Tag: Epistemology

Posted on

(Featured) Autonomous AI Systems in Conflict: Emergent Behavior and Its Impact on Predictability and Reliability

Autonomous AI Systems in Conflict: Emergent Behavior and Its Impact on Predictability and Reliability

Daniel Trusilo investigates the concept of emergent behavior in complex autonomous systems and its implications in dynamic, open context environments such as conflict scenarios. In a nuanced exploration of the intricacies of autonomous systems, the author employs two hypothetical case studies—an intelligence, surveillance, and reconnaissance (ISR) maritime swarm system and a next-generation autonomous humanitarian notification system—to articulate and elucidate the effects of emergent behavior.

In the case of the ISR swarm system, the author underscores how the autonomous algorithm’s unpredictable micro-level behavior can yield reliable macro-level outcomes, enhancing the system’s robustness and resilience against adversarial interventions. Conversely, the humanitarian notification system emphasizes how such systems’ unpredictability can fortify International Humanitarian Law (IHL) compliance, reducing civilian harm, and increasing accountability. Thus, the author emphasizes the dichotomy of emergent behavior: it enhances system reliability and effectiveness while posing novel challenges to predictability and system certification.

Navigating these challenges, the author calls attention to the implications for system certification and ethical interoperability. With the potential for these systems to exhibit unforeseen behavior in actual operations, traditional testing, evaluation, verification, and validation methods seem inadequate. Instead, the author suggests adopting dynamic certification methods, allowing the systems to be continually monitored and adjusted in complex, real-world environments, thereby accommodating emergent behavior. Ethical interoperability, the concurrence of ethical AI principles across different organizations and nations, presents another conundrum, especially with differing ethical guidelines governing AI use in defense.

In its broader philosophical framework, the article contributes to the ongoing discourse on the ethics and morality of AI and autonomous systems, particularly within the realm of futures studies. It underscores the tension between the benefits of autonomous systems and the ethical, moral, and practical challenges they pose. The emergent behavior phenomenon can be seen as a microcosm of the larger issues in AI ethics, reflecting on themes of predictability, control, transparency, and accountability. The navigation of these ethical quandaries implies the need for shared ethical frameworks and standards that can accommodate the complex, unpredictable nature of these systems without compromising the underlying moral principles.

In terms of future research, there are several critical avenues to explore. The implications of emergent behavior in weaponized autonomous systems need careful examination, questioning acceptable risk confidence intervals for such systems’ predictability and reliability. Moreover, the impact of emergent behavior on operator trust and the ongoing issue of machine explainability warrants further exploration. Lastly, it would be pertinent to identify methods of certifying complex autonomous systems while addressing the burgeoning body of distinct, organization-specific ethical AI principles. Such endeavors would help operationalize these principles in light of emergent behavior, thereby contributing to the development of responsible, accountable, and effective AI systems.

Abstract

The development of complex autonomous systems that use artificial intelligence (AI) is changing the nature of conflict. In practice, autonomous systems will be extensively tested before being operationally deployed to ensure system behavior is reliable in expected contexts. However, the complexity of autonomous systems means that they will demonstrate emergent behavior in the open context of real-world conflict environments. This article examines the novel implications of emergent behavior of autonomous AI systems designed for conflict through two case studies. These case studies include (1) a swarm system designed for maritime intelligence, surveillance, and reconnaissance operations, and (2) a next-generation humanitarian notification system. Both case studies represent current or near-future technology in which emergent behavior is possible, demonstrating that such behavior can be both unpredictable and more reliable depending on the level at which the system is considered. This counterintuitive relationship between less predictability and more reliability results in unique challenges for system certification and adherence to the growing body of principles for responsible AI in defense, which must be considered for the real-world operationalization of AI designed for conflict environments.

Autonomous AI Systems in Conflict: Emergent Behavior and Its Impact on Predictability and Reliability
Posted on

(Featured) Epistemic diversity and industrial selection bias

Epistemic diversity and industrial selection bias

Manuela Fernández Pinto and Daniel Fernández Pinto offer a compelling examination of the role that funding sources play in shaping scientific consensus, focusing specifically on the influence of private industry. Drawing on the work of Holman and Bruner (2017), the authors use a reinforcement learning model, known as a Q-learning model, to explore industrial selection. The central concept of industrial selection posits that, rather than corrupting individual scientists, private industry can subtly steer scientific outcomes towards their interests by selectively funding research. In the authors’ simulation, three different funding scenarios are considered: research funded solely by industry, research funded solely by a random agent, and research jointly funded by industry and a random agent.

Results from the simulations reinforce the effects of industrial selection observed by Holman and Bruner, showing a divergence from correct scientific hypotheses under sole industry funding. When scientists are funded solely by a random agent, the outcomes are closer to the correct hypothesis. Most notably, when funding is a mix of industry and random allocation, the random element appears to counteract, or at least delay, the bias introduced by industry funding. The authors further observe an unexpected and somewhat paradoxical interaction with methodological diversity, a factor traditionally seen as a strength in scientific communities. Industrial funding effectively exploits this diversity to skew consensus towards industry-friendly outcomes.

The authors then introduce a provocative and potentially contentious suggestion based on their simulations. They propose that a random allocation of funding might be a more effective countermeasure against industrial selection bias than the commonly held belief in meritocratic funding systems, which might inadvertently perpetuate industry bias. This suggestion arises from their observation that a random funding agent in the simulation effectively obstructs industrial selection bias. They also consider the merits and drawbacks of a two-stage random allocation system, wherein only research proposals that pass an initial quality assessment are subject to a subsequent funding lottery.

The study raises compelling philosophical considerations on the influence of funding on the direction and integrity of scientific research. It challenges the common narrative that methodological diversity and meritocracy inherently lead to unbiased, high-quality science, suggesting instead that these can be co-opted by industry to push scientific consensus toward commercially advantageous outcomes. It further incites reflection on the ethical implications of allowing commercial interests to potentially manipulate scientific consensus and the responsibility of society to ensure the pursuit of truth in science. The research also ties into broader discussions on the balance between rational decision-making and randomness, and the potential role of randomness as a mitigating factor in decision-making processes rife with bias or undue influence.

Future research could delve deeper into how a random allocation system might be implemented in practice, particularly regarding the initial quality assessment process. It would also be beneficial to explore how such a system could coexist with traditional funding sources, and what percentage of overall funding would need to be randomly allocated to effectively mitigate industrial selection bias. Additionally, more nuanced simulations could help further untangle the complex relationship between methodological diversity and industrial bias, and identify other possible factors that may be manipulated to sway scientific consensus. Ultimately, this research presents a provocative stepping stone for further exploration into the complex and subtle ways commercial interests may influence scientific research, and potential innovative strategies to counteract such influences.

Abstract

Philosophers of science have argued that epistemic diversity is an asset for the production of scientific knowledge, guarding against the effects of biases, among other advantages. The growing privatization of scientific research, on the contrary, has raised important concerns for philosophers of science, especially with respect to the growing sources of biases in research that it seems to promote. Recently, Holman and Bruner (2017) have shown, using a modified version of Zollman (2010) social network model, that an industrial selection bias can emerge in a scientific community, without corrupting any individual scientist, if the community is epistemically diverse. In this paper, we examine the strength of industrial selection using a reinforcement learning model, which simulates the process of industrial decision-making when allocating funding to scientific projects. Contrary to Holman and Bruner’s model, in which the probability of success of the agents when performing an action is given a priori, in our model the industry learns about the success rate of individual scientists and updates the probability of success on each round. The results of our simulations show that even without previous knowledge of the probability of success of an individual scientist, the industry is still able to disrupt scientific consensus. In fact, the more epistemically diverse the scientific community, the easier it is for the industry to move scientific consensus to the opposite conclusion. Interestingly, our model also shows that having a random funding agent seems to effectively counteract industrial selection bias. Accordingly, we consider the random allocation of funding for research projects as a strategy to counteract industrial selection bias, avoiding commercial exploitation of epistemically diverse communities.

Epistemic diversity and industrial selection bias
Posted on

(Featured) Liars and Trolls and Bots Online: The Problem of Fake Persons

Liars and Trolls and Bots Online: The Problem of Fake Persons

Keith Raymond Harris explores of the role of ‘fake persons’—bots and trolls—in online spaces and their deleterious impact on our acquisition and distribution of knowledge. Situating his analysis in a technological ecosystem increasingly swamped by these artificial entities, the author dissects the intricate issues engendered by these ‘fake persons’ into three discernible yet interwoven threats: deceptive, skeptical, and epistemic.

The deceptive threat elucidates how bots and trolls propagate false information and craft misleading representations of consensus through manipulated metrics like shares, likes, and comments. This deceptive veneer engenders a distorted perception of reality, leading to the formulation of misguided beliefs. The skeptical threat, on the other hand, stems from the awareness of the online environment’s infestation with these deceitful entities. This awareness engenders a pervasive sense of skepticism, a defensive mechanism that could result in the dismissal of valid evidence, leading to an overall decrease in the trust placed in online information. This skepticism, though justifiable, can have the unintended effect of isolating individuals from genuine knowledge sources.

Further complicating this scenario is the epistemic threat. The author draws a striking analogy between the online world inhabited by ‘fake persons’ and a natural environment populated by ‘mimic species’. In the latter, the significance of certain traits, often used to identify species, diminishes due to the presence of mimics. Analogously, in an environment teeming with bots and trolls, the perceived value of certain forms of evidence depreciates, impairing the ability to discern ‘real’ persons. In this convoluted digital milieu, the credibility of evidence—along with the authenticity of users and the perceived consensus—becomes questionable.

Grounding these digital threats in the wider philosophical discourse, this research accentuates the intricate entanglement of epistemology and ontology in online spaces. It challenges traditional conceptions of identity, reality, and knowledge, echoing Baudrillard’s premonitions of hyperreality and simulation. The presence of ‘fake persons’ obfuscates the demarcation between the real and the artificial, leading to an epistemic crisis where distinguishing between genuine and fallacious information becomes a Herculean task. Furthermore, these digital distortions provoke a profound skepticism that resonates with Cartesian doubt, while simultaneously illustrating the pervasiveness of misinformation and disinformation, reflecting the post-truth era’s cynicism. This research, hence, not only deepens our understanding of the digital world’s complexities but also underscores the shifting epistemic and ontological paradigms in the internet age.

As we navigate through this rapidly mutating digital landscape, the author’s research underscores the urgent need for further exploration. While technological solutions might offer some respite, they cannot completely eradicate these pervasive threats. Future research, therefore, should venture into developing more robust epistemological frameworks that accommodate these digital complexities. It should aim to delve into the philosophy of digital identities, exploring how they are constructed, perceived, and interacted with. There’s also a pressing need for studies that examine the intersection of ethics, technology, and epistemology, especially in the context of ‘fake persons’. Such research would not only enrich the theoretical discourse but could also guide the creation of more ethical and reliable digital spaces.

Abstract

This paper describes the ways in which trolls and bots impede the acquisition of knowledge online. I distinguish between three ways in which trolls and bots can impede knowledge acquisition, namely, by deceiving, by encouraging misplaced skepticism, and by interfering with the acquisition of warrant concerning persons and content encountered online. I argue that these threats are difficult to resist simultaneously. I argue, further, that the threat that trolls and bots pose to knowledge acquisition goes beyond the mere threat of online misinformation, or the more familiar threat posed by liars offline. Trolls and bots are, in effect, fake persons. Consequently, trolls and bots can systemically interfere with knowledge acquisition by manipulating the signals whereby individuals acquire knowledge from one another online. I conclude with a brief discussion of some possible remedies for the problem of fake persons.

Liars and Trolls and Bots Online: The Problem of Fake Persons
Posted on

(Featured) Against AI Understanding and Sentience: Large Language Models, Meaning, and the Patterns of Human Language Use

Against AI Understanding and Sentience: Large Language Models, Meaning, and the Patterns of Human Language Use

Christoph Durt, Thomas Fuchs, and Tom Froese investigate the astonishing capacities of Large Language Models (LLMs) to mimic human-like responses. They begin by acknowledging the unprecedented feats of these models, particularly GPT-3, which have led some to assert that they possess common-sense reasoning and even sentience. They caution, however, that these claims often overlook the instances where LLMs fail to produce sensical responses. Even as the models evolve and mitigate some of these limitations, the authors urge circumspection regarding the attribution of understanding and sentience to these systems.

The authors argue that the progress of LLMs invites a reassessment of long-standing philosophical debates about the limits of AI. The authors challenge the view, expressed by philosophers such as Hubertus Dreyfus, that AI is inherently incapable of understanding meaning. Given the emergent linguistic capabilities of these models, they query whether these advancements warrant attributing understanding to the computational system. Contrary to Dreyfus’s assertion that any formal system cannot be directly sensitive to the relevance of its situation, the authors propose that LLMs seem to exhibit this sensitivity in a pragmatic sense.

While the article explores the philosophical debates surrounding AI understanding and sentience, it does not definitively conclude whether LLMs truly understand or are sentient. The authors suggest that the human-like behaviour exhibited by these models may lead to the inference of a human-like mind. However, they argue that more nuanced and empirically informed positions are required. The authors further advocate for a more comprehensive assessment of LLM output, rather than relying on selective instances of impressive performance.

This research brings into focus the broader philosophical implications of our interaction with AI, particularly the ontological and epistemological assumptions we make when interacting with LLMs. The debate surrounding AI sentience and understanding illuminates the complexities inherent in defining consciousness and understanding, a philosophical quandary that dates back to Descartes and beyond. It forces us to interrogate the nature of understanding – is it a purely human phenomenon, or can it be replicated, even surpassed, by silicon-based entities? Moreover, it challenges our anthropocentric views of cognition and compels us to consider alternate forms of intelligence and understanding.

Looking forward, the study of AI and philosophy would benefit from an even deeper exploration of these questions. More empirical research is needed to understand the extent and limitations of LLMs’ capacities. Concurrently, philosophical inquiry can help define and refine the metrics by which we measure AI understanding and sentience. As we delve further into the AI era, it is crucial that we continue to scrutinize and challenge our assumptions about AI capabilities, not only to enhance our technological advancements but also to enrich our philosophical understanding of the world.

Abstract

Large language models such as ChatGPT are deep learning architectures trained on immense quantities of text. Their capabilities of producing human-like text are often attributed either to mental capacities or the modeling of such capacities. This paper argues, to the contrary, that because much of meaning is embedded in common patterns of language use, LLMs can model the statistical contours of these usage patterns. We agree with distributional semantics that the statistical relations of a text corpus reflect meaning, but only part of it. Written words are only one part of language use, although an important one as it scaffolds our interactions and mental life. In human language production, preconscious anticipatory processes interact with conscious experience. Human language use constitutes and makes use of given patterns and at the same time constantly rearranges them in a way we compare to the creation of a collage. LLMs do not model sentience or other mental capacities of humans but the common patterns in public language use, clichés and biases included. They thereby highlight the surprising extent to which human language use gives rise to and is guided by patterns.

Against AI Understanding and Sentience: Large Language Models, Meaning, and the Patterns of Human Language Use
Posted on

(Featured) A Loosely Wittgensteinian Conception of the Linguistic Understanding of Large Language Models like BERT, GPT-3, and ChatGPT

A Loosely Wittgensteinian Conception of the Linguistic Understanding of Large Language Models like BERT, GPT-3, and ChatGPT

Reto Gubelmann articulates a “loosely Wittgensteinian” conception of linguistic understanding, particularly in the context of advanced artificial intelligence (AI) models such as BERT, GPT-3, and ChatGPT. The author posits that these transformer-based natural language processing (NNLP) models are closing in on the capacity to genuinely understand language, a claim that is buttressed by both empirical and conceptual arguments. The empirical basis is grounded on the remarkable performance of these AI models on benchmarks like GLUE and SuperGLUE, which evaluate them on tasks that, in a human context, would necessitate a deep understanding of language, such as answering questions about a text, summarizing text, and discerning logical relationships between statements. The conceptual underpinnings of this claim draw upon the works of Glock, Taylor, and Wittgenstein to argue that linguistic understanding, a form of intelligence, is marked by flexibility in handling new tasks and novel inputs, as well as the capability to autonomously adapt to new tasks​​.

The article further navigates through the terrain of philosophical objections to the idea that AI can understand language. The author counters objections raised by Searle, Bender, Koller, Davidson, and Nagel, among others, arguing that understanding language does not necessitate any esoteric or mysterious component such as qualia. Rather, it is dependent on the competencies of the AI model, specifically its autonomous adaptability and performance in a wide array of linguistic tasks in diverse settings. By this definition, the author contends that current transformer-based NNLP models are inching closer to meeting the criteria for linguistic understanding​​.

The author also provides a succinct yet comprehensive overview of the evolution of AI models, from the era of “Good Old-Fashioned AI” (GOFAI), which relied on explicit rules and logical processing, to the emergence of neural network models or connectionist AI, which represent a fundamentally different approach to designing intelligent systems. The distinguishing feature of these neural network models, such as the transformer-based models under discussion, is their learning-based approach, which enables them to adapt to new tasks and exhibit flexibility in the face of novel inputs​​.

Embedding these discussions within broader philosophical issues, the article provides a fruitful platform for exploring the nature of intelligence, understanding, and language. The examination of whether or not AI models can understand language opens up questions about the definition of understanding and the conditions that must be met to ascribe understanding to a being. This interrogation is undergirded by a Wittgensteinian perspective, which has profound implications for our understanding of language, mind, and the possibilities of AI. It also prompts us to reconsider the boundaries we draw between human and machine intelligence.

Future research should continue to explore these Wittgensteinian conceptions of linguistic understanding, particularly as AI models continue to evolve and improve. More empirical work could be conducted to test the adaptability and flexibility of AI models in novel linguistic situations, providing more robust evidence for or against their capacity to understand language. Furthermore, the philosophical debate concerning language understanding in AI should continue to be pushed forward, with deeper explorations of the arguments against AI understanding and the development of new philosophical frameworks that can accommodate the rapidly advancing capabilities of AI. As this field advances, interdisciplinary collaboration between AI researchers, linguists, and philosophers will be vital in order to fully grasp the implications of these transformative technologies.

Abstract

In this article, I develop a loosely Wittgensteinian conception of what it takes for a being, including an AI system, to understand language, and I suggest that current state of the art systems are closer to fulfilling these requirements than one might think. Developing and defending this claim has both empirical and conceptual aspects. The conceptual aspects concern the criteria that are reasonably applied when judging whether some being understands language; the empirical aspects concern the question whether a given being fulfills these criteria. On the conceptual side, the article builds on Glock’s concept of intelligence, Taylor’s conception of intrinsic rightness as well as Wittgenstein’s rule-following considerations. On the empirical side, it is argued that current transformer-based NNLP models, such as BERT and GPT-3 come close to fulfilling these criteria.

A Loosely Wittgensteinian Conception of the Linguistic Understanding of Large Language Models like BERT, GPT-3, and ChatGPT
Posted on

(Featured) The black box problem revisited. Real and imaginary challenges for automated legal decision making

The black box problem revisited. Real and imaginary challenges for automated legal decision making

Bartosz Brożek et al. explore the ethical and practical dilemmas arising from the integration of Artificial Intelligence (AI) in the realm of law. The authors suggest that despite the perceived opacity and unpredictability of AI, these machines can provide rational and justifiable decisions in legal reasoning. By challenging conventional notions of decision-making and justifiability, the paper reframes the discussion around AI’s role in law and provides a compelling argument for AI’s potential to aid in legal reasoning.

The authors delve into the intricacies of legal decision-making, highlighting the contrast between our traditional expectations and the realities of legal reasoning. They argue that while we expect legal decisions to be based on clearly identifiable structures, algorithmic operations on beliefs, and classical logic, the cognitive science research paints a contrasting picture. The authors further suggest that most legal decisions emerge unconsciously, lack a recognizable structure, and are often influenced by emotional reactions and social training. This observation paves the way for a paradigm shift, suggesting that rather than the process, it is the justifiability of the decision ex post that is paramount.

The authors propose a two-module AI system, one intuitive and the other rational. The intuitive module, powered by machine learning, recognizes patterns from large datasets and makes decisions. The rational module, grounded in logic, does not make decisions but justifies those made by the intuitive module. In this framework, AI can be seen as rational if an acceptable justification can be provided for its decisions, despite their unpredictability. This interesting intertwining of machine learning and logic reshapes our understanding of AI’s role in legal decision-making.

This paper touches upon broader philosophical issues surrounding consciousness, rationality, and decision-making. By arguing for a shift from a process-oriented to a result-oriented evaluation of decision-making, the authors challenge the traditional Kantian perspective. The proposed model, in which an AI’s decisions are assessed based on their post-hoc justifiability, aligns more closely with consequentialist philosophy. This emphasis on the end result rather than the means to reach it further stimulates the ongoing debate on the ethical implications of AI use and the re-evaluation of long-held philosophical tenets in the face of technological advancements.

Future research could explore various facets of this proposed two-module AI system, particularly the interplay and potential conflicts between the intuitive and rational modules. Questions around what constitutes an “acceptable justification” in various legal contexts also demand further exploration. Additionally, research could investigate how this approach to AI in law would intersect with other legal principles, such as fairness, transparency, and due process. Ultimately, the paper presents a compelling case for rethinking the role and evaluation of AI in legal decision-making, opening up intriguing possibilities for future philosophical and legal discourse.

Abstract

This paper addresses the black-box problem in artificial intelligence (AI), and the related problem of explainability of AI in the legal context. We argue, first, that the black box problem is, in fact, a superficial one as it results from an overlap of four different – albeit interconnected – issues: the opacity problem, the strangeness problem, the unpredictability problem, and the justification problem. Thus, we propose a framework for discussing both the black box problem and the explainability of AI. We argue further that contrary to often defended claims the opacity issue is not a genuine problem. We also dismiss the justification problem. Further, we describe the tensions involved in the strangeness and unpredictability problems and suggest some ways to alleviate them.

The black box problem revisited. Real and imaginary challenges for automated legal decision making
Posted on

(Featured) Bare statistical evidence and the legitimacy of software-based judicial decisions

Bare statistical evidence and the legitimacy of software-based judicial decisions

Eva Schmidt et al. explore the question of whether evidence provided by software systems can serve as a legitimate basis for judicial decisions, focusing on two primary cases: recidivism predictions and DNA cold hit cases. The authors approach this question by analyzing the nature of bare statistical evidence and its relation to individualized evidence. They argue that while bare statistical evidence is generally considered insufficient to meet the standard of proof in criminal and civil cases, software-generated evidence can be individualized and, thus, meet this standard of proof under certain conditions.

In the case of recidivism predictions, the authors discuss the use of software systems such as COMPAS, which rely on bare statistical evidence to estimate the risk of an individual reoffending. They argue that for a sentence to be just and have the potential to serve as an incentive, it must be based on the specific features of the individual concerned, rather than solely on general features of a group that they belong to, which may correlate with high recidivism risk. The authors maintain that bare statistical evidence alone is insufficient for sentencing decisions.

Regarding DNA cold hit cases, the authors propose that statistical evidence generated by software systems like TrueAllele can be individualized through abductive reasoning or inference to the best explanation when it comes to cases of extreme probability. They argue that when the best explanation for the evidence is that the defendant is the source of the crime scene DNA, the evidence can be considered individualized and, thus, meet the standard of proof for criminal cases. This aligns with the normic account of individualized evidence, which posits that the best explanation of a piece of evidence is also the most normal explanation.

The authors’ analysis raises broader philosophical questions concerning the nature of evidence, the role of statistical reasoning in judicial decision-making, and the ethical implications of using software systems in the courtroom. It highlights the importance of distinguishing between different types of support, such as abductive (normic) support and probabilistic support, and of understanding the connections and disconnections between these concepts. Moreover, the paper touches on issues related to transparency, explainability, and fairness in the use of software systems as decision-making aids in the legal context.

Future research could further explore the implications of using software systems for other types of legal evidence and decision-making processes, as well as the ethical and epistemological challenges that these systems pose. Additionally, investigating the relationship between individualized evidence and statistical reasoning could shed light on the nature of evidence itself and the standards of proof required in various legal contexts. Finally, future work could focus on the development of guidelines and best practices for the implementation and evaluation of software systems in the courtroom, addressing issues such as transparency, explainability, and the appropriate weighting of statistical and individualized evidence.

Abstract

Can the evidence provided by software systems meet the standard of proof for civil or criminal cases, and is it individualized evidence? Or, to the contrary, do software systems exclusively provide bare statistical evidence? In this paper, we argue that there are cases in which evidence in the form of probabilities computed by software systems is not bare statistical evidence, and is thus able to meet the standard of proof. First, based on the case of State v. Loomis, we investigate recidivism predictions provided by software systems used in the courtroom. Here, we raise problems for software systems that provide predictions that are based on bare statistical evidence. Second, by examining the case of People v. Chubbs, we argue that the statistical evidence provided by software systems in cold hit DNA cases may in some cases suffice for individualized evidence, on a view on which individualized evidence is evidence that normically supports the relevant proposition (Smith, in Mind 127:1193–1218, 2018).

Bare statistical evidence and the legitimacy of software-based judicial decisions
Posted on

(Featured) Might text-davinci-003 have inner speech?

Might text-davinci-003 have inner speech?

Stephen Francis Mann and Daniel Gregory endeavor to explore the possibility of inner speech in artificial intelligence, specifically within an AI assistant. The researchers employ a Turing-like test, which involves a conversation with a chatbot to assess its linguistic competence, creativity, and reasoning. Throughout the experiment, the chatbot is asked a series of questions designed to probe its capabilities and discern whether it possesses the capacity for inner speech.

The researchers find mixed evidence to support the presence of inner speech in the AI chatbot. While the chatbot claims to have inner speech, its performance on sentence completion tasks somewhat corroborates this assertion. However, its inconsistent performance on rhyme-detection tasks, particularly when involving non-words, raises doubts regarding the presence of inner speech. The authors also note that the chatbot’s responses can be explained by its highly advanced autocomplete capabilities, which further complicates the evaluation of its inner speech.

Ultimately, the paper questions the efficacy of Turing-like tests as a means to determine mental states or mind-like properties in artificial agents. It suggests that linguistic competence alone may not be sufficient to ascertain whether AI possesses mind-like properties such as inner speech. The authors imply that those who argue against the plausibility of mental states in AI agents might reason that the absence of minds in conversation agents proves that linguistic competence is an insufficient test for mind-like properties.

This research taps into broader philosophical issues, such as the nature of consciousness and the criteria required to attribute mental states to artificial agents. As AI continues to advance, the demarcation between human and machine becomes increasingly blurred, forcing us to reevaluate our understanding of concepts like inner speech and consciousness. The question of whether AI can possess inner speech underscores the need for a more robust philosophical framework that can accommodate the unique characteristics and capabilities of artificial agents.

Future research in this domain could benefit from exploring alternative methods for evaluating inner speech in AI, going beyond Turing-like tests. For instance, researchers might investigate the AI’s decision-making processes or the mechanisms that underpin its creativity. Additionally, interdisciplinary collaboration with fields such as cognitive science and neuroscience could shed light on the cognitive processes at play in both humans and AI agents, thus providing a richer context for understanding the nature of inner speech in artificial agents. By expanding the scope of inquiry, we can better assess the extent to which AI agents possess mind-like properties and develop a more nuanced understanding of the implications of such findings for the future of AI and human cognition.

Abstract

In November 2022, OpenAI released ChatGPT, an incredibly sophisticated chatbot. Its capability is astonishing: as well as conversing with human interlocutors, it can answer questions about history, explain almost anything you might think to ask it, and write poetry. This level of achievement has provoked interest in questions about whether a chatbot might have something similar to human intelligence or even consciousness. Given that the function of a chatbot is to process linguistic input and produce linguistic output, we consider the question whether a sophisticated chatbot might have inner speech. That is: Might it talk to itself, internally? We explored this via a conversation with ‘Playground’, a chatbot which is very similar to ChatGPT but more flexible in certain respects. We asked it questions which, plausibly, can only be answered if one first produces some inner speech. Here, we present our findings and discuss their philosophical significance.

Might text-davinci-003 have inner speech?
Posted on

(Featured) Deepfakes and the epistemic apocalypse

Deepfakes and the epistemic apocalypse

Joshua Habgood-Cooter critically examines the common perception that deepfakes represent a unique and unprecedented threat to our epistemic landscape. They argue that such a viewpoint is misguided and that deepfakes should be understood as a social problem rather than a purely technological one. The author offers three main lines of criticism to counter the narrative of deepfakes as harbingers of an epistemic apocalypse. First, they propose that the knowledge we gain from recordings is a special case of knowledge from instruments, which relies on social practices around the design, operation, and maintenance of recording technology. Second, they present historical examples of manipulated recordings to demonstrate that deepfakes are not a novel phenomenon, and that social practices have been employed in the past to address similar issues. Third, they contend that technochauvinism and the post-truth narrative have obscured potential social measures to address deepfakes.

The author argues that deepfakes are embedded in a techno-social context and should be treated as part of the broader social practices involved in the production of knowledge and ignorance. They suggest that examining historical episodes of deceptive recordings can provide valuable insights into how social norms and community policing could be utilized to address the challenges posed by deepfakes. Moreover, the author emphasizes that the most serious harms associated with deepfake videos are likely to be consequences of established ignorance-producing social practices affecting minority and marginalized groups.

By reframing deepfakes as a social problem, the paper challenges the notion that the technology itself is inherently dangerous and urges us to consider how our social practices contribute to the production and dissemination of manipulated recordings. This approach highlights the interdependence between technology and society, and offers a more nuanced understanding of the ethical, political, and epistemic implications of deepfakes.

In the broader philosophical context, this paper raises important questions about the nature of knowledge, the role of trust in our epistemic practices, and the relationship between technology and the social dynamics of knowledge production. It also contributes to ongoing debates in social epistemology, emphasizing the collective nature of knowledge and the responsibility that society bears in shaping our epistemic landscape.

Future research could explore other historical episodes of manipulated recordings and the social responses that emerged to address them, further informing our understanding of how to manage the challenges posed by deepfakes. Additionally, scholars could investigate the role of institutional actors, such as governments and media organizations, in shaping and reinforcing norms and practices around the production and dissemination of recordings. This line of inquiry could lead to a more comprehensive understanding of the techno-social context in which deepfakes operate and inform policy recommendations for mitigating their potential harms.

Abstract

It is widely thought that deepfake videos are a significant and unprecedented threat to our epistemic practices. In some writing about deepfakes, manipulated videos appear as the harbingers of an unprecedented epistemic apocalypse. In this paper I want to take a critical look at some of the more catastrophic predictions about deepfake videos. I will argue for three claims: (1) that once we recognise the role of social norms in the epistemology of recordings, deepfakes are much less concerning, (2) that the history of photographic manipulation reveals some important precedents, correcting claims about the novelty of deepfakes, and (3) that proposed solutions to deepfakes have been overly focused on technological interventions. My overall goal is not so much to argue that deepfakes are not a problem, but to argue that behind concerns around deepfakes lie a more general class of social problems about the organisation of our epistemic practices.

Deepfakes and the epistemic apocalypse
Posted on

(Featured) The epistemic impossibility of an artificial intelligence take-over of democracy

The epistemic impossibility of an artificial intelligence take-over of democracy

Daniel Innerarity explores the limits of algorithmic governance in relation to democratic decision-making. They argue that algorithms function with a 0/1 logic that is the opposite of ambiguity, and they are unable to handle complex problems that are not well-structured or quantifiable. The authors argue that politics consists of making decisions in the absence of indisputable evidence and that algorithms are of limited utility in such circumstances. Algorithmic rationality reduces the complexity of social phenomena to numbers, whereas political decisions are rarely based on binary categories. The authors suggest that the epistemological principle of uncertainty is central to democratic institutions and that our democratic institutions are a recognition of our ignorance.

The author highlights the limitations of algorithms in decision-making and suggest that they are appropriate only for well-structured and quantifiable problems. In contrast, political decisions are rarely based on binary categories, and politics consists of making decisions in the absence of indisputable evidence. The authors argue that algorithmic rationality reduces the complexity of social phenomena to numbers, which is inappropriate for democratic decision-making. Instead, they suggest that democratic institutions are a recognition of our ignorance and the importance of uncertainty in decision-making.

The author suggests that the epistemological principle of uncertainty is central to democratic institutions. They argue that democracy exists precisely because our knowledge is so limited, and we are so prone to error. Precisely where our knowledge is incomplete, we have greater need for institutions and procedures that favour reflection, debate, criticism, independent advice, reasoned argumentation, and the competition of ideas and visions. Our democratic institutions are not an exhibition of how much we know but a recognition of our ignorance.

The research presented in this paper is significant for broader philosophical issues related to the relationship between knowledge, power, and democratic decision-making. It raises questions about the role of algorithms in decision-making and the limits of rationality in politics. It also highlights the importance of uncertainty, ambiguity, and contingency in democratic decision-making, which has important implications for the legitimacy of democratic institutions.

Future research could explore the implications of these findings for the development of democratic institutions and the role of algorithms in decision-making. It could also explore the role of uncertainty, ambiguity, and contingency in decision-making more broadly and its relationship to different philosophical traditions. Furthermore, it could explore the implications of these findings for the development of more participatory and deliberative forms of democracy that allow for greater reflection, debate, and criticism.

Abstract

Those who claim, whether with fear or with hope, that algorithmic governance can control politics or the whole political process or that artificial intelligence is capable of taking charge of or wrecking democracy, recognize that this is not yet possible with our current technological capabilities but that it could come about in the future if we had better quality data or more powerful computational tools. Those who fear or desire this algorithmic suppression of democracy assume that something similar will be possible someday and that it is only a question of technological progress. If that were the case, no limits would be insurmountable on principle. I want to challenge that conception with a limit that is less normative than epistemological; there are things that artificial intelligence cannot do, because it is unable to do them, not because it should not do them, and this is particularly apparent in politics, which is a peculiar decision-making realm. Machines and people take decisions in a very different fashion. Human beings are particularly gifted at one type of situation and very clumsy in others. The part of politics that is, strictly speaking, political is where this contrast and our greatest aptitude are most apparent. If that is the case, as I believe, then the possibility that democracy will one day be taken over by artificial intelligence is, as a fear or as a desire, manifestly exaggerated. The corresponding counterpart to this is: if the fear that democracy could disappear at the hands of artificial intelligence is not realistic, then we should not expect exorbitant benefits from it either. For epistemic reasons that I will explain, it does not seem likely that artificial intelligence is capable of taking over political logic.

The epistemic impossibility of an artificial intelligence take-over of democracy