The ISC’s guide offers a comprehensive framework designed to bridge the gap between high-level principles and practical, actionable policy. It responds to the urgent need for a common understanding of both the opportunities and risks presented by emerging technologies. This is an essential document for those working at the policy nexus in our rapidly changing digital era.
The framework explores the potential of AI and its derivatives through a comprehensive lens, encompassing human and societal wellbeing along with external factors like economics, politics, the environment, and security. Some aspects of the checklist may be more relevant than others, depending on the context, but better decisions seem more likely if all domains are considered, even if some can be quickly identified as irrelevant in particular cases. This is the inherent value of a checklist approach.
“In an era marked by rapid technological innovation and complex global challenges, the ISC’s framework for comprehensive and multidimensional analysis of the potential impacts empowers leaders to make informed, responsible decisions. It ensures that as we advance technologically, we do so with a careful consideration of the ethical, social, and economic implications”.
Peter Gluckman, ISC President
While high-level principles have been promulgated by UNESCO, the OECD, the European Commission and the UN, amongst others, and various discussions continue regarding issues of potential governance, regulation, ethics and safety, there is a large gap between such principles and a governance or regulatory framework. The ISC address this need through its new guide for policy-makers.
This guide for policy-makers is not intended to proscribe a regulatory regime, but rather to suggest an adaptive and evolving analytical framework which could underpin any assessment and regulatory processes that might be developed by stakeholders, including governments and the multilateral system.
“The framework is a critical step in the global conversation on AI as it provides a basis from which we can build consensus on the implications of the technology for both now and into the future”.
Hema Sridhar, Former Chief Science Adviser, Ministry of Defence, New Zealand and now Senior Research Fellow, University of Auckland, New Zealand.
Since October 2023, there have been several significant national and multilateral initiatives with further consideration of the ethics and safety of AI. The implications of AI on the integrity of some of our critical systems, including financial, government, legal and education, as well as different knowledge systems (including scientific and indigenous knowledge), are of increasing concern. The framework further reflects these aspects.
The feedback received from ISC Members and the international policy-making community to date is reflected in the revised version of the analytical framework, which is now released as a guide to policy-makers.
A guide for policy-makers: Evaluating rapidly developing technologies including AI, large language models and beyond
This discussion paper provides the outline of an initial framework to inform the multiple global and national discussions taking place related to AI.
Download the framework for use in your organization
Here we provide the framework tool as an editable Excel sheet for use in your organization. If you prefer an open source format, please contact secretariat@council.science.
Introduction
Rapidly emerging technologies present challenging issues when it comes to their use, governance and potential regulation. The ongoing policy and public debates on artificial intelligence (AI) and its use have brought these issues into acute focus. Broad principles for AI have been announced by UNESCO, the OECD, the UN and others, including the United Kingdom’s Bletchley Declaration, and there are emerging jurisdictional attempts at regulation of aspects of the technology through, for example, the European Union (EU) AI Act or the recent United States AI Executive Order.
While the use of AI is discussed at length in these and other fora, across geopolitical divides and in countries at all income levels, there remains an ontological gap between the development of high level principles and their incorporation into practice through either regulatory, policy, governance or stewardship approaches. The path from principle to practice is poorly defined, but given the nature and cadence of AI development and application, the variety of interest involved and the range of possible applications, any approach cannot be overly generic or prescriptive.
For these reasons, the non-governmental scientific community continues to play a particular role. The International Science Council (ISC) – with its pluralistic membership from the social and natural sciences – released a discussion paper in October 2023 presenting a preliminary analytical framework that considered the risks, benefits, threats and opportunities associated with rapidly moving digital technology. While it was developed to consider AI, it is inherently technology agnostic and can be applied to a range of emerging and disruptive technologies, such as synthetic biology and quantum. That discussion paper invited feedback from academics and policy-makers. The overwhelming feedback made conducting such an analysis necessary and stood as a valuable approach to address emerging technologies like AI.
The purpose of the framework is to provide a tool to inform all stakeholders – including governments, trade negotiators, regulators, civil society and industry – of the evolution of these technologies to help them frame how they might consider the implications, positive or negative, of the technology itself, and more specifically its particular application. This analytical framework has been developed independent of government and industry interests. It is maximally pluralistic in its perspectives, encompassing all aspects of the technology and its implications based on extensive consultation and feedback.
This discussion paper for policy-makers is not intended to proscribe a regulatory regime, but rather to suggest an adaptive and evolving analytical framework which could underpin any assessment and regulatory processes that might be developed by stakeholders, including governments and the multilateral system.
As decision-makers globally and nationally consider appropriate policy settings and levers to balance the risks and rewards of a new technology such as AI, the analytical framework is intended as a complementary tool to ensure the full suite of potential implications are adequately reflected.
Background: why an analytical framework?
The rapid emergence of technologies with the complexity and implications of AI is driving many claims of great benefit. However, it also provokes fears of significant risks, from individual to geostrategic level.1 Much of the discussion to date has been considered in a binary sense as publicly expressed views tend to take place at the extreme ends of the spectrum. The claims made for or against AI are often hyperbolic and – given the nature of the technology – difficult to assess.
A more pragmatic approach is necessary where hyperbole is replaced with calibrated and more granular evaluations. AI technology will continue to evolve, and history shows that virtually every technology has both beneficial and harmful uses. The question is, therefore: how can we achieve the beneficial outcomes from this technology, while reducing the risk of harmful consequences, some of which could be existential in magnitude?
The future is always uncertain, but there are sufficient credible and expert voices regarding AI and generative AI to encourage a relatively precautionary approach. In addition, a systems approach is necessary as AI is a class of technologies with broad use and application by multiple types of users. This means that the full context must be considered when considering the implications of an any AI use for individuals, social life, civic life, societal life and in the global context.
Unlike most other technologies, for digital and related technologies, the time between development, release and application is extremely short, largely driven by the interests of the production companies or agencies. By its very nature – and given it is based on the digital backbone – AI will have applications that are rapidly pervasive, as has already been seen with the development of large language models. As a result, some properties may only become apparent after release, meaning there is the risk of unforeseen consequences, both malevolent and benevolent.
Important societal values dimensions, particularly across different regions and cultures, will influence how any use is perceived and accepted. Furthermore, geostrategic interests are already dominating discussion, with sovereign and multilateral interests continuously intersecting and thus driving competition and division.
To date, much of the regulation of a virtual technology has largely been seen through the lens of ‘principles’ and voluntary compliance, although with the EU AI Act2 and similar we are seeing a shift to more enforceable but somewhat narrow regulations. Establishing an effective global or national technology governance and/or regulatory system remains challenging and there is no obvious solution. Multiple layers of risk-informed decision-making will be needed along the chain, from inventor to producer, to user, to government and to the multilateral system.
While high-level principles have been promulgated by UNESCO, the OECD, the European Commission and the UN, amongst others, and various high-level discussions continue regarding issues of potential governance, regulation, ethics and safety, there is a large gap between such principles and a governance or regulatory framework. This needs to be addressed.
As a starting point, the ISC considers developing a taxonomy of considerations that any developer, regulator, policy adviser, consumer or decision-maker could reference. Given the broad implications of these technologies, such a taxonomy must consider the totality of implications rather than a narrowly focused framing. Global fragmentation is increasing due to the influence of geostrategic interests on decision making, and given the urgency of this technology, it is essential for independent and neutral voices to persistently champion a unified and inclusive approach.
1) Hindustan Times. 2023. G20 must set up an international panel on technological change.
https://www.hindustantimes.com/opinion/g20-must-set-up-an-international-panel-on-technological-change-101679237287848.html
2) The EU Artificial Intelligence Act. 2023. https://artificialintelligenceact.eu
The development of an analytical framework
The ISC is the primary global non-governmental organization integrating natural and social sciences. Its global and disciplinary reach means it is well placed to generate independent and globally relevant advice to inform the complex choices ahead, particularly as the current voices in this arena are largely from industry or from the policy and political communities of the major technological powers.
Following a period of extensive discussion, which included the consideration of a non-governmental assessment process, the ISC concluded that its most useful contribution would be to produce an adaptive analytic framework that can be used as the basis for discourse and decision-making by all stakeholders, including during any formal assessment processes that emerge.
The preliminary analytical framework, which was released for discussion and feedback in October 2023, took the form of an overarching checklist designed for use by both government and non-governmental institutions. The framework identified and explored the potential of a technology such as AI and its derivatives through a wide lens that encompasses human and societal wellbeing, as well as external factors such as economics, politics, the environment and security. Some aspects of the checklist may be more relevant than others, depending on the context, but better decisions seem more likely if all domains are considered, even if some can be quickly identified as irrelevant in particular cases. This is the inherent value of a checklist approach.
The preliminary framework was derived from previous work and thinking, including the International Network for Governmental Science Advice’s (INGSA) report on digital wellbeing3 and the OECD Framework for the Classification of AI Systems,4 to present the totality of the potential opportunities, risks and impacts of AI. These previous products were more restricted in their intent given their time and context; there is a need for an overarching framework that presents the full range of issues both in the short and longer term.
Since its release, the discussion paper has received significant support from many experts and policy-makers. Many have specifically endorsed the recommendation to develop an adaptive framework that allows for deliberate and proactive consideration of the risks and implications of the technology, and in doing so, always considers the totality of dimensions from the individual to society and systems.
One key observation made through the feedback was acknowledgement that several of the implications considered in the framework are inherently multifaceted and extend across multiple categories. For example, disinformation could be considered from both the individual and geostrategic lens; thus, the consequences would be wide ranging.
The option to include case studies or exemplars to test the framework was also suggested. This could be used to develop guidelines to demonstrate how it could be used in practice in different contexts. However, this would be a significant undertaking and may confine how different groups perceive the use of this framework. It is best done by policy-makers working with experts in specific jurisdictions or contexts.
Since October 2023, there have been several significant national and multilateral initiatives with further consideration of the ethics and safety of AI. The implications of AI on the integrity of some of our critical systems, including financial, government, legal and education, as well as different knowledge systems (including scientific and indigenous knowledge), are of increasing concern. The revised framework further reflects these aspects.
The feedback received to date is reflected in the revised version of the analytical framework, which is now released as a guide to policy-makers.
While the framework is presented in the context of AI and related technologies, it is immediately transferable to the considerations of other rapidly emerging technologies such as quantum and synthetic biology.
3) Gluckman, P. and Allen, K. 2018. Understanding wellbeing in the context of rapid digital and associated transformations. INGSA.
https://ingsa.org/wp-content/uploads/2023/01/INGSA-Digital-Wellbeing-Sept18.pdf
4) OECD. 2022. OECD Framework for the Classification of AI systems. OECD Digital Economy Papers, No. 323,#. Paris, OECD Publishing.
https://oecd.ai/en/classificatio
The Framework
The following table presents the dimensions of a putative analytic framework. Examples are provided to illustrate why each domain may matter; in context, the framework would require contextually relevant expansion. It is also important to distinguish between the generic issues that arise during platform developments and those that may emerge during specific applications. No single consideration included here should be treated as a priority and, as such, all should be examined.
The issues are broadly grouped into the following categories as outlined below:
The table details dimensions that might need to be considered when evaluating a new technology.
🔴 INGSA. 2018. Understanding wellbeing in the context of rapid digital and associated transformations.
https://ingsa.org/wp-content/uploads/2023/01/INGSA-Digital-Wellbeing-Sept18.pdf
🟢 New descriptors (sourced through the extensive consultation and feedback and literature review)
🟡 OECD Framework for the Classification of AI Systems: a tool for effective AI policies.
https://oecd.ai/en/classification
| Criteria | Examples of how this may be reflected in analysis |
| 🟡Users’ AI competency | How competent and aware of the system’s properties are the likely users who will interact with the system? How will they be provided with the relevant user information and cautions? |
| 🟡 Impacted stakeholder | Who are the primary stakeholders that will be impacted by the system (individuals, communities, vulnerable, sectoral workers, children, policy-makers, professionals etc.)? |
| 🟡 Optionality | Are users provided with the opportunity to opt out of the system or are they given opportunities to challenge or correct the output? |
| 🟡Risks to human rights and democratic values | Does the system impact fundamentally on human rights, including but not limited to privacy, freedom of expression, fairness, non-discriminatory etc.? |
| 🟡Potential effects on people’s wellbeing | Do the system impact areas relate to the individual user’s wellbeing (job quality, education, social interactions, mental health, identity, environment etc.)? |
| 🟡 Potential for human labour displacement | Is there a potential for the system to automate tasks or functions that were being executed by humans? If so, what are the downstream consequences? |
| 🟡 Potential for identity, values or knowledge manipulation | Is the system designed or potentially able to manipulate the user’s identity or values set, or spread disinformation? |
| 🔴 Opportunities for self-expression and self-actualization | Is there a potential for artifice and self-doubt? Is there a potential for false or unverifiable claims of expertise? |
| 🔴 Measures of self-worth | Is there pressure to portray idealized self? Could automation replace a sense of personal fulfilment? Is there pressure to compete with the system in the workplace? Is individual reputation harder to protect against disinformation? |
| 🔴 Privacy | Are there diffused responsibilities for safeguarding privacy and are there any assumptions being made on how personal data are used? |
| 🔴 Autonomy | Could the AI system affect human autonomy by generating over-reliance by end-users? |
| 🔴 Human development | Is there an impact on acquisition of key skills for human development, such as executive functions or interpersonal skills, or changes in attention time affecting learning, personality development, mental health concerns etc.? |
| 🔴 Personal health care | Are there claims of self-diagnosis or personalized health care solutions? If so, are they validated to regulatory standards? |
| 🔴 Mental health | Is there a risk of increased anxiety, loneliness or other mental health issues, or can the technology ameliorate such impacts? |
| 🟢 Human evolution | Could large language models and artificial general intelligence change the course of human evolution? |
| 🟢 Human-machine interaction | Could the use lead to de-skilling and dependency over time for individuals? Are there impacts on human interaction? |
| Criteria | Examples of how this may be reflected in analysis |
| 🔴 Societal values | Does the system fundamentally change the nature of society, enable normalization of ideas previously considered anti-social, or breach societal values of the culture in which it is being applied? |
| 🔴 Social interactions | Is there an effect on meaningful human contact, including emotional relationships? |
| 🔴 Population health | Is there a potential for the system to advance or undermine population health intentions? |
| 🔴 Cultural expression | Is an increase in cultural appropriation or discrimination likely or more difficult to address? Does reliance on the system for decision-making exclude or marginalize culturally relevant sectional ties of society? |
| 🔴 Public education | Is there an effect on teacher roles or education institutions? Does the system emphasize or reduce the digital divide and inequity among students? Is the intrinsic value of knowledge or critical understanding advanced or undermined? |
| 🟢 Distorted realities | Are the methods used to discern what is true still applicable? Is the perception of reality compromised? |
| Criteria | Examples of how this may be reflected in analysis |
| 🟡 Industrial sector | In which industrial sector is the system deployed (finance, agriculture, health care, education, defence etc.)? |
| 🟡 Business model | In which business function is the system employed and in what capacity? Where is the system used (private, public, non-profit)? |
| 🟡 Impacts on critical activities | Would a disruption of the system’s function or activity affect essential services or critical infrastructures? |
| 🟡Breadth of deployment | How is the system deployed (narrow use within unit vs. widespread nationally/international)? |
| 🟡 Technical maturity | How technically mature is the system? |
| 🟢 Interoperability | Are there likely to be silos, nationally or globally, that inhibit free trade and impact cooperation with partners? |
| 🟢 Technological sovereignty | Is a desire for technological sovereignty driving behaviours, including control over the entire AI supply chain? |
| 🔴 Income redistribution and national fiscal levers | Could the core roles of the sovereign state be compromised (e.g., reserve banks)? Will the state’s ability to meet citizens’ expectations and implications (social, economic, political etc.) be advanced or reduced? |
| 🟢 Digital divide (AI divide) | Are existing digital inequalities exacerbated or new ones created? |
| Criteria | Examples of how this may be reflected in analysis |
| 🔴 Governance and public service | Could the governance mechanisms and global governance system be affected positively or negatively? |
| 🔴 News media | Is public discourse likely to become polarized and entrenched at a population level? Will there be an effect on the levels of trust in the Fourth Estate? Will conventional journalist ethics and integrity standards be further affected? |
| 🔴 Rule of law | Will there be an effect on the ability to identify individuals or organizations to hold accountable (e.g., what kind of accountability to assign to an algorithm for adverse outcomes)? Is there a loss of sovereignty created (environmental, fiscal, social policy, ethics etc.)? |
| 🔴Politics and social cohesion | Is there a possibility of more entrenched political views and less opportunity for consensus building? Is there the possibility of further marginalizing groups? Are adversarial styles of politics made more or less likely? |
| 🟢 Social licence | Are there privacy concerns, trust issues and moral concerns that need to be considered for stakeholder acceptance of the use? |
| 🟢 Indigenous knowledge | Could Indigenous knowledge and data be corrupted or misappropriated? Are there adequate measures to protect against misrepresentation, misinformation and exploitation? |
| 🟢 Scientific system | Is academic and research integrity compromised? Is there a loss of trust in science? Are there possibilities of misuse, overuse or abuse? What is the consequence of the practice of science? |
| Criteria | Examples of how this may be reflected in analysis |
| 🟢 Precision surveillance | Are the systems trained on individual behavioural and biological data and could they be used to exploit individuals or groups? |
| 🟢 Digital competition | Could state or non-state actors (e.g. large technology companies) harness systems and data to understand and control other countries’ populations and ecosystems, or undermine jurisdictional control? |
| 🟢 Geopolitical competition | Could the system stir competition between nations over harnessing individual and group data for economic, medical and security interests? |
| 🟢 Shift in global powers | Is the status of nation-states as the world’s primary geopolitical actors under threat? Do technology companies wield power once reserved for nation-states and have they become independent, sovereign actors (emerging technopolar world order)? |
| 🟢 Disinformation | Would the system facilitate the production and dissemination of disinformation by state and non-state actors with an impact on social cohesion, trust and democracy? |
| 🟢 Dual-use applications | Is there a possibility for both military application as well as civilian use? |
| 🟢 Fragmentation of global order | Could silos or clusters of regulation and compliance develop that hinder cooperation, lead to inconsistencies in application and create room for conflict? |
| Criteria | Examples of how this may be reflected in analysis |
| 🟢 Energy and resource consumption (carbon footprint) | Do the system and requirements increase uptake of energy and resource consumption over and above the efficiency gains obtained through the application? |
| 🟢Energy source | Where is the energy sourced from for the system (renewable vs. fossil fuels etc.)? |
| Criteria | Examples of how this may be reflected in analysis |
| 🟡 Direction and collection | Are the data and input collected by humans, automated sensors or both? |
| 🟡 Provenance of the data | Are the data and input from experts provided, observed, synthetic or derived? Are there watermark protections to confirm provenance? |
| 🟡 Dynamic nature of the data | Are the data dynamic, static, dynamic updated from time to time or real-time? |
| 🟡 Rights | Are the data proprietary, public or personal (related to identifiable individuals)? |
| 🟡 Identifiability and personal data | If personal, are the data anonymized or pseudonymized? |
| 🟡 Structure of the data | Are the data structured, semi-structured, complex structured or unstructured? |
| 🟡 Format of the data | Is the format of the data and metadata standardized or non-standardized? |
| 🟡 Scale of the data | What is the dataset’s scale? |
| 🟡 Appropriateness and quality of the data | Is the dataset fit for purpose? Is the sample size adequate? Is it representative and complete enough? How noisy are the data? Is it error prone? |
| Criteria | Examples of how this may be reflected in analysis |
| 🟡 Information availability | Is any information available about the system’s model? |
| 🟡 Type of AI model | Is the model symbolic (human-generated rules), statistical (uses data) or hybrid? |
| 🟡 Rights associated with model | Is the model open-source or proprietary, self- or third-party managed? |
| 🟡 Single of multiple models | Is the system composed of one model or several interlinked models? |
| 🟡 Generative or discriminative | Is the model generative, discriminative or both? |
| 🟡 Model building | Does the system learn based on human-written rules, from data, through supervised learning or through reinforcement learning? |
| 🟡 Model evolution (AI drift) | Does the model evolve and/or acquire abilities from interacting with data in the field? |
| 🟡 Federated or central learning | Is the model trained centrally or in several local servers or ‘edge’ devices? |
| 🟡 Development/maintenance | Is the model universal, customizable or tailored to the AI actor’s data? |
| 🟡 Deterministic or probabilistic | Is the model used in a deterministic or probabilistic manner? |
| 🟡 Model transparency | Is information available to users to allow them to understand model outputs and limitations or use constraints? |
| 🟢 Computational limitation | Are there computational limitations to the system? Is it possible to predict capability jumps or scaling laws? |
| Criteria | Examples of how this may be reflected in analysis |
| 🟡 Task(s) performed by system | What tasks does the system perform (recognition, event detection, forecasting etc.)? |
| 🟡 Combining tasks and actions | Does the system combine several tasks and actions (content generation systems, autonomous systems, control systems etc.)? |
| 🟡 System’s level of autonomy | How autonomous are the system’s actions and what role do humans play? |
| 🟡 Degree of human involvement | Is there some human involvement to oversee the overall activity of the AI system and the ability to decide when and how to use the AI system in any situation? |
| 🟡 Core application | Does the system belong to a core application area such as human language technologies, computer vision, automation and/or optimization or robotics? |
| 🟡 Evaluation | Are standards or methods available for evaluating system output? |
How could this framework be used?
This framework could be used in many ways, including:
A way forward
In summary, the analytical framework is provided as the basis of a toolkit that could be used by stakeholders to comprehensively look at the any significant developments either of platforms or of use in a consistent and systematic manner. The dimensions presented in this framework have relevance from technology assessment to public policy, from human development to sociology, and futures and technology studies. While developed for AI, this analytical framework has much broader application to any other emerging technology.
6 UN AI Advisory Board. 2023. Interim Report: Governing AI for Humanity. https://www.un.org/sites/un2.un.org/files/ai_advisory_body_interim_report.pd
Acknowledgements
Many people have been consulted and provided feedback in the development of both the initial discussion paper and feedback following its release. Both papers were drafted by Sir Peter Gluckman, President, the ISC and Hema Sridhar, former Chief Science Adviser, Ministry of Defence, New Zealand and now Senior Research Fellow, University of Auckland, New Zealand.
In particular, the ISC Lord Martin Rees, former President of the Royal Society and Co-Founder of the Centre for the Study of Existential Risks, University of Cambridge; Professor Shivaji Sondhi, Professor of Physics, University of Oxford; Professor K Vijay Raghavan, former Principal Scientific Adviser to the Government of India; Amandeep Singh Gill, the UN Secretary General’s Envoy on Technology; Seán Ó hÉigeartaigh, Executive Director, Centre for the Study of Existential Risks, University of Cambridge; Sir David Spiegelhalter, Winton Professor of the Public Understanding of Risk, University
of Cambridge; Amanda-June Brawner, Senior Policy Adviser and Ian Wiggins, Director of International Affairs, Royal Society, United Kingdom; Dr Jerome Duberry, Managing Director and Dr Marie-Laure Salles, Director, Geneva Graduate Institute; Chor Pharn Lee, Centre for Strategic Futures, Prime Minister’s Office, Singapore; Barend Mons and Dr Simon Hodson, the Committee on Data (CoDATA); Professor Yuko Harayama, Former Executive Director, RIKEN; Professor
Rémi Quirion, President, INGSA; Dr Claire Craig, University of Oxford and Former Head of Foresight, Government Office of Science; Prof Yoshua Bengio, UN Secretary General’s Scientific Advisory Board and at Université de Montréal; and the many others who provided feedback to the ISC on the initial discussion paper.
Preparing National Research Ecosystems for AI: Strategies and progress in 2024
This working paper from the ISC’s think tank, the Centre for Science Futures, provides fundamental information and access to resources from countries from all parts of the world, at various stages of integrating AI into their research ecosystems.
Read more about Snapshots of reform: Researcher evaluation within science organizations
Read more about Report on the ISC Meeting of Pacific Scholars
Read more about The Key Principles for Scientific Publishing
