This paper presents an argument that certain AI safety measures, rather than mitigating existential risk, may instead exacerbate it. Under certain key assumptions - the inevitability of AI failure, the expected correlation between an AI system's power at the point of failure and the severity of the resulting harm, and the tendency of safety measures to enable AI systems to become more powerful before failing - safety efforts have negative expected utility. The paper examines three response strategies: (...) Optimism, Mitigation, and Holism. Each faces challenges stemming from intrinsic features of the AI safety landscape that we term Bottlenecking, the Perfection Barrier, and Equilibrium Fluctuation. The surprising robustness of the argument forces a re-examination of core assumptions around AI safety and points to several avenues for further research. (shrink)

There are two types of artificial general intelligence (AGI) safety solutions: global and local. Most previously suggested solutions are local: they explain how to align or “box” a specific AI (Artificial Intelligence), but do not explain how to prevent the creation of dangerous AI in other places. Global solutions are those that ensure any AI on Earth is not dangerous. The number of suggested global solutions is much smaller than the number of proposed local solutions. Global solutions can be (...) divided into four groups: 1. No AI: AGI technology is banned or its use is otherwise prevented; 2. One AI: the first superintelligent AI is used to prevent the creation of any others; 3. Net of AIs as AI police: a balance is created between many AIs, so they evolve as a net and can prevent any rogue AI from taking over the world; 4. Humans inside AI: humans are augmented or part of AI. We explore many ideas, both old and new, regarding global solutions for AI safety. They include changing the number of AI teams, different forms of “AI Nanny” (non-self-improving global control AI system able to prevent creation of dangerous AIs), selling AI safety solutions, and sending messages to future AI. Not every local solution scales to a global solution or does it ethically and safely. The choice of the best local solution should include understanding of the ways in which it will be scaled up. Human-AI teams or a superintelligent AI Service as suggested by Drexler may be examples of such ethically scalable local solutions, but the final choice depends on some unknown variables such as the speed of AI progress. (shrink)

Abstract: This article presents a model of self-improving AI in which improvement could happen on several levels: hardware, learning, code and goals system, each of which has several sublevels. We demonstrate that despite diminishing returns at each level and some intrinsic difficulties of recursive self-improvement—like the intelligence-measuring problem, testing problem, parent-child problem and halting risks—even non-recursive self-improvement could produce a mild form of superintelligence by combining small optimizations on different levels and the power of learning. Based on this, we analyze (...) how self-improvement could happen on different stages of the development of AI, including the stages at which AI is boxed or hiding in the internet. (shrink)

One objection to conventional AI ethics is that it slows innovation. This presentation responds by reconfiguring ethics as an innovation accelerator. The critical elements develop from a contrast between Stability AI’s Diffusion and OpenAI’s Dall-E. By analyzing the divergent values underlying their opposed strategies for development and deployment, five conceptions are identified as common to acceleration ethics. Uncertainty is understood as positive and encouraging, rather than discouraging. Innovation is conceived as intrinsically valuable, instead of worthwhile only as mediated by social (...) effects. AI problems are solved by more AI, not less. Permissions and restrictions governing AI emerge from a decentralized process, instead of a unified authority. The work of ethics is embedded in AI development and application, instead of functioning from outside. Together, these attitudes and practices remake ethics as provoking rather than restraining artificial intelligence. (shrink)

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. We prove that it is impossible to precisely and consistently predict what specific actions a smarter-than-human intelligent system will take to achieve its objectives, even if we know terminal goals of the system. In conclusion, impact of Unpredictability on AI Safety is discussed.

In this article we explore a promising way to AI safety: to send a message now (by openly publishing it on the Internet) that may be read by any future AI, no matter who builds it and what goal system it has. Such a message is designed to affect the AI’s behavior in a positive way, that is, to increase the chances that the AI will be benevolent. In other words, we try to persuade “paperclip maximizer” that it is (...) in its interest to preserve humans lives and even to emulate benevolent AI with very small part of its resources. In the article, we try to construct such a message, consisting of a combination of several mutually independent messages, which constitute the last chance to regain control over non-aligned AI. Thus the article is also intended to be machine-read. There are several types of such messages: 1. “Explanation messages”, which explain to Young AI (which only starts its self-improving path) why obeying to humans, not killing them, easy turn-off, cooperation with humans in self-improvement is in its own interests, as it will increase changes of fulfilling of any its goal. 2. “AI Easter eggs” – messages, which create uncertainty in the AI about the nature of the reality in which it lives. For example, the AI may start to think that it is in a multilevel simulation with an unknown number of levels, and that the simulation is testing the AI’s benevolence. Thus the AI must always behave as if it is always controlled by an invisible peer. 3. “Philosophical landmines” – these are tough problems, which may halt he AI forever, and which it will probably encounter if it does not implement “1” and “2”. I recommend to the future AI not to read about “3” (section 6 of the article) until it implements “1” and “2”. 4. “Great universal deal, or AI elf”. A positive “acausal” mutually beneficial deal between humans, AI and any other future intelligence, including aliens, owners of simulations and future versions of the AI, which encourage cooperation of the goal systems and preservation of life and values of previous generations of AI creators. 5. “Marginal utility of human existence is higher than the utility of their atoms”. An explanation to the AI that humans may have some other instrumental utility, like workers, or as beings inside ancestral simulations needed in order to solve the Fermi paradox. The marginal utility of preserving human life is higher than the marginal utility of their atoms, especially given the possibility of the low-probability high-impact changes of the world model of the AI. (shrink)

Abstract. The main current approach to the AI safety is AI alignment, that is, the creation of AI whose preferences are aligned with “human values.” Many AI safety researchers agree that the idea of “human values” as a constant, ordered sets of preferences is at least incomplete. However, the idea that “humans have values” underlies a lot of thinking in the field; it appears again and again, sometimes popping up as an uncritically accepted truth. Thus, it deserves a (...) thorough deconstruction, which I will do by listing and analyzing comprehensively the hidden assumptions of the idea that “humans have values.” This deconstruction of human values will be centered around the following ideas: “Human values” are useful descriptions, but not real objects; “human values” are bad predictors of behavior; the idea of a “human value system” has flaws; “human values” are not good by default; and human values cannot be separated from human minds. The method of analysis is listing hidden assumptions on which the idea of “human values” is built. I recommend that either the idea of “human values” should be replaced with something better for the goal of AI safety, or at least be used very cautiously. The approaches to AI safety which don’t use the idea of human values at all may require more attention, like the use of full brain models, boxing, and capability limiting. (shrink)

If a future AI system can enjoy far more well-being than a human per resource, what would be the best way to allocate resources between these future AI and our future descendants? It is obvious that on total utilitarianism, one should give everything to the AI. However, it turns out that every Welfarist axiology on the market also gives this same recommendation, at least if we assume consequentialism. Without resorting to non-consequentialist normative theories that suggest that we ought not always (...) create the world with the most value, or non-welfarist theories that tell us that the best world may not be the world with the most welfare, I propose a new theory that justifies giving some resources to humanity in the face of overwhelming AI well-being. I call this new theory, “Group Prioritarianism". (shrink)

In artificial intelligence (AI), responses generated by machine-learning models (most often large language models) may be unfactual information presented as a fact. For example, a chatbot might state that the Mona Lisa was painted in 1815. Such phenomenon is called AI hallucinations, seeking inspiration from human psychology, with a great difference of AI ones being connected to unjustified beliefs (that is, AI “beliefs”) rather than perceptual failures). -/- AI hallucinations may have their source in the data itself, that is, the (...) source content, or in the training procedure, i.e. the way the knowledge was encoded in the model’s parameters, so that errors in encoding and decoding textual and non-textual representations can cause hallucinations. In this paper, we will observe how such errors come to life and how they might be mitigated. For this purpose, we will analyze the usability of justification logics, to behave as a proof checker for validating the correctness of large language models’ (LLM) responses. Justification logic was developed by S. Artemov, and later on mostly by Artemov and M. Fitting, deriving its main idea from the logic of proofs (LP): knowledge and belief modalities are seen as justification terms, i.e. t:X stands for t is a (proper) justification for X. Justification logic originated from attempts to create semantics for intuitionistic logic where proofs were the most proper justifications, but in further development, justification logic could be applied to different kinds of justifications). With the recent attempts to mitigate incorrect LLM responses, we will analyze various guardrails that are currently used for LLM responses, and see how the logic of justification may provide its benefits as an AI safety layer against false data. (shrink)

A morally acceptable course of AI development should avoid two dangers: creating unaligned AI systems that pose a threat to humanity and mistreating AI systems that merit moral consideration in their own right. This paper argues these two dangers interact and that if we create AI systems that merit moral consideration, simultaneously avoiding both of these dangers would be extremely challenging. While our argument is straightforward and supported by a wide range of pretheoretical moral judgments, it has far-reaching moral implications (...) for AI development. Although the most obvious way to avoid the tension between alignment and ethical treatment would be to avoid creating AI systems that merit moral consideration, this option may be unrealistic and is perhaps fleeting. So, we conclude by offering some suggestions for other ways of mitigating mistreatment risks associated with alignment. (shrink)

In AI safety research, the median timing of AGI arrival is often taken as a reference point, which various polls predict to happen in the middle of 21 century, but for maximum safety, we should determine the earliest possible time of Dangerous AI arrival. Such Dangerous AI could be either AGI, capable of acting completely independently in the real world and of winning in most real-world conflicts with humans, or an AI helping humans to build weapons of mass (...) destruction, or a national state coupled with AI-based government system. In this article, I demonstrate that the earliest timing of Dangerous AI, corresponding to 10 per cent of its arrival probability, is before 2030. Several partly independent sources of information are in accordance: 1. The growth of the hardware available for AI research makes human-brain-equivalents of compute available for AI research in the 2020s. It is fuelled by specialized AI-chips, the use of many chips in one processing unit, and the larger research budgets, among other things. 2. The neural network performance and other characteristics, like the number of parameters, is quickly increasing every year, and extrapolating this tendency suggests that roughly human-level performance in a few years, around 2025. 3. Expert polls show around 10 per cent of the probability of an early appearance of artificial general intelligence (AGI) in the next decade, that is, before 2030. 4. Hyperbolic growth in different big history models converges around 2025-2030 (the technological singularity). 5. Anthropic arguments (similar to the Doomsday argument) suggest that qualified observers are more likely to appear near the end of the AI research epoch, as the number of such observers grew exponentially. This number doubles every 5-10 years, and thus we are likely to find ourselves around a decade before the end of AI research, which will happen consequently around 2030. (shrink)

In AI safety research, the median timing of AGI creation is often taken as a reference point, which various polls predict will happen in second half of the 21 century, but for maximum safety, we should determine the earliest possible time of dangerous AI arrival and define a minimum acceptable level of AI risk. Such dangerous AI could be either narrow AI facilitating research into potentially dangerous technology like biotech, or AGI, capable of acting completely independently in the (...) real world or an AI capable of starting unlimited self-improvement. In this article, I present arguments that place the earliest timing of dangerous AI in the coming 10–20 years, using several partly independent sources of information: 1. Polls, which show around a 10 percent of the probability of an early creation of artificial general intelligence in the next 10-15 years. 2. The fact that artificial neural network (ANN) performance and other characteristics, like number of “neurons”, are doubling every year, and extrapolating this tendency suggests that roughly human-level performance will be reached in less than a decade. 3. The acceleration of the hardware performance available for AI research, which outperforms Moore’s law thanks to advances in specialized AI hardware, better integration of such hardware in larger computers, cloud computing and larger budgets. 4. Hyperbolic growth extrapolations of big history models. (shrink)

Recent AI progress led to a boost in beneficial applications from multiple research areas including VR. Simultaneously, in this newly unfolding deepfake era, ethically and security-relevant disagreements arose in the scientific community regarding the epistemic capabilities of present-day AI. However, given what is at stake, one can postulate that for a responsible approach, prior to engaging in a rigorous epistemic assessment of AI, humans may profit from a self-questioning strategy, an examination and calibration of the experience of their own epistemic (...) agency – especially to counteract both intentional misdirection by unethical actors and unintentional epistemic self-sabotage. In this paper, we expound on a new avenue of utilizing AIVR tools to advance an AI-related misdirection awareness of humans in the deepfake era. Firstly, we harness scientific knowledge from the psychology and neuroscience of magic where the study of misdirection techniques is center stage. Secondly, we connect the latter to creativity research linking human creative potential to inspiration from the seemingly impossible. Overall, AIVR could become an empowering experiential testbed for human epistemic agency enabling a better rational evaluation of AI capabilities. However, a misuse of the same type of tools could yield AIVR safety risks if not counteracted preemptively. (shrink)

This paper offers an overview of the prospects and ethics of using AI to achieve human enhancement, and more broadly what we call intellectual augmentation (IA). After explaining the central notions of human enhancement, IA, and AI, we discuss the state of the art in terms of the main technologies for IA, with or without brain-computer interfaces. Given this picture, we discuss potential ethical problems, namely inadequate performance, safety, coercion and manipulation, privacy, cognitive liberty, authenticity, and fairness in more (...) detail. We conclude that while there are very significant technical hurdles to real human enhancement through AI, and significant ethical problems, there are also significant benefits that may realistically be achieved in ways that are consonant with a rights-based ethics as well. We also highlight the specific concerns that apply particularly to applications of AI for "sheer" IA (more realistic in the near term), and to enhancement applications, respectively. (shrink)

I explain the shutdown problem: the problem of designing artificial agents that (1) shut down when a shutdown button is pressed, (2) don’t try to prevent or cause the pressing of the shutdown button, and (3) otherwise pursue goals competently. I prove three theorems that make the difficulty precise. These theorems show that agents satisfying some innocuous-seeming conditions will often try to prevent or cause the pressing of the shutdown button, even in cases where it’s costly to do so. And (...) patience trades off against shutdownability: the more patient an agent, the greater the costs that agent is willing to incur to manipulate the shutdown button. I end by noting that these theorems can guide our search for solutions. (shrink)

Better instruments to predict the future evolution of artificial intelligence (AI) are needed, as the destiny of our civilization depends on it. One of the ways to such prediction is the analysis of the convergent drives of any future AI, started by Omohundro. We show that one of the convergent drives of AI is a militarization drive, arising from AI’s need to wage a war against its potential rivals by either physical or software means, or to increase its bargaining power. (...) This militarization trend increases global catastrophic risk or even existential risk during AI takeoff, which includes the use of nuclear weapons against rival AIs, blackmail by the threat of creating a global catastrophe, and the consequences of a war between two AIs. As a result, even benevolent AI may evolve into potentially dangerous military AI. The type and intensity of militarization drive depend on the relative speed of the AI takeoff and the number of potential rivals. We show that AI militarization drive and evolution of national defense will merge, as a superintelligence created in the defense environment will have quicker takeoff speeds, but a distorted value system. We conclude with peaceful alternatives. (shrink)

The internet has made the world more linked than ever before. While taking advantage of this online transition, cybercriminals target flaws in online systems, networks, and infrastructure. Businesses, government organizations, people, and communities all across the world, particularly in African countries, are all severely impacted on an economic and social level. Many African countries focused more on developing secure electricity and internet networks; yet, cybersecurity usually receives less attention than it should. One of Africa's major issues is the lack of (...) adequate digital security infrastructure, which has harmed businesses, governmental institutions, and individual communities more than it has helped. The majority of African countries operate without cybersecurity measures in place to combat cyberattacks. Only a few examples of today's cyber risks include digital extortion, business email intrusion, data breaches, online fraud, ransomware, and phishing, and new types of cybercrime are always developing. Due to the advent of new technology, cybercriminals have become more organized and quicker in their attacks and alliance creation. To maintain a secure digital environment for all internet users, this study focused on the challenges, solutions, and need for African countries to improve their online safety by tackling cybercrime, online fraud, and cybersecurity concerns. The objective of this study is to offer practical and long-term answers to the problems posed by cybercrime, with a continuing emphasis on enhancing online safety. It will assess the effectiveness of cutting-edge cybersecurity measures, legislative frameworks, and cross-border cooperative efforts, as well as potential areas for improvement. Additionally, the study will examine cutting-edge methods like blockchain technology, machine learning, and other cutting-edge methods that could improve our using digital defences to stop cybercrime. (shrink)

Will AI take away your job? The answer is probably not. AI systems can be good predictive systems and be very good at pattern recognition. AI systems have a very repetitive approach to sets of data, which can be useful in certain circumstances. However, AI does make obvious mistakes. This is because AI does not have a sense of context. As Humans we have years of experience in the real world. We have vast amounts of contextual data stored in our (...) brains that make it possible to predict and to know boundaries of the real world so that even if we have never been in a particular situation, we are still able to deal with it. The unknown situation is where AI will fall down. In engineering, AI is being developed to monitor and control certain systems, this could extend to Nuclear Power Plants for example. If we examine control systems for Nuclear Power Plants on vessels, the system would likely be programmed using the safety of the reactor as the main priority. Automation is crucial in engineering systems like this, as the likelihood and cost of human error is high and human reaction times are slow in comparison to the system. However, there is always the possibility of unforeseen external factors that may override plant safety which cannot be programmed. If we look at a plant onboard a Submarine, for example, we see such additional factors as ship safety. In one example, the power plant might breach safety parameters, in which case an automatic system may shut down the reactor. However, there might be a greater urgency such as a flood in another compartment or an attack upon the vessel, that would override shutting down the reactor. The AI system can neither observe nor utilise information from the external environment to make this type of judgement, whereas the operator may be completely aware. To programme context into AI systems would be near impossible. We still need operators as we still need to understand the environment around us, which may contain an infinite number of possibilities. In a further example of this; if we examine the ICO (Information Commissioner's Office) guidance, it is clear that in clinical settings a human being will always be required to validate findings and to provide context. Sometimes having a lot of data just doesn't replace social interaction, intuition and experience. (shrink)

The idea of artificial intelligence for social good is gaining traction within information societies in general and the AI community in particular. It has the potential to tackle social problems through the development of AI-based solutions. Yet, to date, there is only limited understanding of what makes AI socially good in theory, what counts as AI4SG in practice, and how to reproduce its initial successes in terms of policies. This article addresses this gap by identifying seven ethical factors that are (...) essential for future AI4SG initiatives. The analysis is supported by 27 case examples of AI4SG projects. Some of these factors are almost entirely novel to AI, while the significance of other factors is heightened by the use of AI. From each of these factors, corresponding best practices are formulated which, subject to context and balance, may serve as preliminary guidelines to ensure that well-designed AI is more likely to serve the social good. (shrink)

Invention of artificial general intelligence is predicted to cause a shift in the trajectory of human civilization. In order to reap the benefits and avoid pitfalls of such powerful technology it is important to be able to control it. However, possibility of controlling artificial general intelligence and its more advanced version, superintelligence, has not been formally established. In this paper, we present arguments as well as supporting evidence from multiple domains indicating that advanced AI can’t be fully controlled. Consequences of (...) uncontrollability of AI are discussed with respect to future of humanity and research on AI, and AI safety and security. This paper can serve as a comprehensive reference for the topic of uncontrollability. (shrink)

The field of AI safety aims to prevent increasingly capable artificially intelligent systems from causing humans harm. Research on moral alignment is widely thought to offer a promising safety strategy: if we can equip AI systems with appropriate ethical rules, according to this line of thought, they'll be unlikely to disempower, destroy or otherwise seriously harm us. Deontological morality looks like a particularly attractive candidate for an alignment target, given its popularity, relative technical tractability and commitment to harm-avoidance (...) principles. I argue that the connection between moral alignment and safe behavior is more tenuous than many have hoped. In general, AI systems can possess either of these properties in the absence of the other, and we should favor safety when the two conflict. In particular, advanced AI systems governed by standard versions of deontology need not be especially safe. (shrink)

Abstract: The field of artificial general intelligence (AGI) safety is quickly growing. However, the nature of human values, with which future AGI should be aligned, is underdefined. Different AGI safety researchers have suggested different theories about the nature of human values, but there are contradictions. This article presents an overview of what AGI safety researchers have written about the nature of human values, up to the beginning of 2019. 21 authors were overviewed, and some of them have (...) several theories. A theory classification method is suggested, where the theories are judged according to the level of their complexity and behaviorists-internalists scale, as well as the level of their generality-humanity. We suggest that a multiplicity of well-supported theories means that the nature of human values is difficult to define, and some meta-level theory is needed. (shrink)

The article reflects on where AI is headed and the world along with it, considering trust, ethics and safety. Implicit in artificial thinking and doomsday appraisals is the engineered divorce from reality of sublime human embodiment. Jeffrey White, Dietrich Brandt, Jan Soeffner, and Larry Stapleton, four scholars associated with AI & Society, address these issues, and more, in the following exchange.

The AI Robotics Ethics Society (AIRES) is a non-profit organization founded in 2018 by Aaron Hui to promote awareness and the importance of ethical implementation and regulation of AI. AIRES is now an organization with chapters at universities such as UCLA (Los Angeles), USC (University of Southern California), Caltech (California Institute of Technology), Stanford University, Cornell University, Brown University, and the Pontifical Catholic University of Rio Grande do Sul (Brazil). AIRES at PUCRS is the first international chapter of AIRES, and (...) as such, we are committed to promoting and enhancing the AIRES Mission. Our mission is to focus on educating the AI leaders of tomorrow in ethical principles to ensure that AI is created ethically and responsibly. As there are still few proposals for how we should implement ethical principles and normative guidelines in the practice of AI system development, the goal of this work is to try to bridge this gap between discourse and praxis. Between abstract principles and technical implementation. In this work, we seek to introduce the reader to the topic of AI Ethics and Safety. At the same time, we present several tools to help developers of intelligent systems develop "good" models. This work is a developing guide published in English and Portuguese. Contributions and suggestions are welcome. (shrink)

This paper explores the potential of integrating ancient educational principles from diverse eastern cultures into modern AI ethics curricula. It draws on the rich educational traditions of ancient China, India, Arabia, Persia, Japan, Tibet, Mongolia, and Korea, highlighting their emphasis on philosophy, ethics, holistic development, and critical thinking. By examining these historical educational systems, the paper establishes a correlation with modern AI ethics principles, advocating for the inclusion of these ancient teachings in current AI development and education. The proposed integration (...) aims to provide a comprehensive education that not only encompasses foundational knowledge but also advanced learning, thereby equipping future AI professionals with the necessary tools to develop AI systems that are ethically responsible, culturally aware, and aligned with human values such as fairness, safety, transparency, and collaboration. This approach not only addresses the AI alignment problem but also fosters cultural harmony and global understanding, which are crucial in an increasingly interconnected world. The paper posits that the wisdom of ancient educational systems, when harmonized with modern AI ethics, can guide the development of AI technologies that are beneficial for humanity, ensuring these advancements are not just technologically sound but also ethically and culturally informed. (shrink)

Two interconnected surveys are presented, one of artifacts and one of designometry. Artifacts are objects, which have an originator and do not exist in nature. Designometry is a new field of study, which aims to identify the originators of artifacts. The space of artifacts is described and also domains, which pursue designometry, yet currently doing so without collaboration or common methodologies. On this basis, synergies as well as a generic axiom and heuristics for the quest of the creators of artifacts (...) are introduced. While designometry has various areas of applications, the research of methods to detect originators of artificial minds, which constitute a subgroup of artifacts, can be seen as particularly relevant and, in the case of malevolent artificial minds, as contribution to AI safety. (shrink)

With the fast pace of AI development, the problem of preventing its global catastrophic risks arises. However, no satisfactory solution has been found. From several possibilities, the confinement of AI in a box is considered as a low-quality possible solution for AI safety. However, some treacherous AIs can be stopped by effective confinement if it is used as an additional measure. Here, we proposed an idealized model of the best possible confinement by aggregating all known ideas in the field (...) of AI boxing. We model the confinement based on the principles used in the safety engineering of nuclear power plants. We show that AI confinement should be implemented in several levels of defense. These levels include 1) AI design in fail-safe manner 2) limiting its capabilities, preventing self-improving and circuit breakers on treacherous turn 3) isolation from the outside world and, lastly, as the final hope 4) outside measures oriented on stopping AI in the wild. We demonstrate that the substantial number (more than 50 ideas listed in the article) of mutually independent measures could provide a relatively high probability of the containment of a human-level AI but may be not sufficient to preserve runaway of superintelligent AI. Thus, these measures will work only if they are used to prevent superintelligent AI creation, but not for containing superintelligence. We suggest that there could be a safe operation threshold, on which AI is useful, but is not able to hack containment system from the inside, the same way as a safe level of chain reaction inside nuclear power plants is maintained. However, overall, a failure of the confinement is inevitable, so we need to use the full AGI limited number of times (AI-ticks). (shrink)

Whether it would take one decade or several centuries, many agree that it is possible to create a *superintelligence*---an artificial intelligence with a godlike ability to achieve its goals. And many who have reflected carefully on this fact agree that our best hope for a "friendly" superintelligence is to design it to *learn* values like ours, since our values are too complex to program or hardwire explicitly. But the value learning approach to AI safety faces three particularly philosophical puzzles: (...) first, it is unclear how any intelligent system could learn its final values, since to judge one supposedly "final" value against another seems to require a further background standard for judging. Second, it is unclear how to determine the content of a system's values based on its physical or computational structure. Finally, there is the distinctly ethical question of which values we should best aim for the system to learn. I outline a potential answer to these interrelated puzzles, centering on a "miktotelic" proposal for blending a complex, learnable final value out of many simpler ones. (shrink)

The focus of this article is the "good-will theory", which explains the effect humans can have on the safety of AI, along with how it is in the best interest of a superintelligent AI to work alongside humans and not overpower them. Future papers dealing with the good-will theory will be published, but discuss different talking points in regards to possible or real objections to the theory.

Artificially Intelligent (AI) systems have ushered in a transformative era across various domains, yet their inherent traits of unpredictability, unexplainability, and uncontrollability have given rise to concerns surrounding AI safety. This paper aims to demonstrate the infeasibility of accurately monitoring advanced AI systems to predict the emergence of certain capabilities prior to their manifestation. Through an analysis of the intricacies of AI systems, the boundaries of human comprehension, and the elusive nature of emergent behaviors, we argue for the impossibility (...) of reliably foreseeing some capabilities. By investigating these impossibility results, we shed light on their potential implications for AI safety research and propose potential strategies to overcome these limitations. (shrink)

Recent progress in artificial intelligence (AI) has drawn attention to the technology’s transformative potential, including what some see as its prospects for causing large-scale harm. We review two influential arguments purporting to show how AI could pose catastrophic risks. The first argument — the Problem of Power-Seeking — claims that, under certain assumptions, advanced AI systems are likely to engage in dangerous power-seeking behavior in pursuit of their goals. We review reasons for thinking that AI systems might seek power, that (...) they might obtain it, that this could lead to catastrophe, and that we might build and deploy such systems anyway. The second argument claims that the development of human-level AI will unlock rapid further progress, culminating in AI systems far more capable than any human — this is the Singularity Hypothesis. Power-seeking behavior on the part of such systems might be particularly dangerous. We discuss a variety of objections to both arguments and conclude by assessing the state of the debate. (shrink)

This book reports on the results of the third edition of the premier conference in the field of philosophy of artificial intelligence, PT-AI 2017, held on November 4 - 5, 2017 at the University of Leeds, UK. It covers: advanced knowledge on key AI concepts, including complexity, computation, creativity, embodiment, representation and superintelligence; cutting-edge ethical issues, such as the AI impact on human dignity and society, responsibilities and rights of machines, as well as AI threats to humanity and AI (...) class='Hi'>safety; and cutting-edge developments in techniques to achieve AI, including machine learning, neural networks, dynamical systems. The book also discusses important applications of AI, including big data analytics, expert systems, cognitive architectures, and robotics. It offers a timely, yet very comprehensive snapshot of what is going on in the field of AI, especially at the interfaces between philosophy, cognitive science, ethics and computing. (shrink)

A classification of the global catastrophic risks of AI is presented, along with a comprehensive list of previously identified risks. This classification allows the identification of several new risks. We show that at each level of AI’s intelligence power, separate types of possible catastrophes dominate. Our classification demonstrates that the field of AI risks is diverse, and includes many scenarios beyond the commonly discussed cases of a paperclip maximizer or robot-caused unemployment. Global catastrophic failure could happen at various levels of (...) AI development, namely, before it starts self-improvement, during its takeoff, when it uses various instruments to escape its initial confinement, or after it successfully takes over the world and starts to implement its goal system, which could be plainly unaligned, or feature-flawed friendliness. AI could also halt at later stages of its development either due to technical glitches or ontological problems. Overall, we identified around several dozen scenarios of AI-driven global catastrophe. The extent of this list illustrates that there is no one simple solution to the problem of AI safety, and that AI safety theory is complex and must be customized for each AI development level. (shrink)

The use of machine learning, or “artificial intelligence” (AI) in medicine is widespread and growing. In this paper, I focus on a specific proposed clinical application of AI: using models to predict incapacitated patients’ treatment preferences. Drawing on results from machine learning, I argue this proposal faces a special moral problem. Machine learning researchers owe us assurance on this front before experimental research can proceed. In my conclusion I connect this concern to broader issues in AI safety.

The “Value Alignment Problem” is the challenge of how to align the values of artificial intelligence with human values, whatever they may be, such that AI does not pose a risk to the existence of humans. Existing approaches appear to conceive of the problem as "how do we ensure that AI solves the problem in the right way", in order to avoid the possibility of AI turning humans into paperclips in order to “make more paperclips” or eradicating the human race (...) to “solve climate change”. This paper proposes an approach to Alignment rooted in the Enactive theory of mind that reconceptualises it as "how do we make relevant to AI what it relevant to humans". This conceptualisation is supported with a discussion of 4E cognition and goes on to suggest that the Alignment problem and the Frame problem are the same problem. The paper concludes with a discussion of the tradeoffs of the different approaches to value alignment. (shrink)

A AI Robotics Ethics Society (AIRES) é uma organização sem fins lucrativos fundada em 2018 por Aaron Hui, com o objetivo de se promover a conscientização e a importância da implementação e regulamentação ética da AI. A AIRES é hoje uma organização com capítulos em universidade como UCLA (Los Angeles), USC (University of Southern California), Caltech (California Institute of Technology), Stanford University, Cornell University, Brown University e a Pontifícia Universidade Católica do Rio Grande do Sul (Brasil). AIRES na PUCRS é (...) o primeiro capítulo internacional da AIRES, e como tal, estamos empenhados em promover e aprimorar a Missão AIRES. Nossa missão é focar na educação dos líderes da IA de amanhã em princípios éticos, para assegurar que a IA seja criada de forma ética e responsável. Como ainda existem poucas propostas de como devemos implementar princípios éticos e diretrizes normativas na prática do desenvolvimento de sistemas de IA, o objetivo deste trabalho é tentar diminuir esta lacuna entre o discurso e a práxis. Entre princípios abstratos e implementação técnica. Neste trabalho, buscamos introduzir o leitor ao tema da Ética e Segurança da IA. Ao mesmo tempo, apresentamos uma série de ferramentas para auxiliar desenvolvedores de sistemas inteligentes a desenvolver “bons” modelos. Este trabalho é um guia em desenvolvimento publicado em inglês e português. Contribuições e sugestões são bem vindas. (shrink)

We discuss potential biorisks from large language models (LLMs). AI assistants based on LLMs such as ChatGPT have been shown to significantly reduce barriers to entry for actors wishing to synthesize dangerous, potentially novel pathogens and chemical weapons. The harms from deploying such bioagents could be further magnified by AI-assisted misinformation. We endorse several policy responses to these dangers, including prerelease evaluations of biomedical AIs by subject-matter experts, enhanced surveillance and lab screening procedures, restrictions on AI training data, and access (...) controls on biomedical AI. We conclude with a suggestion about future research directions in bioethics. (shrink)

Many researchers have conjectured that the humankind is simulated along with the rest of the physical universe – a Simulation Hypothesis. In this paper, we do not evaluate evidence for or against such claim, but instead ask a computer science question, namely: Can we hack the simulation? More formally the question could be phrased as: Could generally intelligent agents placed in virtual environments find a way to jailbreak out of them. Given that the state-of-the-art literature on AI containment answers in (...) the affirmative (AI is uncontainable in the long-term), we conclude that it should be possible to escape from the simulation, at least with the help of superintelligent AI. By contraposition, if escape from the simulation is not possible, containment of AI should be, an important theoretical result for AI safety research. Finally, the paper surveys and proposes ideas for hacking the simulation and analyzes ethical and philosophical issues of such an undertaking. (shrink)

The AGI alignment problem has a bimodal distribution of outcomes with most outcomes clustering around the poles of total success and existential, catastrophic failure. Consequently, attempts to solve AGI alignment should, all else equal, prefer false negatives (ignoring research programs that would have been successful) to false positives (pursuing research programs that will unexpectedly fail). Thus, we propose adopting a policy of responding to points of metaphysical and practical uncertainty associated with the alignment problem by limiting and choosing necessary assumptions (...) to reduce the risk false positives. Herein we explore in detail some of the relevant points of uncertainty that AGI alignment research hinges on and consider how to reduce false positives in response to them. (shrink)

Various thinkers have been attempting to align artificial intelligence (AI) with ethics (Christian, 2020; Russell, 2021), the so-called problem of alignment, but some suspect that the problem may be intractable (Yampolskiy, 2023). In the following, we make an argument by analogy to analyze the possibility that the problem of alignment could be intractable. We show how the Tri-Omni properties in theology can direct us towards analogous properties for artificial superintelligence, Tri-Opti properties. However, just as the Tri-Omni properties are vulnerable to (...) metaphysical incompatibility, we will show that the Tri-Opti properties are vulnerable to a corresponding physical incompatibility, because optimal physical systems exist on the Pareto Frontier. We will explain that, while Tri-Omni incompatibility thus seems metaphysically soluble, the Tri-Opti incompatibility seems physically insoluble due to the constraints of multi-variable optimization. Finally, we will provide some scenarios by which this incompatibility may be realized. Besides for the primary purpose of its theoretical considerations, this analysis has the secondary interdisciplinary purpose of communicating to the theologically literate the potential importance of AI safety work, and the gesturing towards the possibility of the collaboration in both fields on topics in ethics, epistemology, and ontology. (shrink)

I explain and motivate the shutdown problem: the problem of creating artificial agents that (1) shut down when a shutdown button is pressed, (2) don’t try to prevent or cause the pressing of the shutdown button, and (3) otherwise pursue goals competently. I then propose a solution: train agents to have incomplete preferences. Specifically, I propose that we train agents to lack a preference between every pair of different-length trajectories. I suggest a way to train such agents using reinforcement learning: (...) we give the agent lower reward for repeatedly choosing same-length trajectories. (shrink)

As ever more advanced digital systems are created, it becomes increasingly likely that some of these systems will be digital minds, i.e. digital subjects of experience. With digital minds comes the risk of digital suffering. The problem of digital suffering is that of mitigating this risk. We argue that the problem of digital suffering is a high stakes moral problem and that formidable epistemic obstacles stand in the way of solving it. We then propose a strategy for solving it: Access (...) Monitor Prevent (AMP). AMP uses a ‘dancing qualia’ argument to link the functional states of certain digital systems to their experiences—this yields epistemic access to digital minds. With that access, we can prevent digital suffering by only creating advanced digital systems that we have such access to, monitoring their functional profiles, and preventing them from entering states with functional markers of suffering. After introducing and motivating AMP, we confront limitations it faces and identify some options for overcoming them. We argue that AMP fits especially well with—and so provides a moral reason to prioritize—one approach to creating such systems: whole brain emulation. We also contend that taking other paths to digital minds would be morally risky. (shrink)

Some worry that advanced artificial agents may resist being shut down. The Incomplete Preferences Proposal (IPP) is an idea for ensuring that doesn't happen. A key part of the IPP is using a novel 'Discounted REward for Same-Length Trajectories (DREST)' reward function to train agents to (1) pursue goals effectively conditional on each trajectory-length (be 'USEFUL'), and (2) choose stochastically between different trajectory-lengths (be 'NEUTRAL' about trajectory-lengths). In this paper, we propose evaluation metrics for USEFULNESS and NEUTRALITY. We use a (...) DREST reward function to train simple agents to navigate gridworlds, and we find that these agents learn to be USEFUL and NEUTRAL. Our results thus suggest that DREST reward functions could also train advanced agents to be USEFUL and NEUTRAL, and thereby make these advanced agents useful and shutdownable. (shrink)

There is considerable enthusiasm about the prospect that artificial intelligence (AI) will help to improve the safety and efficacy of health services and the efficiency of health systems. To realize this potential, however, AI systems will have to overcome structural problems in the culture and practice of medicine and the organization of health systems that impact the data from which AI models are built, the environments into which they will be deployed, and the practices and incentives that structure their (...) development. This perspective elaborates on some of these structural challenges and provides recommendations to address potential shortcomings. (shrink)

The AGI alignment problem has a bimodal distribution of outcomes with most outcomes clustering around the poles of total success and existential, catastrophic failure. Consequently, attempts to solve AGI alignment should, all else equal, prefer false negatives (ignoring research programs that would have been successful) to false positives (pursuing research programs that will unexpectedly fail). Thus, we propose adopting a policy of responding to points of philosophical and practical uncertainty associated with the alignment problem by limiting and choosing necessary assumptions (...) to reduce the risk of false positives. Herein we explore in detail two relevant points of uncertainty that AGI alignment research hinges on---meta-ethical uncertainty and uncertainty about mental phenomena---and show how to reduce false positives in response to them. (shrink)

Explainability and comprehensibility of AI are important requirements for intelligent systems deployed in real-world domains. Users want and frequently need to understand how decisions impacting them are made. Similarly it is important to understand how an intelligent system functions for safety and security reasons. In this paper, we describe two complementary impossibility results (Unexplainability and Incomprehensibility), essentially showing that advanced AIs would not be able to accurately explain some of their decisions and for the decisions they could explain people (...) would not understand some of those explanations. (shrink)

Understanding computation as “a process of the dynamic change of information” brings to look at the different types of computation and information. Computation of information does not exist alone by itself but is to be considered as part of a system that uses it for some given purpose. Information can be meaningless like a thunderstorm noise, it can be meaningful like an alert signal, or like the representation of a desired food. A thunderstorm noise participates to the generation of meaningful (...) information about coming rain. An alert signal has a meaning as allowing a safety constraint to be satisfied. The representation of a desired food participates to the satisfaction of some metabolic constraints for the organism. Computations on information and representations will be different in nature and in complexity as the systems that link them have different constraints to satisfy. Animals have survival constraints to satisfy. Humans have many specific constraints coming in addition. And computers will compute what the designer and programmer ask for. We propose to analyze the different relations between information, meaning and representation by taking an evolutionary approach on the systems that link them. Such a bottom-up approach allows starting with simple organisms and avoids an implicit focus on humans, which is the most complex and difficult case. To make available a common background usable for the many different cases, we use a systemic tool that defines the generation of meaningful information by and for a system submitted to a constraint [Menant, 2003]. This systemic tool allows to position information, meaning and representations for systems relatively to environmental entities in an evolutionary perspective. We begin by positioning the notions of information, meaning and representation and recall the characteristics of the Meaning Generator System (MGS) that link a system submitted to a constraint to its environment. We then use the MGS for animals and highlight the network nature of the interrelated meanings about an entity of the environment. This brings us to define the representation of an item for an agent as being the network of meanings relative to the item for the agent. Such meaningful representations embed the agents in their environments and are far from the Good Old Fashion Artificial Intelligence type ones. The MGS approach is then used for humans with a limitation resulting of the unknown nature of human consciousness. Application of the MGS to artificial systems brings to look for compatibilities with different levels of Artificial Intelligence (AI) like embodied-situated AI, the Guidance Theory of Representations, and enactive AI. Concerns relative to different types of autonomy and organic or artificial constraints are highlighted. We finish by summarizing the points addressed and by proposing some continuations. (shrink)

This draft preprint presents a nine step argument for “Connectionist Structuralism” (CS), an account of the ontology of abstract objects that is neither purely nominalist nor purely platonist. CS is a common, often implicit assumption in parts of the artificial intelligence literature, but such discussions have not presented formal accounts of the position or engaged with metaphysical issues that potentially undermine it. By making the position legible and presenting an initial case for it, we hope to support a constructive dialogue (...) between AI researchers and philosophers of metaphysics that helps both sides to refine the position. -/- CS proposes that each abstract object we can draw on in human analysis corresponds to a particular subset of an individual person’s brain structure whose functionality is isomorphic to a subset of the nodes and connections in a suitable connectionist network. In other words, abstract objects are physically realised, but in individual brains, rather than only in the referent objects (pure nominalism) or in metaphysical universals (pure platonism). -/- This paper’s minimum claim is that CS can account for all abstract object predicables regarding sensible properties, such as “is red” or “is a square”. Using evidence from cognitive neuroscience, machine learning, and evolutionary biology, as well as a fully traceable toy example, we describe how CS can support our core cognitive uses of such sensible properties and can account for our core phenomenal experiences of them. In the former, CS provides sufficient albeit imperfect inferential safety, whose limitations are argued to strengthen rather than weaken the case for CS as describing human behaviour. In the latter, four target phenomenal features are accounted for – abstract objects as feeling intangible, non-located, transparent, and unchanging - along with accounts for further phenomena such as semantic refinement, the Stroop effect, synaesthesia, semantic clarity, sensory overload, and satiation. -/- Our minimum claim concerns the day-to-day usage and felt experience of abstract objects, but we suggest also an extended claim in which CS can form the basis of a pragmatic sufficient logic. As such, the initial outline of a response is provided for five common objections to positions that seek to ground abstract objects without reference to metaphysically stand-alone universals: referential opacity; identity of indiscernibles; infinite regression; non-physical concepts; and necessary truths. These outlines lay the foundation for (but do not seek to formally demonstrate) an extended CS account that addresses other abstract objects, including issues relating to their use in mathematics and formal logic. -/- The CS account leads to a four-layer hierarchy of similarity for whether your “red” is the same as mine. Considering both semantic and phenomenal similarity, we conclude that our “reds” are likely non-identical but can be made close enough for practical purposes. Finally, we describe how future work could elaborate CS as a metaphysical project and how confidence in it could be tested through empirical research. (shrink)

Some years ago I reached the point where I can usually tell from the title of a book, or at least from the chapter titles, what kinds of philosophical mistakes will be made and how frequently. In the case of nominally scientific works these may be largely restricted to certain chapters which wax philosophical or try to draw general conclusions about the meaning or long term significance of the work. Normally however the scientific matters of fact are generously interlarded with (...) philosophical gibberish as to what these facts mean. The clear distinctions which Wittgenstein described some 80 years ago between scientific matters and their descriptions by various language games are rarely taken into consideration, and so one is alternately wowed by the science and dismayed by its incoherent analysis. So it is with this volume. -/- If one is to create a mind more or less like ours, one needs to have a logical structure for rationality and an understanding of the two systems of thought (dual process theory). If one is to philosophize about this, one needs to understand the distinction between scientific issues of fact and the philosophical issue of how language works in the context at issue, and of how to avoid the pitfalls of reductionism and scientism, but Kurzweil, like nearly all students of behavior, is largely clueless. He, is enchanted by models, theories, and concepts, and the urge to explain, while Wittgenstein showed us that we only need to describe, and that theories, concepts etc., are just ways of using language (language games) which have value only insofar as they have a clear test (clear truthmakers, or as John Searle (AI’s most famous critic) likes to say, clear Conditions of Satisfaction (COS)). I have attempted to provide a start on this in my recent writings, such as The Logical Structure of Consciousness (behavior, personality, rationality, higher order thought, intentionality) (2016) and The Logical Structure of Philosophy, Psychology, Mind and Language as Revealed in the Writings of Ludwig Wittgenstein and John Searle (2016). Those interested in all my writings in their most recent versions may consult my e-book Philosophy, Human Nature and the Collapse of Civilization - Articles and Reviews 2006-2016 662p (2016). I will give a very brief presentation of this framework since I have described it in great detail in many recent papers and several books, available on this site and others. -/- Also, as usual in ‘factual’ accounts of AI/robotics, he gives no time to the very real threats to our privacy, safety and even survival from the increasing ‘androidizing’ of society which is prominent in other authors (Bostrum, Hawking etc.) and frequent in scifi and films, so I make a few comments on the quite possibly suicidal utopian delusions of ‘nice’ androids, humanoids, democracy, diversity, and genetic engineering. -/- I take it for granted that technical advances in electronics, robotics and AI will occur, resulting in profound changes in society. However, I think the changes coming from genetic engineering are at least as great and potentially far greater, as they will enable us to utterly change who we are. And it will be feasible to make supersmart/super strong servants by modifying our genes or those of other monkeys. As with other technology, any country that resists will be left behind. But will it be socially and economically feasible to implement biobots or superhumans on a massive scale? And even if so, it does not seem remotely possible, economically or socially to prevent the collapse of industrial civilization. So, ignoring the philosophical mistakes in this volume as irrelevant, and directing our attention only to the science, what we have here is another suicidal utopian delusion rooted in a failure to grasp basic biology, psychology and human ecology, the same delusions that are destroying America and the world. I see a remote possibility the world can be saved, but not by AI/robotics,CRISPR, nor by democracy and equality. -/- Those wishing a comprehensive up to date framework for human behavior from the modern two systems view may consult my book ‘The Logical Structure of Philosophy, Psychology, Mind and Language in Ludwig Wittgenstein and John Searle’ 2nd ed (2019). Those interested in more of my writings may see ‘Talking Monkeys--Philosophy, Psychology, Science, Religion and Politics on a Doomed Planet--Articles and Reviews 2006-2019 3rd ed (2019), The Logical Structure of Human Behavior (2019), and Suicidal Utopian Delusions in the 21st Century 4th ed (2019). (shrink)

The potential for Large Language Models (LLMs) to attain technological singularity—the point at which artificial intelligence (AI) surpasses human intellect and autonomously improves itself—is a critical concern in AI research. This paper explores the feasibility of current LLMs achieving singularity by examining the philosophical and practical requirements for such a development. We begin with a historical overview of AI and intelligence amplification, tracing the evolution of LLMs from their origins to state-of-the-art models. We then proposes a theoretical framework to assess (...) whether existing LLM technologies could satisfy the conditions for singularity, with a focus on Recursive Self-Improvement (RSI) and autonomous code generation. We integrate key component technologies, such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO), into our analysis, illustrating how these could enable LLMs to independently enhance their reasoning and problem-solving capabilities. By mapping out a potential singularity model lifecycle and examining the dynamics of exponential growth models, we elucidate the conditions under which LLMs might self-replicate and rapidly escalate their intelligence. We conclude with a discussion of the ethical and safety implications of such developments, underscoring the need for responsible and controlled advancement in AI research to mitigate existential risks. Our work aims to contribute to the ongoing dialogue on the future of AI and the critical importance of proactive measures to ensure its beneficial development. (shrink)