Publications

@misc{Ganesh2024,

title = {Policy Approaches for Building a Responsible Ecosystem: Contextualising AI Governance Challenges Within Other Regulatory/Governance Sectors and Histories},

author = {Bhargavi Ganesh},

url = {https://zenodo.org/records/13132896},

doi = {https://doi.org/10.5281/zenodo.13132896},

year  = {2024},

date = {2024-07-30},

urldate = {2024-07-30},

abstract = {This report explains the historical role that governments have assumed in creating a more responsible ecosystem around new technologies, and protecting the safety and fundamental rights of citizens. This role is particularly crucial in the context of AI, given how ubiquitously it is used and how it is now almost impossible to opt out of its impacts. Analogies to past governance challenges, such as regulations targeting smog/air pollution, can help policymakers understand their role in the presence of externalised risks. Additionally, examining past governance responses to harms and risks associated with new technologies, enables policymakers to build AI governance mechanisms based on existing governance tools, rather than approaching AI governance as an unprecedented and daunting task. Moreover, governance remains an iterative exercise, in which initial governance efforts are typically crucial in enabling governments to revise and improve subsequent efforts.



Governance responses to AI are important because the failure to meet the concerns of affected parties risks creating responsibility gaps, or outcomes for which society bears the costs but no-one ultimately faces the consequences. It also inhibits the relational practice of responsibility, in which affected parties can consistently feed into the process of improving AI systems. The ability to have concerns heard and responded to is the hallmark of any working democracy. Persistent responsibility gaps threaten social trust and solidarity, by diminishing the trust that individuals and communities have both in institutions and each other. Although the concept of trust is more often used to describe the loss of trust by entities looking to use or adopt AI, the impacts of losing social trust and solidarity present even greater challenges, by threatening broader social well-being and political stability. },

keywords = {},

pubstate = {published},

tppubtype = {misc}

}

Close

Simkute, Auste

Explainability in Expert Contents: Challenges and Limitations in Supporting Domain Experts in AI-driven Decision-making Miscellaneous

2024.

Abstract | Links | BibTeX

Jones, Elliot; Hardalupas, Mahi; Agnew, William

Under the radar? Examining the evaluation of foundation models Miscellaneous

2024.

Abstract | Links | BibTeX

@misc{nokey,

title = {Under the radar? Examining the evaluation of foundation models},

author = {Elliot Jones and Mahi Hardalupas and William Agnew},

url = {https://www.adalovelaceinstitute.org/report/under-the-radar/},

year  = {2024},

date = {2024-07-25},

urldate = {2024-07-25},

abstract = {Global policy proposals for ensuring the safety of advanced artificial intelligence (AI) systems have centred on foundation model evaluations as an important method to identify and mitigate the risks these systems pose. The core goals of foundation model evaluations are to understand the foundation model and / or its impacts, including the model’s capabilities, risks, performance, behaviour and social impact.



Policymakers are seeking to use evaluations to provide clarity on appropriate and responsible uses of foundation models. They are incorporating evaluations into emerging regulatory proposals in the EU, UK and USA, and creating both voluntary and legally mandated requirements for developers to evaluate AI systems for different kinds of risks.



The EU’s newly passed AI Act requires developers of foundation models and general-purpose AI models to evaluate these systems for ‘systemic risks’. The Act has established an AI Office, which also has a mandate to evaluate general purpose AI models.



In the USA and UK, governments have secured voluntary commitments from major AI companies to allow external evaluations of their foundation models by newly established national AI safety institutes. France, Canada, Japan and Singapore have their own AI safety institutes with similar mandates to develop and run evaluations of foundation models.



Both governments and technology companies have described evaluations as a necessary component of effective foundation model governance. Many foundation model developers have hired dedicated evaluation teams to construct evaluations and test their models, and there is also a growing third-party evaluation industry in which contracted third parties can construct test models on behalf of a developer.



However, our research indicates that evaluations alone are not sufficient for determining the safety of foundation models, the systems built from them and their applications for people and society in real-world conditions. There is no agreed terminology or set of methods for evaluating foundation models, and evaluations need to be used alongside other tools including codes of practice, incident reporting and post-market monitoring. In practice, AI model evaluations are currently voluntary and subject to company discretion, leading to inconsistencies in quality and limited access for evaluators without pre-existing company relationships. Current policy proposals allow companies to selectively choose what evaluations to conduct, and fail to ensure evaluation results lead to meaningful action that prevents unsafe products from entering the market.},

keywords = {},

pubstate = {published},

tppubtype = {misc}

}

Close

Global policy proposals for ensuring the safety of advanced artificial intelligence (AI) systems have centred on foundation model evaluations as an important method to identify and mitigate the risks these systems pose. The core goals of foundation model evaluations are to understand the foundation model and / or its impacts, including the model’s capabilities, risks, performance, behaviour and social impact.

Policymakers are seeking to use evaluations to provide clarity on appropriate and responsible uses of foundation models. They are incorporating evaluations into emerging regulatory proposals in the EU, UK and USA, and creating both voluntary and legally mandated requirements for developers to evaluate AI systems for different kinds of risks.

The EU’s newly passed AI Act requires developers of foundation models and general-purpose AI models to evaluate these systems for ‘systemic risks’. The Act has established an AI Office, which also has a mandate to evaluate general purpose AI models.

In the USA and UK, governments have secured voluntary commitments from major AI companies to allow external evaluations of their foundation models by newly established national AI safety institutes. France, Canada, Japan and Singapore have their own AI safety institutes with similar mandates to develop and run evaluations of foundation models.

Both governments and technology companies have described evaluations as a necessary component of effective foundation model governance. Many foundation model developers have hired dedicated evaluation teams to construct evaluations and test their models, and there is also a growing third-party evaluation industry in which contracted third parties can construct test models on behalf of a developer.

However, our research indicates that evaluations alone are not sufficient for determining the safety of foundation models, the systems built from them and their applications for people and society in real-world conditions. There is no agreed terminology or set of methods for evaluating foundation models, and evaluations need to be used alongside other tools including codes of practice, incident reporting and post-market monitoring. In practice, AI model evaluations are currently voluntary and subject to company discretion, leading to inconsistencies in quality and limited access for evaluators without pre-existing company relationships. Current policy proposals allow companies to selectively choose what evaluations to conduct, and fail to ensure evaluation results lead to meaningful action that prevents unsafe products from entering the market.

Close

Tollon, Fabio

Technology and the Situationist Challenge to Virtue Ethics Journal Article

In: Science and Engineering Ethics, vol. 30, iss. 10, 2024.

Abstract | Links | BibTeX