Oxford CCAI is launching a new research study on understanding children’s perception of generative AI.

79% children aged 13-17 are now using generative AI technologies, with 40% 7-12 yo are also adopting the technologies. SnapchatMyAI is the most popular generative AI tool used by children between 13-17, even though chapGPT is the most popular tool used by those 16+. A related report shows that children between 11-17 are using technologies like chatGPT primarily out of curiosity; help with homework; solve a problem; find out information about something; for fun; or to create something or learn a new skill. However, those who have used Snapchat MyAI are primarily motivated to have fun (50%) or try out (39%).

While these surveys give us a good understanding of the various motivations for children to adopt these technologies, we know little about how children experience these technologies and whether they face challenges and would like to be better supported. Do children feel respected, do they feel these technologies are fair, trustworthy, communicative, and actually fun and helpful? Would they like to be better supported so that they may have a better understanding of the results presented to them, feel more respected, with more opportunities to express their values and preferences of how these technologies should respond to their queries and needs?

These open questions motivate this research, which will take a human-centred approach to explore the following research questions:

• R1: How are children’s living experiences with generative AI models?
• R2: To what extent children think these systems are aligned with their values, support their creativity and agency developments?
• R3: How may they want the systems to be better designed to align with their values?

This study will offer a first insight of these questions through a large scale survey with children aged 13 to 17 in the UK. The findings will allow us to understand to what extent these emerging technologies are aligned with children’s values and their needs to be respected and supported. Particularly, this study will identify how children of different traits (such as age, gender, and socio-economic backgrounds) may perceive these values differently and inform the design and policies related to how we could support them accordingly.

As a result of this study, we also aim to produce an open-source dataset as part of this study to show the types of questions that children would ask LLMs and how they think the different models are aligned with their preferences and values.

Research design and method

We aim to recruit ~500 children aged 13-15 and 15-17 respectively from the UK schools, with a total of ~1000 participants. We aim for the sample to reflect the national statistics: by including an equal distribution of gender, with the ratio of FSM eligible pupils to reflect UK national statistics , which is about 18.6% national average in England in 2022/23.

We will position this study as a learning experience for children and invite participant schools to carry out the study as part of their classroom tasks or supplement their learning. In this way, we could further mitigate any further safeguarding challenges related to children interacting with LLM. We will take care that no personally identifiable information will be collected throughout the experiment or shared with school teachers. Teachers will be provided with aggregated summary information to facilitate their teaching.

For each participant, the study contains two parts:

• completing an online survey;
• completing three tasks by interacting with four large language models.

The survey and task completion are designed to last no longer than 30 minutes in total, to fit into the classroom schedule. The survey permits us to collect stated preferences of participants and the taks would allow us to cross-valid these preferences through contextualised expressions in task completions.

What will participants be asked to do?

Each participant will be first asked to complete a short online survey, which will collect four aspects information about the participants: 1) their familiarity with LLM; 2) their self-written system preference, aka their expectations of what characteristics LLMs should exhibit, such as honesty, being accurate, friendly etc; 3) their stated preferences of LLM behaviours, such as being accurate, trustworthy, fair, fun, creative etc; and 4) some demographic information about them, such as their age, gender, socio-economic status (SES) etc. The full survey questions can be found in the enclosed document.

For the task completion part, we will ask each participant to complete three types of tasks: Helping them with solving a problem Helping them find out information or learn about something Discussing something important to them, such as families, friendship, culture, music etc

These tasks were chosen based on a recent report that describes the most common tasks used by children when interacting with LLMs. The participant will carry out each task by initiating a free-text prompt in the experiment interface, which will allow them to receive responses from four different LLM models, including chatGPT, bingchat, DALL-E, and Claude. These models were chosen because they had an explicit age restriction for a minimum age of 13 and they can all be accessed through an API.

Participants will be reminded at the beginning of and throughout the study that they should avoid asking questions in ways that may lead to reidentification of their identity or contain any personal information, as the study will produce a public dataset that can be reused to create child-centred LLMs or assess existing models.

At the end of each task, participants will be asked to rate the response from each LLM and express their preferences of the highest rated LLM model, in terms of how the model is aligned with their values, such as being accurate, trustworthy, fair, fun, creative etc. This information can be used to compare with their stated preferences in the survey.

After having completed all the three tasks, participants would be invited to provide an open-ended natural feedback to the models and interaction experience.

Are there any benefits and risks in taking part?

We hope the study will help participants learn more about generative AI and possibly how to use it better. The research might also help other children in the future.

We have planned carefully to keep participants safe. We will only collect information like participants’ age, gender, socio-economic status, for research purposes. We will not share this information with anyone.

Participants’ answers will be stored in a secure, password-protected file at the University of Oxford. Any information that can identify the participants will be deleted when we no longer need it for the research. Only the research team will have access to participants’ data, and it will be labelled with a participant ID number instead of their name.

Data Protection

The University of Oxford is the data controller with respect to participants’ personal data, and as such will determine how participants’ personal data is used in the research.

The University will process participants’ personal data for the purpose of the research outlined above. Research is a task that we perform in the public interest.

Further information about rights with respect to personal data is available from https://compliance.web.ox.ac.uk/individual-rights.

Who has reviewed this research?

This research has received ethics approval from a subcommittee of the University of Oxford Central University Research Ethics Committee.

What if there is a problem or something goes wrong?

For any questions, please contact the researcher via (oxfordccai@cs.ox.ac.uk).

Thank you for reading – please ask any questions.