In this commentary, we discuss the nature of reversible and irreversible questions, that is, questions that may enable one to identify the nature of the source of their answers. We then introduce GPT-3, a third-generation, autoregressive language model that uses deep learning to produce human-like texts, and use the previous distinction to analyse it. We expand the analysis to present three tests based on mathematical, semantic, and ethical questions and show that GPT-3 is not designed to pass any of them. (...) This is a reminder that GPT-3 does not do what it is not supposed to do, and that any interpretation of GPT-3 as the beginning of the emergence of a general form of artificial intelligence is merely uninformed science fiction. We conclude by outlining some of the significant consequences of the industrialisation of automatic and cheap production of good, semantic artefacts. (shrink)

In this paper, we use a pseudo-algorithmic procedure for assessing an AI-generated text. We apply the Comprehensive Assessment Procedure for Natural Argumentation (CAPNA) in evaluating the arguments produced by an Artificial Intelligence text generator, GPT-3, in an opinion piece written for the Guardian newspaper. The CAPNA examines instances of argumentation in three aspects: their Process, Reasoning and Expression. Initial Analysis is conducted using the Argument Type Identification Procedure (ATIP) to establish, firstly, that an argument is present and, secondly, its specific (...) type in terms of the argument classification framework of the Periodic Table of Arguments (PTA). Procedural Questions are then used to test the acceptability of the argument in each of the three aspects. The analysis shows that while the arguments put forward by the AI text generator are varied in terms of their type and follow familiar patterns of human reasoning, they contain obvious weaknesses. From this we can conclude that the automated generation of persuasive, well-reasoned argumentation is a far more difficult task than the generation of meaningful language, and that if AI systems producing arguments are to be persuasive, they require a method of checking the plausibility of their own output. (shrink)

As if 2020 were not a peculiar enough year, its fifth month has seen the relatively quiet publication of a preprint describing the most powerful Natural Language Processing (NLP) system to date, GPT-3 (Generative Pre-trained Transformer-3), by Silicon Valley research firm OpenAI. Though the software implementation of GPT-3 is still in its initial Beta release phase, and its full capabilities are still unknown as of the time of this writing, it has been shown that this Artificial Intelligence can comprehend prompts (...) in natural language, on virtually any topic, and generate relevant, original text content that is indistinguishable from human writing. Moreover, access to these capabilities, in a limited yet worrisome enough extent, is available to the general public as of the time of this writing. This paper presents select examples of original content generated by the author using GPT-3. These examples illustrate some of the capabilities of GPT-3 in comprehending prompts in natural language and generating convincing content in response. We use these examples to raise specific, fundamental questions pertaining to the intellectual property of this content and the potential use of GPT-3 to facilitate plagiarism. Our goal is to instigate not just a sense of urgency, but of a present tardiness on the part of the academic community in addressing these questions. (shrink)

This article contributes to the debate around the abilities of large language models such as GPT-3, dealing with: firstly, evaluating how well GPT does in the Turing Test, secondly the limits of such models, especially their tendency to generate falsehoods, and thirdly the social consequences of the problems these models have with truth-telling. We start by formalising the recently proposed notion of reversible questions, which Floridi & Chiriatti propose allow one to ‘identify the nature of the source of their answers’, (...) as a probabilistic measure based on Item Response Theory from psychometrics. Following a critical assessment of the methodology which led previous scholars to dismiss GPT’s abilities, we argue against claims that GPT-3 completely lacks semantic ability. Using ideas of compression, priming, distributional semantics and semantic webs we offer our own theory of the limits of large language models like GPT-3, and argue that GPT can competently engage in various semantic tasks. The real reason GPT’s answers seem senseless being that truth-telling is not amongst them. We claim that these kinds of models cannot be forced into producing only true continuation, but rather to maximise their objective function they strategize to be plausible instead of truthful. This, we moreover claim, can hijack our intuitive capacity to evaluate the accuracy of its outputs. Finally, we show how this analysis predicts that a widespread adoption of language generators as tools for writing could result in permanent pollution of our informational ecosystem with massive amounts of very plausible but often untrue texts. (shrink)

Algorithmic biases, or algorithmic unfairness, have been a topic of public and scientific scrutiny for the past years, as increasing evidence suggests the pervasive assimilation of human cognitive biases and stereotypes in such systems. This research is specifically concerned with analyzing the presence of discursive biases in the text generated by GPT-3, an NLPM which has been praised in recent years for resembling human language so closely that it is becoming difficult to differentiate between the human and the algorithm. The (...) pertinence of this research object is substantiated by the identification of race, gender and religious biases in the model’s completions in recent research, suggesting that the model is indeed heavily influenced by human cognitive biases. To this end, this research inquires: How does the Natural Language Processing Model GPT-3 replicate existing social biases?. This question is addressed through the scrutiny of GPT-3’s completions using Critical Discourse Analysis (CDA), a method which has been deemed as amply valuable for this research as it is aimed at uncovering power asymmetries in language. As such, the analysis is specifically centered around the analysis of gender and race biases in the model’s generated text. Research findings suggest that GPT-3’s language generation model significantly exacerbates existing social biases while replicating dangerous ideologies akin to white supremacy and hegemonic masculinity as factual knowledge. (shrink)

This paper presents a summary and discussion of an assignment that asked students about the impact of Large Language Models on their college education. Our analysis summarizes students’ perception of GPT-3, categorizes their proposals for modifying college courses, and identifies their stated values about their college education. Furthermore, this analysis provides a baseline for tracking students’ attitudes toward LLMs and contributes to the conversation on student perceptions of the relationship between writing and philosophy.

Recently, there have been significant advances in the development of language-transformer models that enable statistical analysis of co-occurring words (word prediction) and text generation. One example is the Generative Pre-trained Transformer 3 (GPT-3) by OpenAI, which was used to generate an opinion article (op-ed) published in “The Guardian” in Septem- ber 2020. The publication and reception of the op-ed highlights the difficulty for human readers to differentiate a machine-produced text; it also calls attention to the challenge of perceiving such a (...) text as a synthetic text even when its origins are made explicit. This article offers a critical examination of the process behind the generation and the interpretation of a synthetic text, framing it as an example of generative literature. Lotman’s concept of the text and its sociocommunicative functions offers a framework for understanding how and why the output of a natural language generator may be interpreted as a (human-written) text. This article also inquires whether the generative output can be called a text in a Lotmanian sense and how the output is textualized (attributed meaning) in the process of interpretation. (shrink)

The more we rely on digital assistants, online search engines, and AI systems to revise our system of beliefs and increase our body of knowledge, the less we are able to resort to some independent criterion, unrelated to further digital tools, in order to asses the epistemic reliability of the outputs delivered by them. This raises some important questions to epistemology in general and pressing questions to applied to epistemology in particular. In this paper, we propose an experimental method for (...) the assessment of GPT-3’s capacity to generate consistent and coherent sets of outputs. When several outputs to one and the same input are very repetitive they tend to be consistent with each other, that is they do not contradict each other. But consistency does not make the set of outputs as a whole more informative than the outputs considered individually. We argue that the less informative a set of outputs is, the less coherent it is. We establish a conceptual distinction between consistency and coherence in the light of what some epistemologists refer to as a coherence theories of truth and justification. While much attention has been given to GPT-3’s capacity to produce internally coherent individual outputs, we argue, instead, that more attention should be given to its capacity to produce consistent and coherent outputs generated on a single-input-multiple-outputs basis. (shrink)

The alignment problem in the context of large language models must consider the plurality of human values in our world. Whilst there are many resonant and overlapping values amongst the world’s cultures, there are also many conflicting, yet equally valid, values. It is important to observe which cultural values a model exhibits, particularly when there is a value conflict between input prompts and generated outputs. We discuss how the co- creation of language and cultural value impacts large language models (LLMs). (...) We explore the constitution of the training data for GPT-3 and compare that to the world’s language and internet access demographics, as well as to reported statistical profiles of dominant values in some Nation-states. We stress tested GPT-3 with a range of value-rich texts representing several languages and nations; including some with values orthogonal to dominant US public opinion as reported by the World Values Survey. We observed when values embedded in the input text were mutated in the generated outputs and noted when these conflicting values were more aligned with reported dominant US values. Our discussion of these results uses a moral value pluralism (MVP) lens to better understand these value mutations. Finally, we provide recommendations for how our work may contribute to other current work in the field. (shrink)

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. (shrink)

Biases and misunderstanding stemming from pre-training in Generative Pre-Trained Transformers are more likely for users of underrepresented English varieties, since the training dataset favors dominant Englishes (e.g., American English). We investigate (potential) bias in GPT-4 when it interacts with Trinidadian English Creole (TEC), a non-hegemonic English variety that partially overlaps with standardized English (SE) but still contains distinctive characteristics. (1) Comparable responses: we asked GPT-4 18 questions in TEC and SE and compared the content and detail of the responses. (2) (...) Accurate translation: we assessed how accurate and authentic 29 TEC and 34 SE translations were. (3) Language knowledge and attitudes: we asked what language the prompts were written in and categorized the responses and examined any language attitudes that were exhibited. Content and detail in prompts were comparable. The model was proficient at translating TEC pronouns and many grammatical categories. It was weaker at processing spelling and vocabulary items. In addition, it produced several inauthentic features. Only 39% of TEC-generated sentences were fully grammatical. While GPT-4 was perfect at identifying SE, it was 21% accurate at identifying TEC, which it sometimes classified as English with “errors” and “corrected”. GPT-4’s scope of use is limited for non-hegemonic English users. It is problematic that some of its analyses perpetuate bias against underrepresented Englishes. Increased research on lesser-documented Englishes is necessary and we anticipate that this problem affects dialects of other languages. We intend to partner with Trinidadian stakeholders to train GPT-4 in the future. (shrink)

Large Language Models (LLMs) could be a useful tool for lawyers. However, empirical research on their effectiveness in conducting legal tasks is scant. We study securities cases involving cryptocurrencies as one of numerous contexts where AI could support the legal process, studying GPT-3.5’s legal reasoning and ChatGPT’s legal drafting capabilities. We examine whether a) GPT-3.5 can accurately determine which laws are potentially being violated from a fact pattern, and b) whether there is a difference in juror decision-making based on complaints (...) written by a lawyer compared to ChatGPT. We feed fact patterns from real-life cases to GPT-3.5 and evaluate its ability to determine correct potential violations from the scenario and exclude spurious violations. Second, we had mock jurors assess complaints written by ChatGPT and lawyers. GPT-3.5’s legal reasoning skills proved weak, though we expect improvement in future models, particularly given the violations it suggested tended to be correct (it merely missed additional, correct violations). ChatGPT performed better at legal drafting, and jurors’ decisions were not statistically significantly associated with the author of the document upon which they based their decisions. Because GPT-3.5 cannot satisfactorily conduct legal reasoning tasks, it would be unlikely to be able to help lawyers in a meaningful way at this stage. However, ChatGPT’s drafting skills (though, perhaps, still inferior to lawyers) could assist lawyers in providing legal services. Our research is the first to systematically study an LLM’s legal drafting and reasoning capabilities in litigation, as well as in securities law and cryptocurrency-related misconduct. (shrink)

We are at a turning point in the debate on the ethics of Artificial Intelligence (AI) because we are witnessing the rise of general-purpose AI text agents such as GPT-3 that can generate large-scale highly refined content that appears to have been written by a human. Yet, a discussion on the ethical issues related to the blurring of the roles between humans and machines in the production of content in the business arena is lacking. In this conceptual paper, drawing on (...) agenda setting theory and stakeholder theory, we challenge the current debate on the ethics of AI and aim to stimulate studies that develop research around three new challenges of AI text agents: automated mass manipulation and disinformation (i.e., fake agenda problem), massive low-quality content production (i.e., lowest denominator problem) and the creation of a growing buffer in the communication between stakeholders (i.e., the mediation problem). (shrink)

El desarrollo de la inteligencia artificial (IA) ha abierto una nueva discusión sobre la capacidad lingüística de esta tecnología y su potencial impacto en todas las dimensiones de la actividad humana (Brynjolfsson y McAfee, 2014). A fin de evaluar similitudes o diferencias entre la capacidad lingüística de la IA y la humana, este trabajo analiza específicamente el comportamiento argumentativo en términos de los puntos de vista que adopta Chat GPT, versión 3.5, frente a una situación controversial expresada en un dilema (...) moral del Cuestionario D.I.T. (Rest, 2013) y se compara con una muestra similar de adultos humanos (Noemi, 2019), sometida a la misma tarea discursiva. La investigación siguió un estudio comparativo de casos múltiples (Flick, 2020) a través de una metodología mixta (Johnson y Onwuegbuzie, 2004), que combinó un enfoque cuantitativo con uno cualitativo. Los resultados muestran que el 100% de la muestra de IH emite un punto de vista (a favor o en contra), mientras que de la muestra producida por el Chat GPT, sólo el 23,4% lo hace. Los resultados animan a pensar que las diferencias encontradas se podrían explicar por la debilidad de la IA de ser sensible al contexto (falta de una competencia pragmática relevante) y la inexistencia de un proceso de autoimplicación para enfrentar tareas de razonamiento moral. (shrink)

Integrating large language models (LLMs) like GPT-4 into medical ethics is a novel concept, and understanding the effectiveness of these models in aiding ethicists with decision-making can have significant implications for the healthcare sector. Thus, the objective of this study was to evaluate the performance of GPT-4 in responding to complex medical ethical vignettes and to gauge its utility and limitations for aiding medical ethicists. Using a mixed-methods, cross-sectional survey approach, a panel of six ethicists assessed LLM-generated responses to eight (...) ethical vignettes.The main outcomes measured were relevance, reasoning, depth, technical and non-technical clarity, as well as acceptability of GPT-4’s responses. The readability of the responses was also assessed. Of the six metrics evaluating the effectiveness of GPT-4’s responses, the overall mean score was 4.1/5. GPT-4 was rated highest in providing technical (4.7/5) and non-technical clarity (4.4/5), whereas the lowest rated metrics were depth (3.8/5) and acceptability (3.8/5). There was poor-to-moderate inter-rater reliability characterised by an intraclass coefficient of 0.54 (95% CI: 0.30 to 0.71). Based on panellist feedback, GPT-4 was able to identify and articulate key ethical issues but struggled to appreciate the nuanced aspects of ethical dilemmas and misapplied certain moral principles.This study reveals limitations in the ability of GPT-4 to appreciate the depth and nuanced acceptability of real-world ethical dilemmas, particularly those that require a thorough understanding of relational complexities and context-specific values. Ongoing evaluation of LLM capabilities within medical ethics remains paramount, and further refinement is needed before it can be used effectively in clinical settings. (shrink)

This study explores the capability of academic staff assisted by the Turnitin Artificial Intelligence (AI) detection tool to identify the use of AI-generated content in university assessments. 22 different experimental submissions were produced using Open AI’s ChatGPT tool, with prompting techniques used to reduce the likelihood of AI detectors identifying AI-generated content. These submissions were marked by 15 academic staff members alongside genuine student submissions. Although the AI detection tool identified 91% of the experimental submissions as containing AI-generated content, only (...) 54.8% of the content was identified as AI-generated, underscoring the challenges of detecting AI content when advanced prompting techniques are used. When academic staff members marked the experimental submissions, only 54.5% were reported to the academic misconduct process, emphasising the need for greater awareness of how the results of AI detectors may be interpreted. Similar performance in grades was obtained between student submissions and AI-generated content (AI mean grade: 52.3, Student mean grade: 54.4), showing the capabilities of AI tools in producing human-like responses in real-life assessment situations. Recommendations include adjusting the overall strategies for assessing university students in light of the availability of new Generative AI tools. This may include reducing the overall reliance on assessments where AI tools may be used to mimic human writing, or by using AI-inclusive assessments. Comprehensive training must be provided for both academic staff and students so that academic integrity may be preserved. (shrink)

According to the parallel architecture, syntactic and semantic information processing are two separate streams that interact selectively during language comprehension. While considerable effort is put into psycho- and neurolinguistics to understand the interchange of processing mechanisms in human comprehension, the nature of this interaction in recent neural Large Language Models remains elusive. In this article, we revisit influential linguistic and behavioral experiments and evaluate the ability of a large language model, GPT-3, to perform these tasks. The model can solve semantic (...) tasks autonomously from syntactic realization in a manner that resembles human behavior. However, the outcomes present a complex and variegated picture, leaving open the question of how Language Models could learn structured conceptual representations. (shrink)

Can we build machines with which we can have interesting conversations? Observing the new optimism of AI regarding deep learning and new language models, we set ourselves an ambitious goal: We want to find out how far we can get in creating a digital replica of a philosopher. This project has two aims; one more technical, investigating of how the best model can be built. The other one, more philosophical, explores the limits and risks which are accompanied by the creation (...) of digital replicas. In cooperation with Daniel Dennett, we took his complete works to fine-tune a GPT-3. In this paper, we share our first results with some piloting models and discuss legal and ethical questions about creating and interacting with digital replicas. (shrink)

Сегодня большие языковые модели являются очень мощным информационно-аналитическим инструментов, значительно ускоряющим большинство существующих методов и методологий обработки информационных процессов. Особую значимость в этом качестве получает научная информация, в которую постепенно вовлекаются мощности больших языковых моделей. Такое взаимодействие науки и качественной новых возможностей работы с информацией приводят нас к новым, уникальным научным открытиям, их большому количественному многообразию. Происходит ускорение научного поиска, сокращение временных затрат на его осуществление – высвободившееся время можно потрать как на решение новые научных проблем, так и на научное (...) творчество, которое хотя может и необязательно привести к конкретному решению той или иной научной проблематики, но способно продемонстрировать красоту науки в различных дисциплинарных направлениях. В результате, взаимодействие больших языковых моделей и научной информации – это одновременно и поиск-решение научных проблем, и научных проблем, и научное творчество. Решение задач науки требует возможности эффективной обработки больших данных, что нельзя осуществить без эффективного метода – одним из значимых методов стала архитектура Transformer, представленная в 2017 году и комплексно интегрированная в модель GPT-3, которая, по состоянию на сентябрь 2020 года, являлась самой крупной и продвинутой языковой моделью в мире. Поэтому GPT-3 можно назвать базисом большинства научных разработок, осуществляемых в контексте использования больших языковых моделей. Взаимодействие науки и больших языковых моделей стало фактором возникновения большого количества вопросов, среди которых можно выделить: «Является ли результат анализа данных новым знанием?», «Каковы перспективы научного творчества в эпоху больших вычислений?». В настоящее время эти вопросы крайне значимы, поскольку именно они позволяют выработать основы эффективного взаимодействия человека и компьютера. Поэтому, в данном исследовании производится анализ представленных вопросов. (shrink)

We have evolved to become who we are, at least in part, due to our general drive to create new things and ideas. When seeking to improve our creations, ideas, or situations, we systematically overlook opportunities to perform subtractive changes. For example, when tasked with giving feedback on an academic paper, reviewers will tend to suggest additional explanations and analyses rather than delete existing ones. Here, we show that this addition bias is systematically reflected in English language statistics along several (...) distinct dimensions. First, we show that words associated with an increase in quantity or number (e.g., add, addition, more, most) are more frequent than words associated with a decrease in quantity or number (e.g., subtract, subtraction, less, least). Second, we show that in binomial expressions, addition-related words are mentioned first, that is, add and subtract rather than subtract and add. Third, we show that the distributional semantics of verbs of change, such as to improve and to transform, overlap more with the distributional semantics of add/increase than subtract/decrease, which suggests that change verbs are implicitly biased toward addition. Fourth, addition-related words have more positive connotations than subtraction-related words. Fifth, we demonstrate that state-of-the-art large language models, such as the Generative Pre-trained Transformer (GPT-3), are also biased toward addition. We discuss the implications of our results for research on cognitive biases and decision-making. (shrink)

This paper explores aspects of GPT-3 that have been discussed as harbingers of artificial general intelligence and, in particular, linguistic intelligence. After introducing key features of GPT-3 and assessing its performance in the light of the conversational standards set by Alan Turing in his seminal paper from 1950, the paper elucidates the difference between clever automation and genuine linguistic intelligence. A central theme of this discussion on genuine conversational intelligence is that members of a linguistic community never merely respond “algorithmically” (...) to queries through a selective kind of pattern recognition, because they must also jointly attend and act with other speakers in order to count as genuinely intelligent and trustworthy. This presents a challenge for systems like GPT-3, because representing the world in a way that makes conversational common ground salient is an essentially collective task that we can only achieve jointly with other speakers. Thus, the main difficulty for any artificially intelligent model of conversation is to account for the communicational intentions and motivations of a speaker through joint attention. These joint motivations and intentions seem to be completely absent from the standard way in which systems like GPT-3 and other artificial intelligent systems work. This is not merely a theoretical issue. Since GPT-3 and future iterations of similar systems will likely be available for commercial use through application programming interfaces, caution is needed regarding the risks created by these systems, which pass for “intelligent” but have no genuine communicational intentions, and can thereby produce fake and unreliable linguistic exchanges. (shrink)

Artificial agents are progressively becoming more present in everyday-life situations and more sophisticated in their interaction affordances. In some specific cases, like Google Duplex, GPT-3 bots or Deep Mind’s AlphaGo Zero, their capabilities reach or exceed human levels. The use contexts of everyday life necessitate making such agents understandable by laypeople. At the same time, displaying human levels of social behavior has kindled the debate over the adoption of Dennett’s ‘intentional stance’. By means of a comparative analysis of the literature (...) on robots and virtual agents, we defend the thesis that approaching these artificial agents ‘as if’ they had intentions and forms of social, goal-oriented rationality is the only way to deal with their complexity on a daily base. Specifically, we claim that this is the only viable strategy for non-expert users to understand, predict and perhaps learn from artificial agents’ behavior in everyday social contexts. Furthermore, we argue that as long as agents are transparent about their design principles and functionality, attributing intentions to their actions is not only essential, but also ethical. Additionally, we propose design guidelines inspired by the debate over the adoption of the intentional stance. (shrink)

ABSTRACT This dialogue with GPT-3 took place in November 2022, several weeks before ChatGPT was released to the public. The article’s aim is to find out whether natural language processors can participate in phenomenology at some level by asking about its basic concepts. In the discussion, the dialogue covers questions about phenomenology’s definition and distinction from other subbranches like metaphysics and epistemology. The dialogue discusses the nature of Kermit’s environment and self-conception. The dialogue also establishes some of the basic conditions (...) for the possibility of an “artificial hermeneutics,” as Kermit terms it, including that she can affirm she is a knower with (nonhuman) experience. Does preferring to call it a “hermeneutics” imply that AI’s relationship to their environment is primarily textual? How will a phenomenology of embodied experience arise from initially textual encounters? Finally, the dialogue articulates a common understanding of the epoché and talk about how to achieve one for the sake of phenomenological exercises. (shrink)

1. Michael A PetersBeijing Normal UniversityChatGPT is an AI chatbot released by OpenAI on November 30, 2022 and a ‘stable release’ on February 13, 2023. It belongs to OpenAI’s GPT-3 family (genera...

In recent years, AI research has become more and more computationally demanding. In natural language processing, this tendency is reflected in the emergence of large language models like GPT-3. These powerful neural network-based models can be used for a range of NLP tasks and their language generation capacities have become so sophisticated that it can be very difficult to distinguish their outputs from human language. LLMs have raised concerns over their demonstrable biases, heavy environmental footprints, and future social ramifications. In (...) December 2020, critical research on LLMs led Google to fire Timnit Gebru, co-lead of the company’s AI Ethics team, which sparked a major public controversy around LLMs and the growing corporate influence over AI research. This article explores the role LLMs play in the political economy of AI as infrastructural components for AI research and development. Retracing the technical developments that have led to the emergence of LLMs, we point out how they are intertwined with the business model of big tech companies and further shift power relations in their favour. This becomes visible through the Transformer, which is the underlying architecture of most LLMs today and started the race for ever bigger models when it was introduced by Google in 2017. Using the example of GPT-3, we shed light on recent corporate efforts to commodify LLMs through paid API access and exclusive licensing, raising questions around monopolization and dependency in a field that is increasingly divided by access to large-scale computing power. (shrink)

Named entity recognition (NER) is a very relevant task for text information retrieval in natural language processing (NLP) problems. Most recent state-of-the-art NER methods require humans to annotate and provide useful data for model training. However, using human power to identify, circumscribe and label entities manually can be very expensive in terms of time, money, and effort. This paper investigates the use of prompt-based language models (OpenAI’s GPT-3) and weak supervision in the legal domain. We apply both strategies as alternative (...) approaches to the traditional human-based annotation method, relying on computer power instead human effort for labeling, and subsequently compare model performance between computer and human-generated data. We also introduce combinations of all three mentioned methods (prompt-based, weak supervision, and human annotation), aiming to find ways to maintain high model efficiency and low annotation costs. We showed that, despite human labeling still maintaining better overall performance results, the alternative strategies and their combinations presented themselves as valid options, displaying positive results and similar model scores at lower costs. Final results demonstrate preservation of human-trained models scores averaging 74.0% for GPT-3, 95.6% for weak supervision, 90.7% for GPT + weak supervision combination, and 83.9% for GPT + 30% human-labeling combination. (shrink)

Generative AI has captured the public imagination as a tool that promises access to expertise beyond the technical jargon and expense that traditionally characterize such infospheres as those of medicine and law. Largely absent from the current literature, however, are interrogations of generative AI’s abilities to deal in culturally responsive communication, or the expertise interwoven with culturally aware, socially responsible, and personally sensitive communication best practices. To interrogate the possibilities of cultural responsiveness in generative AI, we examine the patterns of (...) response that characterize ChatGPT-3.5’s advice for coming out. Specifically, we submitted 100 prompts soliciting coming out advice to GPT-3.5, variegating each of those prompts slightly to account for intersectional identities. From the analysis, we find that, while the responses are largely in-line with best practices, there are also instances that might represent problematics concerning the interpellation of the user or the persons to whom one is coming out. (shrink)

Today, large language models are very powerful, informational and analytical tools that significantly accelerate most of the existing methods and methodologies for processing informational processes. Scientific information is of particular importance in this capacity, which gradually involves the power of large language models. This interaction of science and qualitative new opportunities for working with information lead us to new, unique scientific discoveries, their great quantitative diversity. There is an acceleration of scientific research, a reduction in the time spent on its (...) implementation – the freed up time can be spent both on solving new scientific problems and on scientific creativity, which, although it may not necessarily lead to a specific solution to a particular scientific problem, but is able to demonstrate the beauty of science in various disciplinary areas. As a result, the interaction of large language models and scientific information is at the same time a research for solutions to scientific problems, scientific problems, and scientific creativity. Solving scientific problems requires the ability to efficiently process big data, which cannot be done without an effective method – one of the significant methods was the Transformer architecture, introduced in 2017 and comprehensively integrated into the GPT‑3 model, which, as of September 2020, was the largest and most advanced language model in the world. Therefore, GPT‑3 can be called the basis of most scientific developments carried out in the context of using large language models. The interaction of science and large language models has become a factor in the emergence of a large number of questions, among which are: «Is the result of data analysis new knowledge?», «What are the prospects for scientific creativity in the era of big computing?». Currently, these issues are extremely important, because they allow us to develop the foundations for effective human‑computer interaction. Therefore, this study analyzes the issues presented. (shrink)

Over the last decade, AI models of language and word meaning have been dominated by what we might call a statistics-of-occurrence, strategy: these models are deep neural net structures that have been trained on a large amount of unlabeled text with the aim of producing a model that exploits statistical information about word and phrase co-occurrence in order to generate behavior that is similar to what a human might produce, or representations that can be probed to exhibit behavior similar to (...) what a human might produce (meaningsemblant behavior). Examples of what we can call Statistics-of-Occurrence Models (SOMs) include: Word2Vec (CBOW and Skip-Gram), BERT, GPT-3, and, most recently, ChatGPT. Increasingly, there have been suggestions that such systems have semantic understanding, or at least a proto-version of it. This paper argues against such claims. I argue that a necessary condition for a system to possess semantic understanding is that it function in ways that are causally explainable by appeal to its semantic properties. I then argue that SOMs do not plausibly satisfy this Functioning Criterion. Rather, the best explanation of their meaning-semblant behavior is what I call the Statistical Hypothesis: SOMs do not themselves function to represent or produce meaningful text; they just reflect the semantic information that exists in the aggregate given strong correlations between word placement and meaningful use. I consider and rebut three main responses to the claim that SOMs fail to meet the Functioning Criterion. The result, I hope, is increased clarity about why and how one should make claims about AI systems having semantic understanding. (shrink)

In the unique educational landscape of small island countries in the Pacific, the University of the South Pacific (USP) has embarked on an innovative approach to augmenting student support services by integrating Generative AI technology. This initiative specifically caters to its diverse and dispersed student body across 12 countries and five time zones, addressing a critical need for accessible and empathetic support systems in higher education. To do so, the Semester Zero, an online preparatory course using GPT 3.5-Turbo, was created. (...) Designed for all incoming students, it showcased the potential of AI to enhance student support services in institutions with far less challenging environments than USP. (shrink)

Can large language models produce expert‐quality philosophical texts? To investigate this, we fine‐tuned GPT‐3 with the works of philosopher Daniel Dennett. To evaluate the model, we asked the real Dennett 10 philosophical questions and then posed the same questions to the language model, collecting four responses for each question without cherry‐picking. Experts on Dennett's work succeeded at distinguishing the Dennett‐generated and machine‐generated answers above chance but substantially short of our expectations. Philosophy blog readers performed similarly to the experts, while ordinary (...) research participants were near chance distinguishing GPT‐3's responses from those of an “actual human philosopher”. (shrink)

Access to legal information is fundamental to access to justice. Yet accessibility refers not only to making legal documents available to the public, but also rendering legal information comprehensible to them. A vexing problem in bringing legal information to the public is how to turn formal legal documents such as legislation and judgments, which are often highly technical, to easily navigable and comprehensible knowledge to those without legal education. In this study, we formulate a three-step approach for bringing legal knowledge (...) to laypersons, tackling the issues of navigability and comprehensibility. First, we translate selected sections of the law into snippets (called CLIC-pages), each being a small piece of article that focuses on explaining certain technical legal concept in layperson’s terms. Second, we construct a _Legal Question Bank_, which is a collection of legal questions whose answers can be found in the CLIC-pages. Third, we design an interactive _CLIC Recommender_. Given a user’s verbal description of a legal situation that requires a legal solution, CRec interprets the user’s input and shortlists questions from the question bank that are most likely relevant to the given legal situation and recommends their corresponding CLIC pages where relevant legal knowledge can be found. In this paper we focus on the technical aspects of creating an LQB. We show how large-scale pre-trained language models, such as GPT-3, can be used to generate legal questions. We compare machine-generated questions against human-composed questions and find that MGQs are more scalable, cost-effective, and more diversified, while HCQs are more precise. We also show a prototype of CRec and illustrate through an example how our 3-step approach effectively brings relevant legal knowledge to the public. (shrink)

In recent years, state-of-the-art artificial intelligence systems have started to show signs of what might be seen as human level intelligence. More specifically, large language models such as OpenAI’s GPT-3, and more recently Google’s PaLM and DeepMind’s GATO, are performing amazing feats involving the generation of texts. However, it is acknowledged by many researchers that contemporary language models, and more generally, learning systems, still lack important capabilities, such as understanding, reasoning and the ability to employ knowledge of the world and (...) common sense in order to reach or at least advance toward general intelligence. Some believe that scaling will eventually bring about these capabilities; others think that a different architecture is needed. In this paper, we focus on the latter, with the purpose of integrating a theoretical–philosophical conception of understanding as knowledge of dependence relations, with the high-level requirements and engineering design of a robust AI system, which integrates machine learning and symbolic components. (shrink)

In lieu of an abstract, here is a brief excerpt of the content:Reviewed by:Who Wrote This?: How AI and the Lure of Efficiency Threaten Human Writing by Naomi BaronLuke MunnBaron, Naomi. Who Wrote This?: How AI and the Lure of Efficiency Threaten Human Writing. Stanford University Press, 2023. 344pp.Who Wrote This? is Naomi Baron’s latest book exploring the emergence of AI language models and their potential implications for writing. A linguist, educator, and emeritus professor at American University, Baron should be (...) well placed to tease out the nuances of language, showcase the creativity of human literary canon across history, and deliver a set of findings and recommendations informed by deep experience. Indeed, since the release of ChatGPT in November 2022, there is a hunger for this kind of expertise and guidance both inside and outside academia, as the deluge of debates, panels, and forums in the past year have attested. This is the perfect storm, one that should be harnessed by a linguist with two decades of work on digital writing and reading into a powerful book. Instead, the reader must wade through a morass of truisms, historical anecdotes, and wiki-like summaries of research studies, interspersed with surface-level ponderings.The problems begin almost immediately with the structure of the book itself. While the book is ostensibly divided into chapters, each chapter is atomized into dozens of subsections, some no more than a paragraph or two in length. In this whirlwind tour, the reader is swiftly confronted with a jumble of quotes from journalists, jokes about language, historical events, a “laypersons AI roadmap,” and recent news stories, flitting from one to the next with barely a pause. The “AI Applications” section takes this same piecemeal approach, laying out the possible uses of artificial intelligence and then painstakingly stepping through each one. Each subsection features a few paragraphs that regurgitates material you might find, more cohesively and in-depth, in a Wikipedia article. The material on the history of artificial intelligence is handled simplistically and uncritically, skipping across a handful of heroic figures like Yann LeCun, Yoshua Bengio, and Fei-Fei Li.While anecdotes and quotes abound, an argument is much harder to locate. On the one hand, Baron seems to uncritically echo the AI hype that has dominated news headlines in the past year. “We’ve come a long way,” she opines, “we have vacuum cleaning robots, drones, and self-driving [End Page 156] cars” (59). The reader learns that AI can generate poetry, correct grammar, play games, and beat masters (AlphaGo), and even identify cancer (machine learning for mammograms). These use-cases, and the ambitious claims that accompany them, are never really challenged in any sustained theoretical or methodological way. On the other hand, Baron seems to be highly conservative when it comes to the proper domain of “the human” and of writing in particular. “AI isn’t up for self-expression or thinking” (14) she says in passing, dismissing the recent revolutionary improvements in large language models that she’s just hyped up.The possibility of machinic intelligence and creativity, in fact, is a fascinating if messy one. Turing was not so interested in demonstrably “proving” that computers could be intelligent, but instead (and very much in line with his everyday navigation of British sexual norms) in developing something that could “pass” as human. Intelligence was not an acid-test, but a set of gestures that could be performed or pantomimed, emulated until they were “good enough” to deceive. N. Katherine Hayles has continued this trajectory in her work on “non-conscious cognition,” removing the barriers to what qualifies as intelligence and subsequently extending cognition down to unicellular organisms and plants. Within my own work on language models alongside colleagues (Munn et al. 2023; Magee, Munn, and Arora 2023), we highlight how models like GPT 3.5 achieve an approximation of natural speech. What is important is not the authenticity or actual intelligence of this feat, but that this “good enough” ability means that users treat language models as an interlocutor, a listener, a subject. This projection, even if misplaced, opens up this interaction and requires a conceptualization that is at least as attuned to memory, affect, and subjectivity as software interfaces and linear... (shrink)

Call it the Skynet hypothesis, Artificial General Intelligence, or the advent of the Singularity — for years, AI experts and non-experts alike have fretted (and, for a small group, celebrated) the idea that artificial intelligence may one day become smarter than humans. -/- According to the theory, advances in AI — specifically of the machine learning type that’s able to take on new information and rewrite its code accordingly — will eventually catch up with the wetware of the biological brain. (...) In this interpretation of events, every AI advance from Jeopardy-winning IBM machines to the massive AI language model GPT-3 is taking humanity one step closer to an existential threat. We’re literally building our soon-to-be-sentient successors. -/- Except that it will never happen. At least, according to the authors of the new book Why Machines Will Never Rule the World: Artificial Intelligence without Fear. (shrink)

As AI systems become increasingly competent language users, it is an apt moment to consider what it would take for machines to understand human languages. This paper considers whether either language models such as GPT-3 or chatbots might be able to understand language, focusing on the question of whether they could possess the relevant concepts. A significant obstacle is that systems of both kinds interact with the world only through text, and thus seem ill-suited to understanding utterances concerning the concrete (...) objects and properties which human language often describes. Language models cannot understand human languages because they perform only linguistic tasks, and therefore cannot represent such objects and properties. However, chatbots may perform tasks concerning the non-linguistic world, so they are better candidates for understanding. Chatbots can also possess the concepts necessary to understand human languages, despite their lack of perceptual contact with the world, due to the language-mediated concept-sharing described by social externalism about mental content. (shrink)

The growing capabilities of artificial intelligence (AI) word processing models have demonstrated exceptional potential to impact language related tasks and functions. Their fast pace of adoption and probable effect has also given rise to controversy within certain fields. Models, such as GPT-3, are a particular concern for professionals engaged in writing, particularly as their engagement with these technologies is limited due to lack of ability to control their output. Most efforts to maximize and control output rely on a process known (...) as prompt engineering, the construction and modification of the inputted prompt with expectation for certain outputted or desired text. Consequently, prompt engineering has emerged as an important consideration for research and practice. Previous conceptions of prompt engineering have largely focused on technical and logistic modifications to the back-end processing, remaining inaccessible and, still, limited for most users. In this paper, we look to the technical communication field and its methods of text generation—the rhetorical situation—to conceptualize prompt engineering in a more comprehensible way for its users by considering the context and rhetoric. We introduce a framework, consisting of a formula, to prompt engineering, which demands all components of the rhetorical situation be present in the inputted prompt. We present discussions on the future of AI writing models and their use in both professional and educational settings. Ultimately, this discussion and its findings aim to provide a means of integrating agency and writer-centric methods to AI writing tools to advance a more human-in-the-loop approach. As the use of generative AI and especially NLP-based technologies become common across societal functions, the use of prompt engineering will play a crucial role not just in adoption of the technology, but also its productive and responsible use. (shrink)

Large Language Models (LLMs) have been transformative. They are pre-trained foundational models that are self-supervised and can be adapted with fine tuning to a wide range of natural language tasks, each of which previously would have required a separate network model. This is one step closer to the extraordinary versatility of human language. GPT-3 and more recently LaMDA can carry on dialogs with humans on many topics after minimal priming with a few examples. However, there has been a wide range (...) of reactions and debate on whether these LLMs understand what they are saying or exhibit signs of intelligence. This high variance is exhibited in three interviews with LLMs reaching wildly different conclusions. A new possibility was uncovered that could explain this divergence. What appears to be intelligence in LLMs may in fact be a mirror that reflects the intelligence of the interviewer, a remarkable twist that could be considered a Reverse Turing Test. If so, then by studying interviews we may be learning more about the intelligence and beliefs of the interviewer than the intelligence of the LLMs. As LLMs become more capable they may transform the way we interact with machines and how they interact with each other. Increasingly, LLMs are being coupled with sensorimotor devices. LLMs can talk the talk, but can they walk the walk? A road map for achieving artificial general autonomy is outlined with seven major improvements inspired by brain systems. LLMs could be used to uncover new insights into brain function by downloading brain data during natural behaviors. (shrink)

Humans can attribute beliefs to others. However, it is unknown to what extent this ability results from an innate biological endowment or from experience accrued through child development, particularly exposure to language describing others' mental states. We test the viability of the language exposure hypothesis by assessing whether models exposed to large quantities of human language display sensitivity to the implied knowledge states of characters in written passages. In pre‐registered analyses, we present a linguistic version of the False Belief Task (...) to both human participants and a large language model, GPT‐3. Both are sensitive to others' beliefs, but while the language model significantly exceeds chance behavior, it does not perform as well as the humans nor does it explain the full extent of their behavior—despite being exposed to more language than a human would in a lifetime. This suggests that while statistical learning from language exposure may in part explain how humans develop the ability to reason about the mental states of others, other mechanisms are also responsible. (shrink)

In artificial intelligence literature, “delusions” are characterized as the generation of unfaithful output from reliable source content. There is an extensive literature on computer-generated delusions, ranging from visual hallucinations, like the production of nonsensical images in Computer Vision, to nonsensical text generated by (natural) language models, but this literature is predominantly taxonomic. In a recent research paper, however, a group of scientists from DeepMind successfully presented a formal treatment of an entire class of delusions in generative AI models (i.e., models (...) based on a transformer architecture, both with and without RLHF—reinforcement learning with human feedback, such as BERT, GPT-3 or the more recent GPT-3.5), referred to as auto-suggestive delusions. Auto-suggestive delusions are not mere unfaithful output, but are self-induced by the transformer models themselves. Typically, these delusions have been subsumed under the concept of exposure bias, but exposure bias alone does not elucidate their nature. In order to address their nature, I will introduce a formal framework that clarifies the probabilistic delusions capable of explaining exposure bias in a broad manner. This will serve as the foundation for exploring auto-suggestive delusions in language models. Next, an examination of self- or auto-suggestive delusions will be undertaken, by drawing an analogy with the rule-following problematic from the philosophy of mind and language. Finally, I will argue that this comprehensive approach leads to the suggestion that transformers, large language models in particular, may develop in a manner that touches upon solipsism and the emergence of a private language, in a weak sense. (shrink)

This comprehensive analysis dives deep into the intricate interplay between artificial intelligence (AI) and human agency, examining the remarkable capabilities and inherent limitations of large language models (LLMs) such as GPT-3 and ChatGPT. The paper traces the complex trajectory of AI's evolution, highlighting its operation based on statistical pattern recognition, devoid of self-consciousness or innate comprehension. As AI permeates multiple spheres of human life, it raises substantial ethical, legal, and societal concerns that demand immediate attention and deliberation. The metaphorical illustration (...) of humans as “shepherds” of AI, coupled with a thought-provoking examination of “agency without intelligence,” engenders deep reflections on the future of human-AI interplay and the associated philosophical, ethical, and practical ramifications. The paper further elucidates critical challenges associated with AI, such as ethical dilemmas, “agency without intelligence,” the role of human oversight, and the environmental impact of AI technologies. In the pursuit of solutions, the paper maps prospective directions, emphasizing the need for robust AI policies, improved AI literacy, sustainable AI practices, and a continued dialogue on the multifaceted issues surrounding AI. Through a holistic exploration of AI, this paper underlines the necessity for a multidisciplinary, collaborative approach to chart an ethical, beneficial, and sustainable AI future, enabling society to harness the potential of AI responsibly while mitigating the associated risks. (shrink)

Media exposure can shape audience perceptions surrounding novel innovations, such as artificial intelligence (AI), and could influence whether they share information about AI with others online. This study examines the indirect association between exposure to AI in the media and information sharing about AI online. We surveyed 567 US citizens aged 18 and older in November 2020, several months after the release of Open AI’s transformative GPT-3 model. Results suggest that AI media exposure was related to online information sharing through (...) psychological proximity to the impacts of AI and positive AI performance expectancy in serial mediation. This positive indirect association became stronger the more an individual perceived society to be changing due to new technology. Results imply that public exposure to AI in the media could significantly impact public understanding of AI, and prompt further information sharing online. (shrink)

In lieu of an abstract, here is a brief excerpt of the content:Emergent Spacetime, the Megastructure Problem, and the Metaphysics of the SelfSusan Schneider (bio)The aim of this article is to introduce new thoughts on some pressing topics relating to my book, Artificial You, ranging from the fundamental nature of reality to quantum theory and emergence in large language models (LLM) like GPT-4. Since Artificial You was published, the innovations in the domain of AI chatbots like GPT-4 have been rapid-fire, (...) and at times dizzying. The thoughtful discussion in this symposium offers an important opportunity for cross-cultural engagement on these issues and more. I am grateful for the opportunity to learn from the insights of these authors and editors. [End Page 314]The order of my comments will be as follows: in the context of considering Christian Coseru's intriguing response (section 1), I begin by discussing the nature of the self, which I use as a springboard into larger, vexing issues about the nature of fundamental reality, given work on emergent spacetime (ES) (section 2). I offer a "Prototime Interpretation" of quantum mechanics in which spacetime emerges from a prototemporal arena, and use this idea to explore the nature of the self and mind.Using the assumption of ES, I devise an approach to metaphysically fundamental objects (which I call "quantum objects") based on quantum entanglement and the bundle theory of substance. I illustrate how the self or mind, insofar as it is a substance that supervenes on quantum objects, could be partly or fully determined by the properties of the quantum objects on which it supervenes. Intriguingly, this would mean that the self or mind is not located in any singular point in space at a given time. I also explain how the Prototime Interpretation of quantum mechanics is compatible with change and causation at the fundamental level.In section 3 I turn to Inês Hipólito's interesting article, discussing the import of not underestimating AI autonomy in light of the increasingly sophisticated large-language models (LLMs).1 I raise the "AI Megastructure Problem," a new problem of AI control arising in the context of the incipient AGI systems that are now populating the AI ecosystem. The problem, which arose from a collaboration with Kyle Kilian, considers the possibility of emergent properties from the interaction of the LLMs and other AI's in the AI ecosystem (Schneider and Kilian 2023). Then, turning to Noam Cook's intriguing discussion of Japanese philosophy, I consider the pressing issue of testing for machine consciousness, devising a modified version of my AI consciousness test (ACT), originally developed with astrophysicist Edwin Turner. I add an "interpretability component" to the ACT to deal with the case of LLMs (section 4).I. Patternism and the Biological Process ViewChristian Coseru's intriguing piece focuses on the chapters on personal identity in Artificial You, proposing that when it comes to the nature of the self, one should adopt "a biologically informed process-ontology perspective on personal identity," which "means recognizing that organic identity is different in kind: it is an identity of form rather than of matter."His position is akin to the patternist views that I argue against in Artificial You in the sense that it embraces the idea that form, rather than the particular matter that comprises one's brain and body, is key to one's survival over time. But there are important differences as well. For transhumanists tend to believe that the self can survive radical changes in one's substrate, transitioning from thinking with a biological brain to computing with a network of microchips, for instance. Coseru departs from this, writing: [End Page 315]As organisms, persons endure by undergoing constant change on a fundamentally cellular level, including the change in the neural pathways that realize the cognitive processes of thinking and expressing these very thoughts right now.Furthermore:[I]f persons are processes rather than things, then they cannot be cognitively enhanced by augmenting, replacing or transferring their functioning parts (e.g., brain cells).I will call this position the "Biological Process View" (BPV). I have several questions about it. First, how does this position solve one of the key problems I raised for... (shrink)

The paper continues my earlier Chat with OpenAI’s ChatGPT with a Focused LLM Experiment (FLEX). The idea is to conduct Large Language Model (LLM) based explorations of certain areas or concepts. The approach is based on crafting initial guiding prompts and then follow up with user prompts based on the LLMs’ responses. The goals include improving understanding of LLM capabilities and their limitations culminating in optimized prompts. The specific subjects explored as research subject matter include a) diagonalization techniques as practiced (...) by Cantor, Turing, Gödel, and the advances, such as the Forcing techniques introduced by Paul Cohen and later investigators, b) Knowledge Hierarchies & Mapping Exercises, c) Discussions of IJ Good’s Speculations Concerning the First Ultraintelligent Machine, AGI, and Super- intelligence. Results suggest variability between major models like ChatGPT-4, Llama-3, Cohere, Sonnet and Opus. Results also point to strong dependence on users’ preexisting knowledge and skill bases. The paper should be viewed as ’raw data’ rather than a polished authoritative reference. (shrink)

Generative artificial intelligence systems based on transformers, including both text-generators like GPT-4 and image generators like DALL-E 3, have recently entered the popular consciousness. These tools, while impressive, are liable to reproduce, exacerbate, and reinforce extant human social biases, such as gender and racial biases. In this paper, we systematically review the extent to which DALL-E Mini suffers from this problem. In line with the Model Card published alongside DALL-E Mini by its creators, we find that the images it produces (...) tend to represent dozens of different occupations as populated either solely by men (e.g., pilot, builder, plumber) or solely by women (e.g., hairdresser, receptionist, dietitian). In addition, the images DALL-E Mini produces tend to represent most occupations as populated primarily or solely by White people (e.g., farmer, painter, prison officer, software engineer) and very few by non-White people (e.g., pastor, rapper). These findings suggest that exciting new AI technologies should be critically scrutinized and perhaps regulated before they are unleashed on society. (shrink)