Every other week in 2023 brought news of a breakthrough in artificial intelligence, and with each new announcement, a fair amount of excitement as well as apprehension followed. Early-adopter types were fascinated while others were shocked and even frightened.

Whether you’re transfixed or terrified, we all continue to learn more about what exactly ChatGPT, Dall-E, Google’s Bard, and other AI tools can do. Yet, we don’t know much about how those capabilities actually shape our perceptions of AI. What makes a tool seem more impressive? Which AI capabilities strike fear in our hearts?

So we set out to investigate which specific abilities and skills cause us to see an AI tool as impressive, dangerous, intelligent, and conscious. We ran an experiment in the US and UK that allows us to say with confidence which AI characteristics directly shape the way we see an AI. We surveyed a nationally representative sample in both countries and presented them with hypothetical AI tools that had different capabilities, exploring the impact of giving a tool different features on perceptions that the tool was, for instance, scary or impressive. This kind of experiment (commonly called a conjoint experiment but more generally a survey experiment) allows us to vary a handful of different attributes and identify which ones are the most influential on what we think.

In both the US and UK, one thing was very clear. A particular set of AI abilities was consistently seen as more impressive, dangerous, intelligent, and even conscious:

Emotional intelligence

We found that:

Emotional intelligence in AI had the greatest impact on how we view it. In our experiment, emotional intelligence included skills such as identifying sentiment in text, recognising emotion in faces, and the ability to empathise.

An AI’s ability to understand and share feelings with the user, comfort them, and have a conversation about how they feel had the greatest influence on how we view an AI tool. This example of advanced emotional intelligence made participants see AI as more impressive, more dangerous, more intelligent, and even more likely to be conscious.

An AI that can understand facial expressions and people's mood from images and speech was also influential. We included this skill as an example of moderate emotional intelligence, and it had a consistent impact on how we perceive an AI.

AI tools that can have complex dialogues with users as well as handling abstract concepts and emotional depth were seen as the most dangerous. We originally included this feature as the most advanced example of engaging with and generating language, but of course, there is a clear emotional component to that skill.

The ability to reason also influenced how we view an AI tool, especially in terms of perceived intelligence. Reasoning skills included solving mathematical problems, writing and editing computer code, as well as producing and testing new scientific theories.

Perhaps unsurprisingly, an AI’s ability to to adjust existing photos and images with filters was both the least impressive and least threatening attribute of everything we tested.

Methodological approach and conjoint analysis

A conjoint experiment allows us to create a profile of someone like a politician, job applicant, or in this case, an AI tool. We can then vary many different attributes at the same time and ultimately identify which attributes directly shape or determine someone’s perception. This kind of experiment allows us to determine the relative importance of each feature and confirm which ones actually shape our perceptions.

In this experiment, we wanted to see which kind of capabilities shaped how we view artificial intelligence - specifically how impressive, dangerous, intelligent, or conscious an AI was perceived to be.

We fielded this survey experiment to 2,027 adults in the US and UK between 15 to 18 December 2023 (approximately half in each country: 1,010 in the UK and 1,017 in the US). Each participant viewed an AI profile with a combination of attributes. We also varied how many attributes the AI had from 3 to 6 capabilities. These attributes could have been either a weak, medium, or strong variation of the AI’s ability to:

Generate visual art
Generate various content, such as text, music, and video
Engage with and generate language
Problem solve and reason
Engage with and understand emotions
Summarise, explain, and interpret information

Participants rated how impressive and dangerous they thought that AI was. They also responded to questions about seeing that AI tool as conscious or “as smart as a human”. Then the participants repeated this process for 3 additional profiles, resulting in a total of 4 profile ratings for each participant. We then assessed which aspects of the profiles caused people to see the AI differently.

This is an example of the AI tools rated by participants in our experiment:

Findings in more detail

High levels of emotional intelligence in an AI consistently made people see it as more impressive, more dangerous, more intelligent, and more likely to be conscious. For instance, an AI that “can empathise with the user to comfort them when they are upset, and have emotional conversations about how it feels” was the ability that influenced how people viewed the AI tool the most. Additional features were also influential, specifically:

Advanced language skills: Can have complex dialogues with users, handling abstract concepts and emotional depth
Advanced reasoning skills: Can produce new scientific theories based on evidence, and test them
Moderate emotional intelligence: Can understand facial expressions, and people's mood from images and speech

The ability to produce certain types of content also shaped how the AI tool was viewed, but these features had a weaker effect, such as:

Advanced artistic creation: Can create complex artwork that resonates with current social and cultural
Advanced video content creation: Can produce photorealistic-looking video content with coherent storytelling
Moderate reasoning: Can write and edit computer code
Advanced information processing: Can explain and interpret the meaning of text or images it is shown, including the metaphors or emotions they convey

While emotional intelligence skills made an AI tool seem more intelligent, reasoning abilities were actually the most important for perceptions of intelligence. The abilities to solve mathematical problems, write computer code, produce scientific theories, and test them had the greatest impact on viewing an AI tool as at least as intelligent as a human.

Who’s afraid of AI?

In addition to the actual skills of an AI tool, we each will bring with us pre-existing thoughts or feelings about AI, which could have an impact on how we view it, especially in terms of how threatening or dangerous we think it could be. With this in mind, we analysed which individual differences, such as demographics and attitudes, might have an effect on our perceptions of AI.

Within this experiment, Americans were slightly more likely to see an AI as threatening than British participants, and women in both countries were also slightly more likely to rate an AI tool as dangerous. Surprisingly, older people in both the US and UK were not more likely to see an AI tool as dangerous.

The biggest difference had to do with whether or not you had used an AI tool previously, such as ChatGPT, Google’s Bard, or Dall-E. About 40% of our sample had previously used an AI tool, and simply having some experience with such a tool made you much less likely to see the hypothetical AI tools in our experiment as dangerous.

Conclusions

Emotions - and not information - are the most impressive for artificial intelligence to understand, process, and engage.

While emotional intelligence was seen as the most valuable, impressive, and scary ability in an AI, reasoning skills make us think the AI tool is more intelligent.

Regardless of its features, people are less likely to be afraid of it if we’ve already had experience with another artificial intelligence tool. While other differences were observable between people, the biggest individual factor was whether or not you had used AI before.

An AI’s ability to have a conversation, write music or stories, make artwork, and summarise information were generally uninfluential on how we view an AI tool. Advances in artificial intelligence over the last year may have made even moderately advanced abilities to create content, engage with language, and process information seem relatively normal.

Impressive

How impressive do you think the AI tool described here is?

Responses ranges from 1 “not at all impressive” to 7 “very impressive”

Dangerous

To what extent do you agree or disagree with this statement? “This Artificial Intelligence tool is dangerous”

Responses ranges from 1 “strongly disagree” to 7 “strongly agree”

Cover image by Cash Macanaya on Unsplash

What makes an artificial intelligence tool impressive or scary?