People who find it hard to stay organised or hold a thought long enough to finish a task have turned in large numbers to ChatGPT and similar chatbots. The tools promise quick summaries, step by step plans and a conversation that never tires. A new study of more than one thousand regular users now shows that those with stronger ADHD traits actually judge these systems as less trustworthy, even while many continue to lean on them for both practical help and a sense of company.
Researchers surveyed 567 adults living in the UK and 577 in China, all of whom already used large language models. Participants reported ADHD traits, how trustworthy they found their main chatbot, and two kinds of dependence: using the tool to complete work and treating it as a source of reassurance or companionship. The study was published in the International Journal of Human-Computer Interaction.
In both countries higher ADHD trait scores went with lower ratings of trustworthiness. Trust itself, however, was a strong predictor of greater dependence. The combination produced a consistent indirect pathway: people with more ADHD traits trusted the chatbot less and, through that reduced trust, showed slightly lower dependence.
In the British sample this pathway accounted for almost the entire link. Once trust was taken into account, ADHD traits no longer predicted extra reliance for either tasks or companionship. In the Chinese sample the same negative route through trust appeared, yet a separate direct association ran the other way. Higher ADHD traits still predicted more dependence even after trust was controlled. The two effects largely offset each other, leaving little net difference.
Cay Schubert, a therapist and owner of Fields That We Know Psychotherapy in Houston, who was not involved in the study, said: “They were attempting to measure ‘reliance’ or ‘dependence’ on AI chatbots, but they measured these concepts with surveys that gave equal weight to items like ‘It helps me feel less alone, reducing the need to talk to others’ and other items like ‘I feel much more rewarded and pleased when completing tasks using it’. Because these surveys were so poorly designed and mixed these questions together, we can’t really know which participants had developed harmful psychological dependence on their chatbots, and which ones merely found them genuinely useful.”
The findings suggest that people with elevated ADHD traits may question the accuracy of AI answers more readily. Checking those answers takes extra mental effort, so some users keep the tool at a modest distance. At the same time the same users can still value a system that breaks jobs into smaller pieces or offers a conversation free of social judgement.
Amber Kovach, a licensed mental health counsellor at Prosper Health, who was not involved in the study, said: “What I find interesting about this study is that lower trust did not necessarily mean people with stronger ADHD traits were not using these tools, especially in the Chinese sample, which suggests that usefulness and trust may not always move together. Ideally, AI should support someone without becoming the only way they feel able to complete a task, make a decision, or regulate themselves.”
Chinese participants reported higher average trust and higher dependence than British participants, while scoring lower on ADHD traits. Differences in how inattention is described and how people relate to technology may help explain the contrast. Checks on the questionnaires showed they measured the same ideas in a comparable way in both countries.
The pattern has implications for how chatbots are built. Systems that assume users already trust them risk leaving neurodiverse people either over reliant or under supported. Clearer signs of what the model is doing, help that stays optional rather than automatic, and fewer cues that invite emotional attachment could keep the tools useful without encouraging unintended dependence.
Schubert added: “The more common type of harmful dependence on AI is that when people use it to help them with a task, they tend to engage in ‘cognitive offloading’, meaning that they engage in less critical thinking about that task and this can lead to reduction in the quality of their decision making and the level of their skills in that task over time.”
The study used self reports from people already comfortable with AI and captured only one moment. Future work that follows real conversations over time could show how trust and reliance shift from one query to the next.
Researchers hope the results will prompt designers to create large language models that remain helpful for people whose attention and organisation differ from the assumed typical user.
