Understanding, trust, and thoughts about regulation

To understand how explainable the British public think a decision made by an AI system should be when explainability trades off with accuracy, we first informed participants that: ​‘Many AI systems are used with the aim of making decisions faster and more accurately than is possible for a human. However, it may not always be possible to explain to a person how an AI system made a decision.’ We then asked people which of the following statements best reflects their personal opinion:

• Making the most accurate AI decision is more important than providing an explanation.

• In some circumstances an explanation should be given, even if that makes the AI decision less accurate.

• An explanation should always be given, even if that makes all AI decisions less accurate.

• Humans, not computers, should always make the decisions and be able to explain them to the people affected.

When there are trade-offs between the explainability and accuracy of AI technologies, the British public value the former over the latter: it is important for people to understand how decisions driven by AI are made.

Figure 6 shows that only 10% of the public feel that ​‘making the most accurate AI decision is more important than providing an explanation’, whereas a majority choose options that reflect a need for explaining decisions. 

Specifically, almost one third (31%) indicate that humans should always make the decisions (and be able to explain them), followed by 26% who think that ​‘sometimes an explanation should be given, even if it reduces accuracy’ and another 22% who choose ​‘an explanation should always be given, even if it reduces accuracy’. 

People’s preferences for explainable AI decisions dovetail with the importance of transparency and accountability demonstrated by people’s specific concerns about each technology. Here, for all technologies (except for driverless cars and virtual health assistants) the proportion of concerns mentioning ​‘it is unclear how decisions are made’ is higher than mentions of ​‘inaccuracy’.

People’s preferences for explainability over accuracy change across age groups.

Older people choose explainability and human involvement over accuracy to a greater extent than younger people. For those aged 18–44, ​‘sometimes an explanation should be given even if it reduces accuracy’ was the most popular response (Figure 7). At the youngest end of the age spectrum (18–24) ​‘humans should always make the decisions and be able to explain them’ is the least popular response, whereas this becomes the first choice from 45+ and above and the highest for respondents aged 65+. 

To find out about people’s views on the regulation of AI, we asked people to indicate what (if anything) would make them more comfortable with AI technologies being used. Participants could select as many they felt applied from a list of seven possible options

Public attitudes suggest a need for regulation that involves redress and the ability to contest AI-powered decisions.

People most commonly indicated that ​‘laws and regulations that prohibit certain uses of technologies and guide the use of all AI technologies’ would increase their comfort with the use of AI, with 62% in favour. People are also largely supportive of ​‘clear procedures for appealing to a human against an AI decision’ (selected by 59%). Adding to the concerns expressed about data security and accountability, 56% of the public want to make sure that ​‘personal information is kept safe and secure’ and 54% want ​‘clear explanations of how AI works’. 

Figure 8 shows the proportion of people selecting each option when asked what, if anything, would make them more comfortable with AI technologies being used. 

We also asked participants who they think should be most responsible for ensuring AI is used safely from a list of seven potential actors. People could select up to two options.

The British public want regulation of AI technologies. ​‘An independent regulator’ is the most popular choice for governance of AI.

Figure 9 shows 41% of people feel that ​‘An independent regulator’ should be responsible for the governance of AI, the most popular choice of the seven presented. Patterns of preferred governance do not change notably depending on whether people feel well informed about new technologies or not.

Results add to a PublicFirst poll conducted in March 2023 with 2,000 UK adult respondents which found that 62% of respondents supported the creation of a new government regulatory agency, similar to the Medicines and Healthcare Products Regulatory Agency (MHRA), to regulate the use of new AI models.

People’s preferences for the governance of AI changes across age groups. 

While people overall most commonly select ​‘an independent regulator’, Figure 10 shows 43% of 18–24-year-olds think that the ​‘companies developing the technology’ should be most responsible for ensuring AI is used safely. In contrast, only 17% of people over 55 select this option. 

This could reflect more in-depth experiences by young people with different technologies and associated risks, and therefore demands for more responsibility on developers. Especially since young people also report the highest exposure to technology driven problems such as online harms’. That 18–24-year-olds most commonly say that the companies developing the technologies should be responsible for ensuring AI is used safely raises questions about private companies’ corporate responsibility alongside regulation.

To understand people’s concerns about who develops AI technologies, we asked people how concerned, if at all, they feel about different actors producing AI technologies. 

We asked this in the context of hospitals asking an outside organisation to produce AI technologies that predict the risk of developing cancer from a scan, and the Department for Work and Pensions (DWP) asking an outside organisation to produce AI technologies for assessing eligibility for welfare benefits. 

We asked people how concerned they are about each of the following groups producing AI in each context: 

• private companies

• not-for-profit organisations (e.g. charities)

• another governmental body or department

• universities/​academic researchers.

For both the use of AI in predicting cancer from a scan, and assessing eligibility for welfare benefits, the British public are most concerned by private companies developing the technologies and least concerned by universities and academic researchers developing the technologies

For the development of AI which may be used to assist the Department for Work and Pensions in assessing eligibility for welfare benefits, the public are most concerned about private companies developing the technology, with 66% being somewhat or very concerned. Just over half, 51%, of people are somewhat or very concerned about another governmental body or department developing the technology, and 46% somewhat or very concerned about not-for-profit organisations developing the technology. 

People are generally least concerned about universities or academic researchers developing this technology, with 43% being somewhat or very concerned. While this is the lowest percentage of concern compared to other stakeholders, this is still a sizable proportion of people expressing concern, which suggests the need for more trusted stakeholders to also be transparent about their role and approach to developing technologies. 

Regarding the development of AI that may help healthcare professionals predict the risk of cancer from a scan, there is a very similar pattern of concerns over who develops the technology. People are most concerned with private companies developing the technology with 61% being somewhat or very concerned, followed by a governmental body (44%). People are less concerned with not-for-profit organisations and universities or academic researchers developing the technology. Overall level of concern about developers was lower for technologies that predict risk of cancer than technologies which help assess eligibility for welfare.

Figure 11 shows the extent to which people feel concerned by the following actors developing new technologies to assess eligibility for welfare benefits and predict the risk of developing cancer: private companies, governmental bodies, not-for-profit organisations and universities/​academic researchers.

While we asked about concerns over the development of a specific technology rather than overall trust, our findings resonate with results from the second wave of a CDEI survey on public attitudes towards AI, which found that on average, respondents most trusted the NHS and academic researchers to use data safely, while trust in government, big tech companies and social media companies was lower.