Periodic Reporting for period 2 - ALGOSOC (Algorithmic Societies: Ethical Life in the Machine Learning Age)
Período documentado: 2022-04-01 hasta 2023-09-30
What is the problem being addressed?
Rapid advancements in machine learning technologies are transforming social and political life in ways that uniquely challenge how we live in relation to others. The life chances of a person are now intimately connected to the data attributes that an algorithm has learned from the data patterns of unknown others. For example, a person’s attributed creditworthiness in the financial system, or their attributed riskiness in the criminal justice system, is increasingly inferred from the patterns learned from the conduct of other people. Understood in this way, processes of machine learning are always also practices of ethical and social connection with others. However, the technical architectures of machine learning algorithms are not commonly treated as containing normative, political and ethical relations to the world. The ALGOSOC project develops a new approach to understanding and responding to how ML algorithms remake the ethical relations that define a society.
Why is it important for society?
The ALGOSOC project examines how 21st century machine learning algorithms are learning to recognize, attribute, and infer the characteristics of entities (people, objects, and scenes). The forms of recognition, attribution and inference that characterised statistical societies are transforming with new techniques for deep learning and generative AI. For example, machine learning models are increasingly multi-modal in their approach to data from images, text, video or language. They are also increasingly able to traverse domains of the social world, building flexible models that can travel and transfer their learning.
What are the overall objectives?
The overarching project aim of ALGOSOC is to advance a new approach to understanding and responding to the consequences of machine learning algorithms for the norms and ethics of contemporary societies. The aim is underpinned by three research objectives, each of which is designed to address a fundamental aspect of the iterative practices of machine learning and to investigate how specific types of algorithmic process reconfigure ethical relations. T
Objective One: To understand how machine learning algorithms are generating a new societal ethics of recognition.
Objective Two: To analyse the ways in which societal differences are generated and negotiated through machine learning algorithms.
Objective Three: To investigate how machine learning generates inferential models of the future, and to understand the consequences of new forms of inference for the ethical relations of contemporary societies.
Rapid advancements in machine learning technologies are transforming social and political life in ways that uniquely challenge how we live in relation to others. The life chances of a person are now intimately connected to the data attributes that an algorithm has learned from the data patterns of unknown others. For example, a person’s attributed creditworthiness in the financial system, or their attributed riskiness in the criminal justice system, is increasingly inferred from the patterns learned from the conduct of other people. Understood in this way, processes of machine learning are always also practices of ethical and social connection with others. However, the technical architectures of machine learning algorithms are not commonly treated as containing normative, political and ethical relations to the world. The ALGOSOC project develops a new approach to understanding and responding to how ML algorithms remake the ethical relations that define a society.
Why is it important for society?
The ALGOSOC project examines how 21st century machine learning algorithms are learning to recognize, attribute, and infer the characteristics of entities (people, objects, and scenes). The forms of recognition, attribution and inference that characterised statistical societies are transforming with new techniques for deep learning and generative AI. For example, machine learning models are increasingly multi-modal in their approach to data from images, text, video or language. They are also increasingly able to traverse domains of the social world, building flexible models that can travel and transfer their learning.
What are the overall objectives?
The overarching project aim of ALGOSOC is to advance a new approach to understanding and responding to the consequences of machine learning algorithms for the norms and ethics of contemporary societies. The aim is underpinned by three research objectives, each of which is designed to address a fundamental aspect of the iterative practices of machine learning and to investigate how specific types of algorithmic process reconfigure ethical relations. T
Objective One: To understand how machine learning algorithms are generating a new societal ethics of recognition.
Objective Two: To analyse the ways in which societal differences are generated and negotiated through machine learning algorithms.
Objective Three: To investigate how machine learning generates inferential models of the future, and to understand the consequences of new forms of inference for the ethical relations of contemporary societies.
Subsection 2: work performed from the beginning of the project to the end of the reporting period.
ALGOSOC’s research achievements to date include:
• Ongoing research across key domains of machine learning as planned in the DoA (work packages 1-7). Results from the ‘recognition’ and ‘attribution’ projects are published in major peer-reviewed academic journals. Results from the ‘inference’ project are under review at a major peer-reviewed journal.
• The development of three additional discrete case studies that allow for the ethico-political investigation of major new aspects of machine learning systems that have emerged during the period of research. These case studies – rules and examples, generative adversarial networks, and transformer models – exemplify the growing capacity of current algorithms to be used in multiple social, economic, or political domains simultaneously. Research in these cases is enabling the ALGOSOC team to examine critically the impacts of generative AI, large language models, and new parallel infrastructures.
• The design and consolidation of a new methodological approach for tracing the social and ethico-political consequences of particular algorithmic models. The results from the methodological research are published in a major peer-reviewed academic journal.
• A series of academic events organised by the ALGOSOC team, developing and disseminating the project research findings.
• Regular presentation and dissemination of research results at major international conferences across a range of academic disciplines and also to audiences of the general public.
ALGOSOC’s research achievements to date include:
• Ongoing research across key domains of machine learning as planned in the DoA (work packages 1-7). Results from the ‘recognition’ and ‘attribution’ projects are published in major peer-reviewed academic journals. Results from the ‘inference’ project are under review at a major peer-reviewed journal.
• The development of three additional discrete case studies that allow for the ethico-political investigation of major new aspects of machine learning systems that have emerged during the period of research. These case studies – rules and examples, generative adversarial networks, and transformer models – exemplify the growing capacity of current algorithms to be used in multiple social, economic, or political domains simultaneously. Research in these cases is enabling the ALGOSOC team to examine critically the impacts of generative AI, large language models, and new parallel infrastructures.
• The design and consolidation of a new methodological approach for tracing the social and ethico-political consequences of particular algorithmic models. The results from the methodological research are published in a major peer-reviewed academic journal.
• A series of academic events organised by the ALGOSOC team, developing and disseminating the project research findings.
• Regular presentation and dissemination of research results at major international conferences across a range of academic disciplines and also to audiences of the general public.
Progress beyond the state-of-the-art and expected results at the end of the project: ALGOSOC is making substantial steps beyond the state of the art for ‘AI ethics’. Beyond the conventional attention to human oversight, datasets and regulation of outputs, the ALGOSOC project provides new social science frameworks for the critique of machine learning models as active participants in forms of world-making and meaning-making. Progress and expected results across the three themes of 'recognition', 'attribution', and 'inference' summarised as follows:
Recognition (WPs 1 & 2): the research team has traced detailed accounts of the deep neural networks that are transforming how people, groups, objects, and scenes are recognised and identified. The building of models has been interrogated in terms of the computer science literature and the practical empirical sites where systems are trialled and deployed. Expected results at the end of the project are that new algorithms for image recognition are substantially reconfiguring the ethical relations involved in relating to others (e.g. in biometric recognition systems at borders). As rules-based algorithms are superseded by examples-based deep learning, the ethics of recognition is also taking a new form. The focus on the relationship between training datasets and recognition has become particularly important with current transformer models such as ChatGPT. Ongoing research will ensure that full account of new GPT models (and impact on ethics of recognition) is taken by the end of the project.
Attribution (WPs 3 & 4): the ALGOSOC team has investigated how the computational 'attribute' is refiguring broader ideas of difference in society. For example, work on the impacts of cluster models on racialised differences has already been published. Work is also ongoing to investigate how attributes are being synthetically generated for flexible and modifiable machine models.
Inference (WPs 5 & 6): the ALGOSOC team are developing a new state-of-the-art in the study of futures-oriented machine learning practices. They are undertaking research on pre-trained models in order to better understand how futures are created through algorithmic techniques. Initial findings have been published, mapping the connections between algorithmic inferences and models of 'society', national and international.
Recognition (WPs 1 & 2): the research team has traced detailed accounts of the deep neural networks that are transforming how people, groups, objects, and scenes are recognised and identified. The building of models has been interrogated in terms of the computer science literature and the practical empirical sites where systems are trialled and deployed. Expected results at the end of the project are that new algorithms for image recognition are substantially reconfiguring the ethical relations involved in relating to others (e.g. in biometric recognition systems at borders). As rules-based algorithms are superseded by examples-based deep learning, the ethics of recognition is also taking a new form. The focus on the relationship between training datasets and recognition has become particularly important with current transformer models such as ChatGPT. Ongoing research will ensure that full account of new GPT models (and impact on ethics of recognition) is taken by the end of the project.
Attribution (WPs 3 & 4): the ALGOSOC team has investigated how the computational 'attribute' is refiguring broader ideas of difference in society. For example, work on the impacts of cluster models on racialised differences has already been published. Work is also ongoing to investigate how attributes are being synthetically generated for flexible and modifiable machine models.
Inference (WPs 5 & 6): the ALGOSOC team are developing a new state-of-the-art in the study of futures-oriented machine learning practices. They are undertaking research on pre-trained models in order to better understand how futures are created through algorithmic techniques. Initial findings have been published, mapping the connections between algorithmic inferences and models of 'society', national and international.