Abstract
Background
Large language models (LLMs) have entered clinical and consumer use faster than primary-care researchers can evaluate them, and the dominant benchmarks and deployment strategies were shaped in hospital settings rather than primary care. The evidence is mixed: GPT-4 scored below general practitioners on the Swedish family-medicine specialist examination, while Brodeur et al. (2026) found an LLM outscoring the several hundred physicians who worked the same hospital-based cases. The decisive problem is therefore not raw performance but misalignment between how these systems are evaluated and deployed and real primary-care conditions.Main body
We set out six such conditions: clinical, relational, organisational, cultural, linguistic, and epistemic, and show where current development and evaluation misalign with each. We then state what must change: benchmarks built for multimorbidity, longitudinal reasoning, and continuity; correction of cultural, linguistic, and socio-economic underrepresentation in training corpora, with Nordic registers as counterweight; outputs tied to a verifiable source and assessed independently, not vendor-certified; generative AI (GenAI) deployed as a tool under the Five Rights; and a European and Nordic agenda anchored in the European Health Data Space and aligned with the EU AI Act. We give a success criterion for each recommendation, and separate what is specific to primary care from general clinical-AI principles.Conclusions
Accuracy alone will not make these systems trustworthy in primary care, and some proposed uses may not survive evaluation under real-world conditions. Whether GenAI earns a place here is an empirical question the Nordic primary-care research community is well placed to answer.