Journalistic Ethics in the Age of Artificial Intelligence: Towards an Update of Deontological Codes in the Ibero-American Context

How Can <i>IJDS</i> Authors, Reviewers, and Editors Use (and Misuse) Generative AI?

during the submission process, as it can lead to productive exchanges, as well as earlier and greater citation of published work (see The Effect of Open Access).

Response to M. Trengove & coll regarding "Attention is not all you need: the complicated case of ethically using large language models in healthcare and medicine".

Purpose This study aims to critically examine the socio-technical, economic and governance challenges emerging at the intersection of Generative artificial intelligence (AI) and Urban AI. By foregrounding the metaphor of “the moon and the ghetto” (Nelson, 1977, 2011), the issue invites contributions that interrogate the gap between technological capability and institutional justice. The purpose is to foster a multidisciplinary dialogue–spanning applied economics, public policy, AI ethics and urban governance – that can inform trustworthy, inclusive and democratically grounded AI practices. Contributors are encouraged to explore not just what GenAI can do, but for whom, how and with what consequences. Design/methodology/approach This study draws upon interdisciplinary literature from public policy, innovation studies, digital governance and urban sociology to frame the emerging governance challenges of Generative AI and Urban AI. It builds a conceptual foundation by synthesizing insights from comparative city case studies, innovation systems theory and normative policy frameworks. The approach is interpretive and exploratory, aiming to situate AI technologies within broader institutional, geopolitical and socio-economic contexts. The study invites contributions that adopt empirical, theoretical or practice-based methodologies addressing the governance of GenAI in cities and regions. Findings This study identifies a critical gap between the rapid technological advancements in Generative AI and the institutional readiness of public governance systems – particularly in urban contexts. It finds that current policy frameworks often prioritize efficiency and innovationism over democratic legitimacy, civic trust and inclusive design. Drawing on comparative global city experiences, it highlights the risk of reinforcing power asymmetries without robust accountability mechanisms. The analysis suggests that trustworthy AI is not a purely technical attribute but a political and institutional achievement, requiring participatory governance architectures and innovation systems grounded in public value and civic engagement. Research limitations/implications As an editorial introduction, this study does not present original empirical data but synthesizes key theoretical frameworks, case studies and policy debates to guide future research. Its analytical scope is conceptual and comparative, offering a foundation for submissions that further investigate Generative and Urban AI through empirical, normative and practice-based lenses. The limitations lie in its broad coverage and reliance on secondary sources. Nonetheless, it provides an agenda-setting contribution by highlighting the urgent need for interdisciplinary research into how AI reshapes public governance, institutional legitimacy and urban democratic futures. Practical implications This editorial offers a structured framework for policymakers, urban planners, technologists and public administrators to critically assess the governance of Generative and Urban AI systems. By highlighting international case studies and conceptual tools – such as public algorithmic infrastructures, civic trust frameworks and anticipatory governance – the article underscores the importance of institutional design, regulatory foresight and civic engagement. It invites practitioners to shift from techno-solutionist approaches toward inclusive, democratic and place-based AI governance. The reflections aim to support the development of trustworthy AI policies that are grounded in legitimacy, accountability and societal needs, particularly in urban and regional contexts. Social implications The editorial underscores that Generative and Urban AI systems are not socially neutral but carry significant implications for equity, representation and democratic legitimacy. These technologies risk reinforcing existing social hierarchies and systemic biases if not governed inclusively. This study calls for reimagining trust not as a technical feature but as a relational, contested dynamic between institutions and citizens. It encourages submissions that examine how AI reshapes the urban social contract, affects marginalized communities and challenges existing civic infrastructures. The goal is to promote AI governance frameworks that are pluralistic, just and reflective of diverse societal values and lived experiences. Originality/value This editorial offers a timely and conceptually grounded intervention into the emerging field of Urban AI and Generative AI governance. By framing the challenges through Richard R. Nelson’s metaphor of The Moon and the Ghetto, this study foregrounds the gap between technical capabilities and enduring societal injustices. The contribution lies in its interdisciplinary synthesis – bridging innovation systems, AI ethics, public policy and urban governance. It introduces a critical framework for assessing “trustworthy AI” not as a technical goal but as a democratic achievement and encourages research that is policy-relevant, equity-oriented and attuned to the institutional realities of AI in cities.

Getting AI Right: Introductory Notes on AI &amp; Society

Purpose This study examines the integration of generative and predictive artificial intelligence (AI) models within smart cities, focusing on how user readiness and technology adoption influence their contribution to sustainable urban development and governance. Design/methodology/approach The study applies a systematic literature review following PRISMA guidelines and synthesizes evidence from 50 peer-reviewed studies (2018–2025) indexed in Scopus and Web of Science. It combines bibliometric mapping using VOSviewer with thematic analysis to examine the drivers, barriers and governance mechanisms shaping the adoption of generative, predictive and hybrid applications in urban contexts. Findings Generative AI fosters participatory engagement, citizen co-design and interactive simulations, advancing SDG 11 (Sustainable Cities and Communities) and SDG 4 (Quality Education) through enhanced digital literacy and inclusive planning. Predictive AI improves operational efficiency, forecasting accuracy and data-driven policymaking, supporting SDG 9 (Industry, Innovation and Infrastructure) and SDG 13 (Climate Action) by promoting sustainable resource use and climate-resilient management. Hybrid AI integrates these strengths, addressing both social and operational aspects of smart city development and aligning with SDG 17 (Partnerships for the Goals) through cross-sector collaboration and shared governance. Collectively, these models contribute to broader sustainability goals, including SDGs 3, 7 and 12. Research limitations/implications This review acknowledges several key limitations. Reliance on Scopus and Web of Science may exclude regionally significant or domain-specific studies not indexed in these databases. The focus on English-language publications introduces potential language bias, possibly overlooking relevant research from non-English-speaking regions. Restricting the timeframe to 2018–2025 captures recent developments but may omit earlier foundational work or the most recent studies not yet indexed. Differences in research design, policy contexts and sample characteristics also affect comparability and limit generalizability. Future research should broaden data sources, include multilingual literature and adopt mixed-methods and longitudinal approaches to enhance contextual diversity and empirical robustness. Practical implications The findings provide practical guidance for policymakers, urban planners and technology developers to design AI governance systems that are transparent, accountable and aligned with the SDGs. Integrating generative and predictive AI can enhance operational efficiency, support participatory planning and promote responsible decision-making. The findings inform the development of adaptive policy frameworks that advance SDG 9 (Industry, Innovation and Infrastructure), SDG 11 (Sustainable Cities and Communities) and SDG 13 (Climate Action) through digital literacy initiatives, cross-sector collaboration and data-informed management. Strengthening these practices enables cities to translate AI’s potential into tangible contributions to inclusive and sustainable urban transformation. Social implications Integrating user readiness and digital literacy into AI adoption is essential for building inclusive and trustworthy smart cities. These efforts support SDG 4 (Quality Education), SDG 10 (Reduced Inequalities) and SDG 16 (Peace, Justice and Strong Institutions). Generative AI encourages citizen participation and collaborative planning, while predictive AI improves service accessibility and data-informed governance. Promoting ethical awareness and community engagement helps narrow digital divides and address bias. Collectively, these elements advance SDG 11 (Sustainable Cities and Communities) and SDG 17 (Partnerships for the Goals) by fostering socially responsive and transparent AI-driven urban development. Originality/value This review is among the first to integrate perspectives on user readiness and technology adoption with comparative insights into generative and predictive AI in smart cities. It advances understanding of how AI-driven urban innovation supports inclusivity, efficiency and sustainability, while outlining policy directions and a future research agenda for equitable and transparent AI governance.

The modern project of creating human-like artificial intelligence (AI) started after World War II, when it was discovered that electronic computers are not just number-crunching machines, but can also manipulate symbols. It is possible to pursue this goal without assuming that machine intelligence is identical to human intelligence. This is known as weak AI. However, many AI researcher have pursued the aim of developing artificial intelligence that is in principle identical to human intelligence, called strong AI. Weak AI is less ambitious than strong AI, and therefore less controversial. However, there are important controversies related to weak AI as well. This paper focuses on the distinction between artificial general intelligence (AGI) and artificial narrow intelligence (ANI). Although AGI may be classified as weak AI, it is close to strong AI because one chief characteristics of human intelligence is its generality. Although AGI is less ambitious than strong AI, there were critics almost from the very beginning. One of the leading critics was the philosopher Hubert Dreyfus, who argued that computers, who have no body, no childhood and no cultural practice, could not acquire intelligence at all. One of Dreyfus’ main arguments was that human knowledge is partly tacit, and therefore cannot be articulated and incorporated in a computer program. However, today one might argue that new approaches to artificial intelligence research have made his arguments obsolete. Deep learning and Big Data are among the latest approaches, and advocates argue that they will be able to realize AGI. A closer look reveals that although development of artificial intelligence for specific purposes (ANI) has been impressive, we have not come much closer to developing artificial general intelligence (AGI). The article further argues that this is in principle impossible, and it revives Hubert Dreyfus’ argument that computers are not in the world.

As artificial intelligence (AI) is becoming more integrated into the field of healthcare, medical students need to learn foundational AI literacy. Yet, traditional, descriptive teaching methods of AI topics are often ineffective in engaging the learners. This article introduces a new application of cinema to teaching AI concepts in medical education. With meticulously chosen movie clips from "Enthiran (Tamil)/Robot (Hindi)/Robo (Telugu)" movie, the students were introduced to the primary differences between artificial narrow intelligence (ANI), artificial general intelligence (AGI), and artificial super intelligence (ASI). This method triggered encouraging responses from students, with learners indicating greater conceptual clarity and heightened interest. Film as an emotive and visual medium not only makes difficult concepts easy to understand but also encourages curiosity, ethical consideration, and higher order thought. This pedagogic intervention demonstrates how narrative-based learning can make abstract AI systems more relatable and clinically relevant for future physicians. Beyond technical content, the method can offer opportunities to cultivate critical engagement with ethical and practical dimensions of AI in healthcare. Integrating film into AI instruction could bridge the gap between theoretical knowledge and clinical application, offering a compelling pathway to enrich medical education in a rapidly evolving digital age.NEW & NOTEWORTHY This article introduces a new learning strategy that employs film to instruct artificial intelligence (AI) principles in medical education. By introducing clips the from "Enthiran (Tamil)/Robot (Hindi)/Robo (Telugu)" movie to clarify artificial narrow intelligence (ANI), artificial general intelligence (AGI), and artificial super intelligence (ASI), the approach converted passive learning into an emotionally evocative and intellectually stimulating experience. Students experienced enhanced comprehension and increased interest in artificial intelligence. This narrative-driven, visually oriented process promises to incorporate technical and ethical AI literacy into medical curricula with enduring relevance and impact.

We may have to soon decide what kind of Artificial General Intelligence (AGI) computers we will build and how they will coexist with humans. Many predictions estimate that artificial intelligence will surpass human intelligence during this century. This poses a risk to humans: computers may cause harm to humans either intentionally or unintentionally. Here we outline a possible democratic society structure that will allow both humans and artificial general intelligence computers to participate peacefully in a common society.There is a potential for conflict between humans and AGIs. AGIs set their own goals which may or may not be compatible with the human society. In human societies conflicts can be avoided through negotiations: all humans have the about the same world view and there is an accepted set of human rights and a framework of international and national legislation. In the worst case, AGIs harm humans either intentionally or unintentionally, or they can deplete the human society of resources.So far, the discussion has been dominated by the view that the AGIs should contain fail-safe mechanisms which prevent conflicts with humans. However, even though this is a logical way of controlling AGIs we feel that the risks can also be handled by using the existing democratic structures in a way that will make it less appealing to AGIs (and humans) to create conflicts.The view of AGIs that we use in this article follows Kantian autonomy where a device sets goals for itself and has urges or drives like humans. These goals may conflict with other actors’ goals which leads to a competition for resources. The way of acting and reacting to other entities creates a personality which can differ from AGI to AGI. The personality may not be like a human personality but nevertheless, it is an individual way of behaviour.The Kantian view of autonomy can be criticized because it neglects the social aspect. The AGIs’ individual level of autonomy determines how strong is their society and how strongly integrated they would be with the human society. The critic of their Kantian autonomy is valid, and it is here that we wish to intervene.In Kantian tradition, conscious humans have free will which makes them morally responsible. Traditionally we think that computers, like animals lack free will or, perhaps, deep feelings. They do not share human values. They cannot express their internal world like humans. This affects the way that AGIs can be seen as moral actors. Often the problem of constraining AGIs has used a technical approach, placing different checks and designs that will reduce the likelihood of adverse behaviour towards humans. In this article we take another point of view. We will look at the way humans behave towards each other and try to find a way of using the same approaches with AGIs.

The fast evolution of Artificial Intelligence (AI) and the developing Artificial General Intelligence (AGI) capabilities transform how education operates, particularly through its effect on teacher training. AI-based systems provide adaptable learning spaces, and they offer both real-time assessment capabilities and data-driven educational method improvements. With its capability for human-level cognitive operations, AGI creates conditions to transform educator skill advancement processes. The article examines AI and AGI integration within teacher education programs by discussing their practical uses and advantages, together with the encountered challenges and ethical dilemmas. The analysis combines evaluative and creative AI tools like Gradescope and ChatGPT, and Carnegie Learning, with developing capabilities in AGI. The article uses detailed analysis, together with tables, along pictorial representations to show the necessity of achieving optimal teacher training through AI-human balanced cooperation. The research finds that AI brings efficiency benefits, but AGI's prospective function needs strict governance together with educational alignment, to maintain ethical, unbiased teacher education.

Technologically advanced artificial intelligence (AI) is making a significant contribution to strengthening national security. AI algorithms facilitate the processing of vast amounts of information, increasing the speed and accuracy of decision-making. Artificial intelligence and machine learning (AI/ML) are crucial for state and integrated hybrid attacks and protecting new threats in cyberspace. Existing AI capabilities have significant potential to impact national security by leveraging existing machine learning technology for automation in labor-intensive activities such as satellite imagery analysis and defense against cyber attacks. This article examines selected aspects of the impact of artificial intelligence on enhancing a state’s ability to protect its interests and its citizens., artificial intelligence through the use of neutron networks, predictive analytics and machine learning algorithms enables security agencies to analyse vast amounts of data and identify patterns indicative of potential threats. Integrating artificial intelligence into surveillance, border control and threat assessment systems enhances the ability to respond preemptively to security challenges. In addition, artificial intelligence algorithms facilitate the processing of vast amounts of information, increasing the speed and accuracy of decision-making by police authorities. The rapid development of AI raises a number of questions for its use in securing not only national security but protecting all citizens. In particular, it is worth answering the question How does artificial intelligence affect national security and clarifying the issue of how law enforcement agencies can use artificial intelligence to maximise the benefits of the new technology in terms of security and protecting communities from rising crime. The analysis is based on a descriptive method in describing the phenomenon; by explaining the concepts and applications of artificial intelligence to determine its role in the national security sphere. An analysis of the usefulness of artificial intelligence in particular in police operations is undertaken, with the aim of defending the thesis that, despite some threats to the protection of human rights from AI, it is becoming the best tool in the fight against all types of crime in the country. Technological advances in AI can also have many positive effects for law enforcement, and useful for law enforcement agencies, for example in facilitating the identification of persons or vehicles, predicting trends in criminal activities, tracking illegal criminal activities or illegal money flows, flagging and responding to fake news. Artificial intelligence (AI) has emerged as one of the biggest threats to information security, but efforts are being made to mitigate this new threat, but also to find solutions on how AI can become an ally in the fight against cyber-security, crime and terrorist threats. Artificial intelligence algorithms search huge datasets of communication traffic, satellite images and social media posts to identify potential cyber security threats, terrorist activities and organized crime. It is advisable, when analyzing the opportunities and threats that AI poses to national and public security, to gain a strategic advantage in the context of rapid technological change and also to manage the many risks associated with AI. The conclusion highlights the impact of AI on national security, creating a range of new opportunities coupled with challenges that government agencies should be prepared for in addressing ethical and security dilemmas. Furthermore, AI improves predictive analytics, thereby enabling security agencies to more accurately anticipate potential threats and enhance their preparedness by identifying vulnerabilities in the national security infrastructure

Artificial intelligence is promoted by means of incomprehensible advocacy through business majors that cannot easily be equated with human consciousness and abilities. Behavioral natural systems are quite different from language models and numeric inferences. This paper reviews through centuries of evolved human knowledge, and the resolutions as referred through the critics of mythology, literature, imagination of celluloid, and technical work products, which are against the intellect of both educative and fear mongering. Human metamorphic abilities are compared against the possible machine takeover and scope of envisaged arguments across both the worlds of ‘Artificial Intelligence’ and ‘Artificial General Intelligence’ with perpetual integrations through ‘Deep Learning’ and ‘Machine Learning’, which are early adaptive to ‘Artificial Narrow Intelligence’ — a cross examination of hypothetical paranoid that is gripping humanity in modern history. The potentiality of a highly sensitive humanoid and sanctification to complete consciousness at par may not be a near probability, but social engineering through the early stages in life may indoctrinate biological senses to a much lower level of ascendancy to Artificial Narrow Intelligence — with furtherance in swindling advancement in processes may reach to a pseudo-Artificial Intelligence {i}. There are no convincing answers to the discoveries from ancient scriptures about the consciousness of archetypal humans against an anticipated replication of a fulfilling Artificial Intelligence {ii}. Human use of lexicon has been the focal of automata for the past few years and the genesis for knowledge, and with the divergence of languages and dialects, scores of dictionaries and tools that perform bidirectional voice and text — contextual services are already influencing the lives, and appeasement to selective humanly incidentals is widely sustainable today {iii}. Synthesizing and harmonizing a pretentious labyrinthine gizmo is the center of human anxiety, but only evaluative research could corroborate that tantamount to genetic consciousness.

To emphasize the journal’s concern with AGI safety, we inaugurate Artificial General Intelligence (AGI) by focusing the first issue on Risks, Governance, and Safety &amp; Alignment Methods. Risks The AI Risk Repository: A Comprehensive Meta-Review, Database, and Taxonomy of Risks from Artificial Intelligence The most comprehensive AI risk taxonomy — 777 specific risks classified into 43 categories — to date has been created by workers collaborating from a half-dozen institutions. We except 11 key pages from the original 79-page report. Their ‘living’ Repository is online and free to download and share. The authors’ intention is to provide a common frame of reference for AI risks. Slattery et al.’s set of ~100 references is excellent and thorough. Thus, pouring through this study for your own specific interest is an efficient way to get on top of the entire current AI risk literature. The highest of their three taxonomy levels, the Causal Taxonomy, is categorized according to the cause of the risk, Human or AI the intention , Intentional action or Unintentional, and timing — Pre-deployment of the AI system or Post-deployment. The Causal Taxonomy can be used “for understanding how, when, or why risks from AI may emerge.” They also call readers’ attention to the AI Incident Database.[1] The Incident Database publishes a monthly roundup here. AI Risk Categorization Decoded (AIR 2024) By examining 8 government and 16 corporate AI risk policies, Zeng et al. seek to provide an AI risk taxonomy unified across public and private sector methodologies. They present 314 risk categories organized into a 4-level hierarchy. The highest level is composed of System &amp; Operational Risks, Content Safety Risks, Societal Risks, and Legal &amp; Rights Risks. Their first takeaway from their analysis is more categories is advantageous, allowing finer granularity in identifying risks and unifying risk categories across methodologies. Thus, indirectly they argue for the Slattery et al. taxonomy with double the categories. This emphasis on fine granularity parallels a comment made to me by Lance Fortnow, Dean of Illinois Institute of Technology College of Computing, on the diversity and specificity of human laws indicating a similar diversity may be necessary to assure AGI safety, and that recent governance proposals may be simplistic. Indeed, Zeng et al.’s second takeaway is that government AI regulation may need significant expansion. Few regulations address foundation models, for instance. And their third takeaway is that comparing AI risk policies from diverse sources is extremely helpful to develop an overall grasp of the issues – how different organizations conceptualize risk, for instance – and how to move toward international cooperation to manage AI risk. AIR-Bench 2024: A Safety Benchmark Based on Risk Categories from Regulations and Policies Applying the work just described above, Zeng et al. constructed an AI safety benchmark aligned with their unified view of private and public sector AI risk policy and specifically targeting the gap in regulation of foundation models they uncovered. They develop and test nearly 6000 risky prompts and find inconsistent responses across foundation models. Zeng et al. give examples of foundation model safety failures in response to various prompts. This work seems a significant advance toward an AGI safety certification conducted by an AI industry consortium or an insurance company consortium along the lines of, e.g., UL Solutions (previously Underwriters’ Laboratory). A Comprehensive Survey of Advanced Persistent Threat Attribution We wanted to publish this important article had to pull it due to a license conflict – please see their arXiv preprint. APT [Advanced Persistent Threat] attacks are attack campaigns orchestrated by highly organized and often state-sponsored threat groups that operate covertly and methodically over prolonged periods. APTs set themselves apart from conventional cyber-attacks by their stealthiness, persistence, and precision in targeting. This systematic review by Rani et al. of 137 papers focuses on the increasing development of automated means to detect AI and ML APTs early and identify the malevolent actors involved. They present the Automated Attribution Framework, which consists of 1) collecting the training data of past attacks, 2) preprocessing and enrichment of the training data, 3) the actual training and pattern recognition on the data, and 4) attribution — applying the trained models to identify the malevolent perpetrating actors. The open research questions summarized by Rani et al. lead toward AI taking an increasing role in APT attribution. Governance Excerpts from Aschenbrenner, Situational Awareness I was pointed to Leopold Aschenbrenner’s 165-page missive by Scott Aaronson’s blog, which said he knew Leopold during his sabbatical at OpenAI and recommended people give it a read and take it seriously. The essence of it is that if we extrapolate from recent AI progress, we will have AGI by 2030, and therefore, for national security, a Manhattan Project-style national AI effort, including nationalizing leading private AGI labs, should be mounted. Here we reprint his Part IV, “The Project,” advocating this controversial effort and describing his vision of how it will occur. I recommend anyone concerned about the dangers of AGI, and especially those working toward AGI, read Aschenbrenner’s entire book. Take a look at the Table of Contents preceding our reprint of “The Project.” And we reprint his Ch. V, “Parting Thoughts,” in our Commentary section. Soft Nationalization: How the US Government Will Control AI Labs Aschenbrenner advocates nationalizing leading AI labs into a high-security, top-secret, US federal government project. OK, how, exactly? A perfect complement to Aschenbrenner’s thoughts is given by Deric Cheng and Corin Katzke of Convergence Analysis. They examine how AGI R&amp;D nationalization could happen realistically, effectively, and efficiently. Their report outlines key issues and initial thoughts as a prelude to their own and others’ detailed proposals to come. It is a beautiful piece of work, IMHO. It is not impossible for private companies to develop AGI responsibly and securely, but the main goal of this journal is to make AGI safety the central debate in the AGI community, and the nationalized, Manhattan-style project point of view must be presented. Further, I find Aschenbrenner’s arguments to be persuasive and Cheng and Katzke’s thoughtful outline of how nationalization could actually occur to be convincing, e.g. (pg. 8): The US may be able to achieve its national security goals with substantially less overhead than total nationalization via effective policy levers and regulation… We argue that various combinations of the policy levers listed below will likely be sufficient to meet US national security concerns, while allowing for more minimal governmental intrusion into private frontier AI development. Acceptable Use Policies for Foundation Models Acceptable use policies are legally binding policies that prohibit specific uses of foundation models. Klyman surveys acceptable use policies from 30 developers encompassing 127 specific use restrictions cited in 184 articles. Like Zeng et al. in “AI Risk Categorization Decoded (AIR 2024),” Klyman highlights the inconsistent number and type of restrictions across developers and lack of transparency behind their motivation and enforcement, indicating the need to for developers to create a unified consensus acceptable use policy. The general motivations are to reduce legal and reputational risk. However, standing in the way of developers working to create a unified policy set is the motivation to use restrictions to hinder competition from using proprietary models. Enforcement can also hinder effective use of a foundation model. Acceptable use policies can be categorized into content restrictions (e.g. the top 4: misinformation, harassment, privacy, discrimination) and end use restrictions, e.g. Anthropic’s restriction on “model scraping,” which is someone training their own AI model on prompts and outputs from Anthropic’s model. Another use restriction is scaling up AI-created content distribution such as automated online posting. As with the Zeng et al. articles, Klyman’s article points the way to create a homogeneous acceptable use policy across a diverse AI ecosystem. Steve Omohundro comments: “…the AI labs’ ‘alignment work’ … is all about the AIs rather than their impact on the world. For goodness sake, the Chinese People's Liberation Army has already fine-tuned Meta's Llama 3.1 to promote Chinese military goals! And Meta's response was ‘that's contrary to our acceptable use policy!’" From the article: Without information about how acceptable use policies are enforced, it is not obvious that they are actually being implemented or effective in limiting dangerous uses. Companies are moving quickly to deploy their models and may in practice invest little in establishing and maintaining the trust and safety teams required to enforce their policies to limit risky uses. Safety Methods Benchmark Early and Red Team Often (Executive Summary excerpt) Two leading methods for uncovering AI safety breaches are 1) inexpensive benchmarking against a standardized test suite, such as prompts for large language models, and 2) longer, higher-cost but more informative intensive, interactive testing by human domain experts (“red-teaming”). Barrett et al., from the UC Berkeley Center for Long-Term Cybersecurity, advocate for this two-pronged approach indicated by the article title. They analyze the methods’ potential for eliminating LLM “dual” use, i.e. corrupting LLMs into creating chemical, biological, radiological, nuclear (CBRN) or cyber or other weaponry or attacks, but the methods apply to less dangerous risk testing as well. Essentially Barrett et al. advocate frequent use of benchmarks until a model attains a high safety score, followed by intensive red-teaming to test the model in more depth and yield more accuracy. Their paraphrase of the article title is: Benchmark Early and Often, and Red-Team Often Enough. Against Purposeful Artificial Intelligence Failures A paper that had to be written, and not surprisingly was, by Yampolskiy, who has sought to cover every aspect of AGI risks, is one arguing that intentionally triggering an AI disaster should not be entertained as an option to alert humanity to the danger of AGI. Models That Prove Their Own Correctness Especially in light of Dalrymple et al.’s governance proposal, Toward Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems, ‘models that prove their own correctness’ seems especially desirable, if not essential. Dalrymple et al. call for 1) a world model, 2) a safety specification, and 3), a means to verify the safety specification, a highly intriguing proposal, but which falls short of providing an example of such a model or means of verification (we hear that Dalrymple is working on an example). Paradise et al. describe two uses of interactive proof systems (IPS) combined with ML to allow a model to prove its own ‘correctness,’ as specified by the user of the model. The first method requires access to a training set of IPS transcripts (the sequence of interactions between the Verifier and Prover) in which the Verifier accepted the Prover’s probabilistic proof. The second method, Reinforcement Learning from Verifier Feedback (RLVF; note their intentional similarity to Reinforcement Learning from Human Feedback, RLHF) avoids the need for the accepted transcripts (which are in essence an external truth oracle) but only after training on such a verified transcript (its ‘base model’) using transcript learning. From then on it can generate its own emulated verified transcripts. The paper opens the door to other innovative applications of ML to IPS. This is a rather deep paper that requires further analysis to judge the realization of its promise. We look forward to a revised version after its peer review at an unspecified journal. We thank Syed Rafi for the pointer to the paper and Quinn Dougherty for inviting Orr Paradise to his safe AGI reading group. Language-Guided World Models: A Model-Based Approach to AI Control Model-based agents are artificial agents equipped with probabilistic “world models” that are capable of foreseeing the future state of an environment (Deisenroth and Rasmussen, 2011; Schmidhuber, 2015). World models endow these agents with the ability to plan and learn in imagination (i.e., internal simulation)…. Citing Dalrymple et al., Zhang et al. likewise extend the capabilities of world models to increase human control over AI. By adjusting the world model, humans can affect many context-sensitive policies simultaneously. However, for the human-AI interaction to be efficient, the world model must process natural language (NLP); hence, language-guided world models (LWMs). NLP also increases the efficiency of model learning by permitting them to read text. World models increase AI transparency, which NL interaction furthers by allowing humans to query models verbally. As an example, in Sec. 5.3, “Application: Agents that discuss plans with humans,” Zhang et al. describe an agent that uses its LWM to plan a task and then ask a human to review it for safety. Commentary Steve Omohundro, “Progress in Superhuman Theorem Proving?” Our co-founding editor Steve Omohundro is a strong proponent of Provably Safe AI, in which automated theorem-proving will play a major role.[2] Here Steve discusses current developments in using proof to lessen LLM hallucinations, the implications of superhuman theorem-proving for safe AGI and resources for interested readers. On Yampolskiy, “Against Purposeful Artificial Intelligence Failures” Topic Editor Jim Miller, Professor of Economics, Game Theory, and Sociology at Smith College, critiques Roman Yampolskiy’s argument against triggering a deliberate AI failure to wake the world up to AI dangers. Leopold Aschenbrenner, Situational Awareness, “Parting Thoughts” Aschenbrenner dismisses his critics as unrealistic and outlines the core tenets of “AI Realism.” Rowan McGovern, “Unhobbling Is All You Need?” Commentary on Aschenbrenner’s Situational Awareness McGovern questions Aschenbrenner’s fundamental assumption that “unhobbling” alone — “fixing obvious ways in which models are hobbled by default, unlocking latent capabilities and giving them tools, leading to step-changes in usefulness” — will result in his extrapolation of recent AI progress to predict the advent of AGI by 2030. McGovern: “Unhobbling conflates computing power with intelligence.” [1] https://incidentdatabase.ai/. “Like similar databases in aviation and computer security, the AI Incident Database aims to learn from experience so we can prevent or mitigate bad outcomes.” [2] Tegmark, M., &amp; Omohundro, S. (2023). Provably safe systems: the only path to controllable AGI. arXiv. Retrieved from https://arxiv.org/abs/2309.01933.

Artificial General Intelligence is the idea that someday an hypothetical agent will arise from artificial intelligence (AI) progresses, and will surpass by far the brightest and most gifted human minds. This idea has been around since the early development of AI. Since then, scenarios on how such AI may behave towards humans have been the subject of many fictional and research works. This paper analyzes the current state of artificial intelligence progresses, and how the current AI race with the ever faster release of impressive new AI methods (that can deceive humans, outperform them at tasks we thought impossible to tackle by AI a mere decade ago, and that disrupt the job market) have raised concerns that Artificial General Intelligence (AGI) might be coming faster that we thought. In particular, we focus on 3 specific families of modern AIs to develop the idea that deep neural networks, which are the current backbone of nearly all artificial intelligence methods, are poor candidates for any AGI to arise due to their many limitations, and therefore that any threat coming from the recent AI race does not lie in AGI but in the limitations, uses, and lack of regulations of our current models and algorithms. This article appears in the AI &amp; Society track.

Current challenges have accelerated the implementation of modern business concepts. Among the many practices of continuous business processes improvement is digitalization. Attention is focused on the benefits of digitalization in companies, which is to improve the processes quality, reduce their passage time, quickly fulfil orders, and hence increase customer loyalty. The concept of artificial intelligence is analysed, its three main types are identified: artificial narrow intelligence, general artificial intelligence, artificial superintelligence. Artificial narrow intelligence is focused on solving a narrowly defined, structured task; general artificial intelligence which is aimed at solving any problem, can respond to different environments and situations. Artificial superintelligence will be able to surpass people in absolutely everything, such as coping with creative tasks, decision-making and maintaining emotional relationships. The advantages of using artificial intelligence (accuracy in data processing, the ability to quickly analyse a large amount of information that will facilitate timely decision-making) are revealed. The main threats to the use of artificial intelligence (jobs disappearance, mass unemployment, loss of control over artificial intelligence – robots’ uncontrollability by humans) are also indicated. The most common technologies of artificial intelligence in enterprises (data science, machine learning, robotization) are considered. The business entities experience in the implementation of various artificial intelligence tools in operational activities, in the medical, legal, space, banking, educational spheres of activity, is presented. It was emphasized in the educational field that the annual increase in artificial intelligence is expected to reach 45% by 2030. It is also highlighted that artificial intelligence contributes to business development and global economic activity. The world's key players in the artificial intelligence market are considered, the top 10 world IT corporations are presented, the growth of their key performance indicators after the introduction of artificial intelligence technologies in goods and services is investigated.