more layers and training data suffice is debated.

    Source: Springs and Winters · p. 3

Abstract

from arXiv · show

Since its beginning in the 1950s, the field of artificial intelligence has cycled several times between periods of optimistic predictions and massive investment ("AI spring") and periods of disappointment, loss of confidence, and reduced funding ("AI winter"). Even with today's seemingly fast pace of AI breakthroughs, the development of long-promised technologies such as self-driving cars, housekeeping robots, and conversational companions has turned out to be much harder than many people expected. One reason for these repeating cycles is our limited understanding of the nature and complexity of intelligence itself. In this paper I describe four fallacies in common assumptions made by AI researchers, which can lead to overconfident predictions about the field. I conclude by discussing the open questions spurred by these fallacies, including the age-old challenge of imbuing machines with humanlike common sense.

Introduction

The chapter links recurring AI optimism and disappointment to overconfidence rooted in mistaken intuitions about intelligence.

Introduction

  • Several prominent 2020 self-driving-car predictions and market targets went unmet.The chapter presents this as a current example of overconfident AI predictions.Source: Introduction · p. 1
  • AI springs bring confident predictions and investment, while stalled progress can bring lost funding, jobs, and an AI winter.McDermott warned of this cycle in 1984, and a new winter followed about five years later.Source: Introduction · p. 1
  • The chapter argues that over-optimism arises partly from fallacies in how people talk about AI and intuit the nature of intelligence.Understanding these fallacies may point toward more robust, trustworthy, and potentially more intelligent systems.Source: Introduction · pp. 1–2

Springs and Winters

The history of AI shows recurring optimism and disappointment, while persistent brittleness and unclear limits complicate predictions about human-level intelligence.

Springs and Winters

  • The 1960s optimism gave way to an AI winter after limitations of perceptrons and negative government assessments led to sharp funding decreases.Enthusiasm rose again in the 1980s around expert systems, Japan’s Fifth Generation project, US Strategic Computing, and neural networks, but those hopes were later dashed.Source: Springs and Winters · p. 2
  • Deep-learning systems can make unpredictable errors on inputs unlike their training data, partly because shortcuts can yield correct answers for the wrong reasons [27] [28].They often fail to learn abstract concepts for transfer [29] and can be misled by adversarial perturbations [30].Source: Springs and Winters · p. 3
  • The sources of deep networks’ brittleness and vulnerability remain incompletely understood, and whether more layers and data suffice is debated.The chapter notes that their decision-making can be opaque and questions whether they understand data in a human sense.Source: Springs and Winters · p. 3
  • New approaches were generating optimism by mid-2021, but the path from narrow advances to human-level AI remained unclear.The chapter expects cycles of AI springs and winters to continue.Source: Springs and Winters · p. 3
  • The chapter identifies four fallacies in AI thinking as sources of false confidence about near-term prospects for truly intelligent machines.It argues that predictions about human-level AI timelines reflect biases and limited understanding of intelligence.Source: Springs and Winters · p. 3

Fallacy 1: Narrow intelligence is on a continuum with general intelligence

The first-step fallacy treats narrow AI advances as progress toward general intelligence, despite the unresolved barrier of common sense.

Fallacy 1: Narrow intelligence is on a continuum with general intelligence

  • The first-step fallacy treats any improvement in narrow AI as movement along a continuum toward general intelligence.Dreyfus’s analogy compares this inference to claiming a tree-climbing monkey is progressing toward landing on the moon.Source: Fallacy 1: Narrow intelligence is on a… · p. 4
  • Common sense has repeatedly been an unexpected obstacle in the assumed continuum from narrow AI advances to general AI.The chapter identifies common sense as a barrier discussed by critics of the assumed progression.Source: Fallacy 1: Narrow intelligence is on a… · p. 4

Fallacy 2: Easy things are easy and hard things are hard

Human intuition misjudges which tasks are hard for machines because many effortless human abilities rely on complex, unconscious processes.

Fallacy 2: Easy things are easy and hard things are hard

  • Moravec’s paradox is that machines can find human-easy perception and mobility harder than games, mathematics, or translation.Tasks humans find difficult, such as chess and Go, have often proved comparatively easier for machines.Source: Fallacy 2: Easy things are easy and… · p. 4
  • The assumption that cognition begins only after perception underlies symbolic AI’s focus on reasoning over already-perceived input.Herbert Simon’s claim placed cognition above a 100-millisecond threshold, overlooking unconscious perceptual processes.Source: Fallacy 2: Easy things are easy and… · p. 4
  • Claims that tasks requiring less than a second of human thought are readily automatable underestimate unconscious perception and thought.The chapter cites Andrew Ng’s 2016 statement as echoing this assumption.Source: Fallacy 2: Easy things are easy and… · p. 4
  • Charades demands acting, language, and theory of mind, abilities described as beyond what AI can accomplish today [43].The comparison challenges treating Go as uniformly one of AI’s most challenging domains.Source: Fallacy 2: Easy things are easy and… · p. 4
  • Humans are largely unaware of the complexity of their own thought processes.Moravec attributes this to extensive unconscious sensory and motor knowledge supporting deliberate reasoning, while Minsky observes that minds are least noticed where they perform best.Source: Fallacy 2: Easy things are easy and… · pp. 4–5

Fallacy 3: The lure of wishful mnemonics

The section argues that intelligence-laden labels for AI systems and benchmarks can overstate what they do. Systems may excel on narrowly named tests without demonstrating the broader human abilities those names suggest.

Fallacy 3: The lure of wishful mnemonics

  • Wishful mnemonics—intelligence-related terms used for programs and evaluations—can make names seem like evidence of the abilities they describe.McDermott warns that naming a program “UNDERSTAND” begs the question of whether it implements understanding.Source: Fallacy 3: The lure of wishful mnemonics · p. 5
  • Anthropomorphic descriptions of Watson and AlphaGo can mislead the public and media, and can also shape AI experts’ views of how humanlike their systems are.The examples attribute reading, understanding, goals, thoughts, and judgments to these systems.Source: Fallacy 3: The lure of wishful mnemonics · p. 5
  • The Stanford Question Answering, RACE Reading Comprehension, and General Language Understanding benchmarks test limited versions of the abilities in their names.The passage reports that leading machines surpassed humans on these particular benchmarks, not on the general abilities their names suggest.Source: Fallacy 3: The lure of wishful mnemonics · pp. 5–6
  • Shortcuts—statistical correlations in training data—can let machines score highly without learning the skill a benchmark is intended to test.The passage notes that machines can outperform humans on particular benchmarks while remaining far from matching the associated general abilities.Source: Fallacy 3: The lure of wishful mnemonics · pp. 5–6

Fallacy 4: Intelligence is all in the brain

The section challenges the assumption that intelligence can be understood or built as brain-only information processing. It argues that cognition is embodied and integrated with emotion, social life, and common sense.

Fallacy 4: Intelligence is all in the brain

  • The assumption that intelligence can be disembodied has deep roots in philosophy and cognitive science and underlies influential AI ideas such as the Physical Symbol System Hypothesis.The hypothesis holds that a digital-computer-like physical symbol system has the necessary and sufficient means for general intelligent action.Source: Fallacy 4: Intelligence is all in the… · p. 6
  • Both symbolic AI and neural approaches have generally treated the body as unnecessary to intelligence, while robots’ physical interactions and feedback remain limited.The passage describes deep neural networks as passively taking in data and outputting behavior, like a “brain-in-a-vat.”Source: Fallacy 4: Intelligence is all in the… · p. 6
  • The brain-only view encourages predictions that matching the brain’s computing capacity and developing suitable software could yield human-level AI.The passage gives a 10^15 FLOP/s estimate and cites predictions about scaling hardware to brain-like resources.Source: Fallacy 4: Intelligence is all in the… · p. 6
  • Research across neuroscience, cognitive psychology, and linguistics supports links between cognition, sensory and motor systems, and body-based concepts.The cited findings include abstract thinking using body-based neural maps and abstract concepts grounded in physical internal models.Source: Fallacy 4: Intelligence is all in the… · p. 7
  • Embodied cognition has gained evidence from several disciplines, but AI has mostly ignored this research, apart from a small group pursuing embodied and developmental approaches.Related areas named include embodied AI, developmental robotics, and grounded language understanding.Source: Fallacy 4: Intelligence is all in the… · p. 7
  • Human intelligence appears strongly integrated with emotions, desires, selfhood, autonomy, and commonsense understanding, so their separability is unclear.The passage argues that speculation about pure rationality lacks support from psychology and neuroscience.Source: Fallacy 4: Intelligence is all in the… · p. 7

Conclusions

The paper concludes that four fallacies expose limits in how AI progress and intelligence are understood, helping explain why humanlike machine intelligence remains harder than expected. It identifies open research questions and calls for a stronger scientific understanding of intelligence.

Conclusions

  • The four fallacies expose flaws in current conceptions of AI and limited intuitions about intelligence, helping explain why humanlike intelligence in machines is harder than expected.Source: Conclusions · p. 8
  • Open questions include how to assess progress toward human-level AI, describe systems’ actual abilities, and determine whether dimensions of human cognition can be disentangled.The paper also asks how to improve intuitions about intelligence.Source: Conclusions · p. 8
  • More effective AI progress assessment requires better vocabulary for machine abilities and scientific understanding of intelligence across natural systems, with deeper engagement across disciplines.Source: Conclusions · p. 8
  • Common sense involves world knowledge used to recognize and predict situations and guide action; capturing it in machines remains an open challenge.The paper identifies core knowledge, abstraction, and analogy as relevant abilities, and points to child development as one encouraging direction.Source: Conclusions · p. 8
  • The paper characterizes AI as alchemy-like and argues that understanding its progress requires developing a scientific understanding of intelligence.Source: Conclusions · p. 8
Loading 2104.12871v2…