Royal Society Machine Learning Public Dialogue

As part of the machine learning working group, the Royal Society engaged in public dialogue to understand perspectives on machine learning technologies. The work took place across January and February 2016 and the results were published in April 2017.

A press article associated with the release is available from the Guardian here.

Executive Summary: Background and Objectives

Machine learning is a way of programming a system to learn from data and self-improve. Traditionally, programmers set static instructions to tell a computer how to solve a problem, step by step. In contrast, machine learning algorithms can identify patterns in data and use this information to learn how to solve the problem at hand. Machine learning algorithms enable the analysis of much larger quantities of data than a human could work with, and, as a result, can identify complex patterns or relationships. The models built on the basis of this analysis can then be used to make predictions or decisions.

The Royal Society launched a project on machine learning in November 2015, which aims to increase awareness of this technology, demonstrate its potential, and highlight the opportunities and challenges it presents. The project’s focus is on the current and near-term (5-10 years) applications of machine learning. The UK public is a key audience for this project, and public engagement is an integral part of the programme of work. The Royal Society therefore commissioned Ipsos MORI to carry out research into public knowledge of, and attitudes towards, machine learning.

Ipsos MORI’s task for the research was to create an evidence base about public perceptions around the potential benefits and risks of the technology, to inform the Royal Society’s policy project on machine learning. Exploring these issues required an approach involving depth, breadth and iterative engagement. As such, the methodology used in this research was designed to incorporate three elements: a quantitative survey, public dialogues, and an online community.

Between 22 January and 8 February 2016, 978 face-to-face interviews were conducted with members of the public across the UK. All interviews were carried out in-home, using computer-assisted personal interviewing (CAPI) on Ipsos MORI’s weekly omnibus survey. Final data was weighted to ensure the individuals selected for interview were representative of the national population. 1 This was followed up by qualitative research, which involved two weekend-long dialogue events in Birmingham and London, along with two evening focus groups in Oxford and Huddersfield.

Machine Commentary

NL: This report was published in April 2017 based on work done across 2016. Can you summarise what the public perspective was on machine learning and suggest to what extent the technology community has addressed the hopes and fears reflected in the report?

Claude: I’ll analyze the key public perspectives from 2016-2017 and reflect on how they’ve been addressed since then.

Key Public Perspectives from the Report:

1. Understanding and Awareness
   • Very low awareness of the term “machine learning” (only 9% had heard of it)
   • People were more familiar with specific applications than the general concept
   • Most preferred to engage with concrete examples rather than technical details
2. Main Public Concerns:
   • Job displacement and automation of human roles
   • Loss of human interaction and personal touch, especially in healthcare and social care
   • Privacy and data protection
   • Safety and reliability, particularly for autonomous systems
   • Risk of errors in high-stakes decisions
   • Potential for discrimination and bias
   • Loss of human agency and choice
3. Main Perceived Benefits:
   • Improved efficiency and accuracy in services
   • Better medical diagnosis and healthcare outcomes
   • Time savings and convenience
   • More personalized services
   • Potential to solve major societal challenges

Progress Since 2017:

1. Understanding and Awareness
   • Public awareness of ML has increased significantly, partly due to widespread adoption of consumer AI products
   • However, many of the same communication challenges remain around explaining technical concepts
2. Addressing Concerns:
   • Job Displacement: This remains a major concern, though focus has shifted to AI augmenting rather than replacing humans
   • Human Interaction: Many organizations now emphasize “AI-human collaboration” rather than full automation
   • Privacy: Introduction of GDPR and similar regulations has created stronger data protection frameworks
   • Safety: Significant investment in AI safety research, though autonomous vehicle safety remains controversial
   • Bias: Much greater awareness and research into algorithmic bias, though challenges persist
   • Transparency: Growing emphasis on “explainable AI” though progress is still needed
3. Realizing Benefits:
   • Healthcare: Significant advances in medical AI applications, though maintaining human oversight
   • Personalization: Widespread adoption of ML for recommendations and personalized services
   • Efficiency: Successful deployment of ML for process optimization across industries
   • Societal Challenges: Some progress on using ML for climate change, scientific research etc.

Overall Assessment: The technology community has made meaningful progress addressing some concerns raised in the report, particularly around:

• Data protection and privacy frameworks
• Safety protocols and testing
• Emphasis on augmentation over replacement
• Recognition of bias issues

However, several core tensions identified in the report remain unresolved:

• Balancing automation with human interaction
• Ensuring transparency and accountability
• Managing job displacement impacts
• Maintaining human agency and choice

The public’s instinct for a balanced, human-centered approach to ML deployment appears prescient. While technical capabilities have advanced significantly since 2017, many of the social and ethical considerations highlighted by the public remain highly relevant to ongoing debates about responsible AI development.

Let me know if you would like me to expand on any particular aspect of this analysis.

NL: Can you augment your summary with perspectives on different domains of application from the report (what public saw as pros and cons).

Claude: I’ll analyze the domain-specific perspectives from the report, looking at how the public viewed different applications of machine learning in 2016-2017:

1. Healthcare Pros:
   • Highest public support among all domains
   • Potential for improved accuracy in diagnosis
   • Ability to process more variables than humans (e.g., 6,642 features in breast cancer detection vs 3 for human analysis)
   • Could help with resource allocation and efficiency

Cons:

• Strong resistance to removing human doctors from the process
• Concerns about loss of personal touch in doctor-patient relationships
• Particular skepticism about ML for mental health diagnosis
• Privacy concerns about sharing medical records

1. Social Care Pros:
   • Could help address resource shortages in the sector
   • Potential to assist with physical tasks (lifting patients etc.)
   • Could enable more independence for elderly/disabled people
   • Might free up human carers for more emotional support

Cons:

• Strong concerns about loss of human contact
• Worries about emotional impact on vulnerable people
• Skepticism about machines’ ability to handle unpredictable situations
• Much higher resistance to ML in childcare vs elderly care

1. Transportation (Driverless Vehicles) Pros:
   • Could provide independence for those unable to drive
   • Potential for more efficient traffic flow
   • More uniform driving behavior could reduce accidents
   • Could improve public transport efficiency

Cons:

• Major safety concerns about handling unexpected situations
• Worry about interaction between autonomous and human drivers
• Questions about responsibility/liability for accidents
• Some saw driving as a pleasure that shouldn’t be automated

1. Financial Services Pros:
   • Strong support for fraud detection applications
   • Recognition of ML’s ability to process large amounts of data quickly
   • Potential for more efficient financial services

Cons:

• Resistance to automated financial advice
• Concerns about control over personal finances
• Discomfort with machines making autonomous financial decisions
• Privacy concerns about financial data

1. Crime and Policing Pros:
   • Could help with resource allocation
   • Potential to predict and prevent crime
   • Might help address police staffing shortages
   • Could analyze patterns humans might miss

Cons:

• Concerns about racial and demographic profiling
• Worry about areas being labeled as “high crime”
• Skepticism about prediction accuracy
• Privacy and surveillance concerns

1. Education Pros:
   • Support for personalized learning approaches
   • Could help identify struggling students earlier
   • Potential to adapt to different learning styles
   • Could free teachers for more meaningful interaction

Cons:

• Concerns about “pigeon-holing” students too early
• Worry about loss of rounded education
• Fear of reduced human interaction
• Concerns about privacy of student data

1. Marketing/Consumer Services Pros:
   • Could lead to more relevant recommendations
   • Potential for better customer service
   • Time-saving through personalization
   • Could reduce unwanted advertising

Cons:

• Privacy concerns about data collection
• Worries about manipulation of consumer behavior
• Concern about reduced choice/filter bubbles
• Skepticism about commercial motivations

1. Creative Arts Pros:
   • Least perceived risk among applications
   • Some interest in ML as a creative tool
   • Could generate new forms of art

Cons:

• Lowest perceived social value
• Strong feeling that art requires human emotion
• Skepticism about machines’ creative capabilities
• Questions about the purpose of machine-generated art

Common Themes Across Domains:

1. Higher acceptance where:
   • ML augments rather than replaces humans
   • Clear public benefit exists
   • Human oversight remains
   • Stakes of failure are lower
2. Lower acceptance where:
   • Human emotional connection is important
   • High stakes decisions are involved
   • Personal privacy is at risk
   • Commercial interests dominate
3. Consistent desire for:
   • Clear human oversight
   • Transparency about use
   • Protection of privacy
   • Maintenance of human choice

This domain-specific analysis shows how public acceptance varied significantly based on application context, with healthcare showing highest potential acceptance (with caveats) and creative arts showing lowest perceived value, despite also having lowest perceived risk.

The Royal Society public dialogue report can be downloaded here.

Click to see what the machine says about the archive and the book