On the 4th of November 2026, Generation AI was joined by Rob Hughes to discuss the building of ParentBench, an open benchmark that puts 100 real parenting questions to 10 AI models and has both AI and professionals score the answers. Dr Anthony Bridgen, project coordinator for Generation AI, reflects on this.
Early childhood represents an important stage of our lives. Before the age of five, the human brain forms connections faster than at any other stage of life, reaching 90% of its final size. As such, this period of development is foundational for children’s ability to acquire more complex abilities later in life. Critical to building this foundation is the existence of stable, secure and nurturing relationships with caregivers. Alongside ensuring children have access to the physical necessities of life, caregivers must support their emotional expression, regulation, and social competence. In the context of societies where intergenerational and community living are deprioritised, caregiver access to embodied knowledge is much reduced. At the same time, other societal constructs which fill these roles, such as schools and the healthcare system have weakened, for example, NHS Health Visitor numbers have declined by 46% in the past decade. Fundamentally, parents and caregivers often experience isolation, anxiety and a pressure to make the ‘correct’ decisions for their children.
It is perhaps unsurprising then, that many caregivers are now turning to AI chatbots to fill this gap. 81% of US adults report using AI tools to help with parenting, including for medical information and behaviour advice. However, whilst we often hear reports on the capabilities of AI at mathematics and coding, we have very little understanding of how good they are at providing appropriate and correct caregiving advice. How then can parents understand which tools they should be using, or if they should be using them at all? Benchmarking is one approach to evaluate models against set ‘best-case’ scenarios but is extremely challenging due to the plurality of what good looks like in this space and the pace at which models change. ParentBench is working do to just this for the caregiving advice large language models provide. By co-creating a set of 100 scenarios, challenging LLMs with these, and evaluating their responses with both LLM and human raters, they have been able to judge model performance in terms of accuracy, helpfulness and safety. Whilst LLMs can answer many questions reasonably well, 1 in 7 responses were flagged as a safety concern by at least one expert. Given that 53% of parents who report using AI for parenting do so for medical advice, this represents considerable concern, an issue highlighted by a current lawsuit against OpenAI alleging that advice provided by ChatGPT around medication dosing led to an individual’s death.
These kinds of evaluations raise important questions around the role we, as a society, want AI to play in parenting. They also pose the challenge of how we can establish a ground truth to evaluate against in the context of scenarios which don’t necessarily have a single, correct answer. Parentbench’s work found that human experts often are not well-aligned with how they rate an LLM’s response to scenarios, perhaps reflecting a diversity in opinions and expertise. It is also vital to understand how caregivers are accessing LLMs, is it directly or through an app, and is there additional context held about the family scenarios which affects helpfulness?
The range of imaginable futures here is vast but to ensure the best possible outcome we should begin thinking about being intentional in use, establishing guardrails, and supporting skill development for caregivers. Above all, we must be cognisant of what this outsourcing of care might do for the relationships between young people and their caregivers.
ParentBench will be formally launched at the Bright Start International Conference on 19-21 of November 2026.
If you are interested in hearing about future seminars, reach out to global.challenges@reuben.ox.ac.uk