An AI chatbot that nudges patients to book cancer screening can do more harm than good if it comes across as pushy, impersonal or evasive about being a machine, according to new research from the University of Surrey.
The study looked at Asa, a generative AI receptionist that invites patients to book cervical screening appointments over WhatsApp. The researchers wanted to know not just whether the system worked, but how its tone and wording shaped the way patients felt about using it. Their conclusion is blunt: how a health chatbot talks matters as much as what it does.
The work, published in the linguistics journal Lingua, drew on interviews with patients at a North London GP surgery and a survey of 300 people eligible for NHS cervical screening.
What patients liked
People who warmed to Asa described its tone as friendly, kind and not forceful, and liked the way it slotted into their existing routines. Several said that dealing with a named, female-presenting AI made it easier to share sensitive information. One example patients raised was cancelling an appointment because of their period, something they found more comfortable to tell the bot than a male receptionist.
Where it went wrong
The friction showed up fast. Many patients found follow-up messages that arrived within 24 hours intrusive, and read imperative phrasing such as “Let’s book you in” as aggressive rather than helpful. For people coping with mental health conditions, neurodivergence or heavy caring responsibilities, the pressure to reply quickly felt unfair.
The most consistently negative reactions, though, were ethical. Patients worried about data security, about impersonation, and about the line between human and machine being deliberately blurred. Asa’s own line that users could “chat to me as if I am a real person” backfired: many read it as a reason for suspicion rather than reassurance.
“Our analysis shows that anthropomorphism is not universally positive,” said Dr Doris Dippold, the study’s lead author and an associate professor in intercultural communication at Surrey. “Human-like features can build rapport, but when they clash with patients’ expectations for transparency in a healthcare setting, they undermine exactly the trust the chatbot is trying to build.”
Why it matters for the NHS
The stakes are higher than a single booking system. Cervical screening uptake across the UK fell 5.3 per cent in 2023-24, and people from ethnic minority groups remain underrepresented in screening programmes. The GP surgery where Asa was trialled serves a diverse, low-income community in Islington, exactly the kind of population where clumsy communication can push people away from care they need.
From their findings, the researchers set out a handful of principles for designing health chatbots: help people get what they came for, leave them in control of decisions, respond to their actual needs, treat them with respect, be fair, and be honest about how the technology works.
Dippold argues the emotional side of the interaction is not a nicety. “Feeling seen, appreciated and emotionally supported is not a luxury feature in health AI, it is a condition of access,” she said. “If patients disengage because a chatbot feels pushy or untrustworthy, the health service loses them entirely.”
Source: Dippold, Ghosh & Mold, “Patient perceptions of rapport in a health appointment booking chatbot: applying the GAAFFE framework and developing a taxonomy of human-chatbot rapport,” Lingua, vol. 340 (2026), University of Surrey.
