Unpredictability of AI: On the Impossibility of Accurately Predicting All Actions of a Smarter Agent

@article{Yampolskiy2020UnpredictabilityOA,
  title={Unpredictability of AI: On the Impossibility of Accurately Predicting All Actions of a Smarter Agent},
  author={Roman V Yampolskiy},
  journal={J. Artif. Intell. Conscious.},
  year={2020},
  volume={7},
  pages={109-118}
}
The young field of AI Safety is still in the process of identifying its challenges and limitations. In this paper, we formally describe one such impossibility result, namely Unpredictability of AI.... 
Unownability of AI : Why Legal Ownership of Artificial Intelligence is Hard
TLDR
It is concluded that it is difficult if not impossible to establish ownership claims over AI models beyond a reasonable doubt.
On Controllability of AI
TLDR
Consequences of uncontrollability of AI are discussed with respect to future of humanity and research on AI, and AI safety and security.
On the Controllability of Artificial Intelligence: An Analysis of Limitations
TLDR
The consequences of uncontrollability of AI are discussed with respect to the future of humanity and research on AI, and AI safety and security.
Impossibility Results in AI: A Survey
TLDR
This paper categorized impossibility theorems applicable to the domain of AI into five categories: deduction, indistinguishability, induction, tradeoffs, and intractability, and concluded that deductive impossibilities deny 100%-guarantees for security.
Impossibility Results for the Online Verification of Ethical and Legal Behaviour of Robots
TLDR
It is argued that, under reasonable assumptions about robot programming and about an observer’s capabilities, no (deterministic) module, algorithmic or otherwise, will enable an observer to tell whether a robot will always follow the rules it is said to obey in its interactions with other robots.
AI Risk Skepticism
TLDR
This work surveys skepticism regarding AI risk and shows parallels with other types of scientific skepticism, and suggests some intervention approaches, which may be successful in reducing AI risk skepticism, at least amongst artificial intelligence researchers.
On the Differences between Human and Machine Intelligence
TLDR
This paper proves that humans are not general intelligences, and widespread implicit assumption of equivalence between capabilities of AGI and HLAI appears to be unjustified.
Human $\neq$ AGI.
TLDR
This paper proves that humans are not general intelligences, and widespread implicit assumption of equivalence between capabilities of AGI and HLAI appears to be unjustified.
Categorization and challenges of utilitarianisms in the context of artificial intelligence
TLDR
A categorization of the various presented utilitarianisms into static utilitarianisms and dynamic utilitarianisms is suggested and a possibility of using a combination of the two categories to resolve most of the challenges without the need to abandon the concept of utilitarianisms as such is suggested.
Four New Horsemen of an Apocalypse? Solar Flares, Super-volcanoes, Pandemics, and Artificial Intelligence
If economists have largely failed to predict or prevent the Global Financial Crisis in 2008, and the more disastrous economic collapse associated with the pandemic of 2020, what else is the
...
...

References

SHOWING 1-10 OF 21 REFERENCES
Impossibility of distributed consensus with one faulty process
TLDR
In this paper, it is shown that every protocol for this problem has the possibility of nontermination, even with only one faulty process.
Predicting future AI failures from historic examples
TLDR
It is suggested that both the frequency and the seriousness of future AI failures will steadily increase and the first attempt to assemble a public data set of AI failures is extremely valuable to AI Safety researchers.
Unpredictability and undecidability in dynamical systems.
  • Moore
  • Physics
    Physical review letters
  • 1990
TLDR
It is shown that motion with as few as three degrees of freedom can be equivalent to a Turing machine, and so be capable of universal computation.
Artificial Intelligence Safety and Security
TLDR
The book would be the first textbook to address challenges of constructing safe and secure advanced machine intelligence, and will be comprised of chapters from leading AI Safety researchers addressing different aspects of the AI control problem as it relates to the development ofsafe and secure artificial intelligence.
Unpredictability and Computational Irreducibility
TLDR
It is proved that, through a robust definition of what means “to be unable to compute the n th step without having to follow the same path than simulating the automaton”, this implies genuinely that if the behavior of an object is computationally irreducible, no computation of its n th state can be faster than the simulation itself.
What to Do with the Singularity Paradox?
The paper begins with an introduction of the Singularity Paradox, an observation that: “Superintelligent machines are feared to be too dumb to possess commonsense”. Ideas from leading researchers in
Computational irreducibility and the predictability of complex physical systems.
TLDR
This work shows how to coarse grain CA in all classes of Wolfram's classification and finds that computationally irreducible physical processes can be predictable and even computationally reducible at a coarse-grained level of description.
Generalized shifts: unpredictability and undecidability in dynamical systems Nonlinearity 4 199-230
TLDR
A class of shift-like dynamical systems is presented that displays a wide variety of behaviours, including periodic points, basins of attraction, and time series, and it is shown that they can be embedded in smooth maps in R2, or smooth flows in R3.
Aggregating Sets of Judgments: An Impossibility Result
Suppose that the members of a certain group each hold a rational set of judgments on some interconnected questions. And imagine that the group itself now has to form a collective, rational set of
The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents
TLDR
The relation between intelligence and motivation in artificial agents is discussed, developing and briefly arguing for two theses that help understand the possible range of behavior of superintelligent agents and point to some potential dangers in building such an agent.
...
...