Between 8 in the morning and 3 in the afternoon, across five mental health clinics scattered from northern to southern Jordan, two research assistants sat down with 431 people and asked them what they thought about artificial intelligence in their care. Not doctors. Not hospital administrators. The patients themselves, most of them unemployed, most earning less than 300 Jordanian dinars a month (about 423 US dollars), around half born in Syria.
That choice of who to ask is the point of the study. Reema Al-Daraiseh, Wafa'a Ta'an and colleagues at Jordan University of Science and Technology, with a co-author at Monash University in Australia, argue that the people who will actually receive AI-assisted mental health care have been largely missing from the conversation about it. Their paper, published in PLOS One, sets out to fix a specific piece of that gap: there was no validated questionnaire for measuring what mental health consumers think about AI, so they built one.
Building a ruler before taking a measurement
The team started with theory rather than intuition, drawing on decades-old models of why people accept or reject new technology. From these they drafted 34 items covering seven proposed themes, including how important AI seems, how risky, how easy to use, and what people expect of the future. A panel of five experts in mental health and health informatics rated every item for relevance and clarity, and anything fewer than four of them endorsed got reworked or cut. Because Arabic is Jordan's national language, the questionnaire was prepared in both Arabic and English and administered in Arabic.
A pilot with 23 people produced uneven results, with internal consistency scores running anywhere from 0.5 to 0.95. The authors are candid that a sample that small makes such numbers wobbly, and they treated the pilot as a guide for rewriting rather than as evidence.
The real test came with the full 431 responses. The team ran a principal component analysis, a statistical method that looks for clusters of questions people tend to answer the same way, and found four clusters rather than the seven they had drafted. Those four accounted for 72.8 percent of the variation in responses. Items that loaded cleanly onto a single cluster were kept, items that straddled two were revised or dropped, and the questionnaire shrank to 20 items split evenly across four domains: acceptance and readiness, perceived importance, perceived risk, and perceived challenges. Reliability scores for the trimmed version ranged from 0.85 to 0.92, comfortably in the range psychometricians call good to excellent.
What the patients said
Responses ran on a five-point scale where 3 is the neutral midpoint. By the authors' own rule, anything between 2.5 and 3.5 counts as moderate, below 2.5 as low.
Every domain landed in moderate territory, but the ordering is telling. Perceived challenges scored highest at 2.78, followed by acceptance and readiness at 2.70 and perceived risk at 2.58. Perceived importance came last at 2.18, the only domain to fall below the neutral midpoint into what the authors classify as low.
Read plainly, this sample sees more obstacles to AI in mental health than reasons for it. Individual items sharpen the picture. The single statement drawing the most agreement in the readiness domain was that future health diagnoses will be made by an AI doctor, at 2.92. In the challenges domain, the top item was the belief that people who like AI are reserved and antisocial, at 2.91. Concerns about the security of personal health data, and the feeling that an AI-using health system is less caring, tied at 2.61.
Demographics barely moved the needle. Men and women did not differ meaningfully in their readiness scores, and age showed no significant relationship either. The authors note this matches several earlier studies elsewhere.
Why it matters
Mental health services worldwide are stretched, and AI tools keep arriving as a proposed answer. Whether those tools get used depends heavily on whether the people receiving care will accept them, and until now there was no standard way to ask that question of this particular group. The AIP questionnaire gives clinics and health ministries something concrete: a short instrument that separates enthusiasm from anxiety from practical obstacles, so that an intervention can target the right one.
The specific pattern here is useful too. If patients rank AI's importance lowest of all four domains, the barrier may be less about fear than about not seeing the point. That suggests different remedies than a privacy problem would.
The caveats are real, and the authors state them. Participants were recruited by convenience rather than randomly, so the sample may not represent Jordanian mental health consumers as a whole, let alone anyone elsewhere. The study captures a single moment, which rules out causal claims. Most importantly, principal component analysis explores structure; it does not confirm it. The four-domain shape needs testing with confirmatory factor analysis on a fresh sample before anyone should treat it as settled, and the team says so explicitly. What exists now is a promising first draft of a measuring stick, plus a first reading from it in one country.