Can AI identify a medicine safely outside the lab?
A review of 91 studies finds rapid progress in pill detection, classification and imprint reading, but little external, prospective or clinical validation. High benchmark accuracy is not evidence that an app prevents medication harm.
The short version
3 key points
- 1.The reviewers searched Scopus, Web of Science, IEEE Xplore and PubMed/MEDLINE for English-language work from 1 January 2016 to 4 September 2026; 895 records yielded 91 included studies.
- 2.The field spans object detection, image classification, OCR and imprint reading, hybrid and multimodal systems, plus few-shot and zero-shot approaches. Dataset and evaluation differences prevent a defensible pooled accuracy estimate.
- 3.External testing, unseen-medication evaluation, prospective deployment and clinical outcomes were limited. A model that names a pictured pill correctly has not thereby proved that it prevents dispensing or administration errors.
The answer: recognition is improving, but safety is still unproven
AI systems can identify pills, read imprints and connect an image with medication information under controlled conditions. The new systematic review shows a technically active field spanning 91 studies. It does not show that a consumer camera app, pharmacy tool or bedside system can yet be trusted as an independent safety check. The strongest evidence remains model performance on particular datasets, not fewer medication errors, fewer adverse drug events or better patient outcomes.
That distinction matters because medicine identification is a high-consequence task. Tablets from different manufacturers may share colour and shape, lighting can obscure an imprint, packaging may be missing, and a correct visual match does not confirm dose, formulation, authenticity, expiry, contraindications or whether the medicine was prescribed for the person holding it. The review's practical message is therefore useful but restrained: recognition can support verification, provided it is linked to validated medicine data, uncertainty is visible and a clinician or pharmacist remains able to check the result.[1]
How the review was conducted
The authors followed PRISMA 2020 and searched four databases—Scopus, Web of Science, IEEE Xplore and PubMed/MEDLINE—for work published between 1 January 2016 and 4 September 2026. They retrieved 895 records, removed 136 duplicates and screened 759 records. After full-text assessment and exclusions, 91 studies remained. The review covered medication detection, classification, optical character recognition, imprint recognition and systems combining images with text, prescriptions, metadata or retrieval tools.
Two reviewers independently assessed the 91 papers across eight domains covering dataset quality, ground truth, data splitting and leakage, evaluation, generalisability, reproducibility, real-world applicability and transparency or safety. Agreement across 728 paired domain judgements was 97.66%, with an aggregate Cohen's kappa of 0.959. That consistency helps show that the rubric was applied similarly, although the authors correctly say their bespoke D1–D8 framework is not a validated certainty-of-evidence scale.[1]
Why headline accuracy numbers do not travel well
Some individual systems report striking benchmark results. The review tables include models trained on anything from a few hundred curated images to millions of images and hundreds of drug classes. They also include different tasks: locating a pill in a scene, naming it, reading an imprint, matching multiple pills, or combining recognition with a medication database. Accuracy, top-k accuracy, mean average precision and OCR scores answer different questions and cannot be arranged as one league table.
Performance can fall when the camera, lighting, background, pill orientation, manufacturer or medication catalogue changes. A class label learned from a development set is especially fragile when the real-world medicine is absent from training. That is why cross-dataset testing and genuinely unseen-medication evaluation matter more than another small increase on a familiar benchmark. The review found those stronger tests remained limited, as did independent multicentre validation and prospective use in clinical or community settings.[1]
What this means for patients, pharmacists and carers
For a patient with sight loss, an older adult managing several prescriptions or a carer sorting medicines, a well-designed recognition tool could make labels and medicine information easier to access. For pharmacists and nurses, it could flag a mismatch for checking rather than replace the professional check. The safer interface is one that says what evidence produced the match, presents plausible alternatives and asks for confirmation when image quality or model confidence is poor.
A wrong answer delivered confidently could be worse than no answer. Consumer tools should not recommend taking an unidentified tablet solely from a photograph, and health services should not count a successful image match as proof of adherence. Deployment also raises privacy and security questions if images, prescriptions or patient identifiers leave the device. Procurement teams need to ask which medicines and manufacturers were tested, how often the model abstains, what happens after an error and whether performance has been measured on the local formulary and devices.[1]
The evidence base has review-level limits too
The review included English-language full-text studies indexed in four databases. Relevant non-English work, unavailable papers and studies framed mainly around pharmacovigilance, safety or Internet-of-Things applications may have been missed. The search was not prospectively registered. Because the 91 studies varied so greatly, the authors used qualitative synthesis and evidence maps rather than a meta-analysis; they did not produce a pooled accuracy, formal publication-bias analysis or quantitative certainty rating.
Many included papers were experimental or proof-of-concept reports, and reporting of privacy, security, interoperability, usability and regulation was inconsistent. The review was supported by Vellore Institute of Technology and reports no external grant or commercial conflict. Those declarations do not remove methodological limitations, but there is no disclosed manufacturer funding that would obviously favour a particular recognition product.[1]
What would change the assessment
Confidence would rise with locked, independent tests covering medicines never seen during training, multiple countries and manufacturers, damaged packaging, difficult lighting, older phones and users with different accessibility needs. Results should report false matches and abstentions, not only average accuracy, and should separate image acquisition failures from classification failures. Public datasets need clear provenance and leakage controls so near-duplicate photographs cannot appear in training and test sets.
The decisive step is a prospective clinical study. Researchers should compare ordinary verification with AI-assisted verification and measure wrong-medication selections, corrected errors, time, workload, user trust and any patient harm. Until such studies show benefit without unacceptable false reassurance, AI medication recognition is best understood as a promising identification layer—not a clinical verdict, prescription check or demonstrated medication-safety intervention.[1]
What this means for people
- Patients and carers may gain accessible medicine information, but should not take an unidentified tablet on the basis of an image match alone.
- Pharmacists and nurses could use recognition as a second check if uncertainty, alternatives and local-formulary coverage are visible.
- Health-service buyers need local external validation and error-handling evidence, not a benchmark accuracy copied from a different dataset.
Global context
The review was led from India but synthesised studies using public and custom datasets from multiple countries. Medication appearance, packaging, regulation and formularies vary internationally, so performance on one country's catalogue cannot be assumed elsewhere. The English-only search may also underrepresent work published in other languages.
What the evidence does not yet show
- The 91 studies were too heterogeneous for a pooled accuracy or direct model ranking.
- External, cross-dataset, unseen-medication, prospective and clinical validation remained limited.
- The review included English-language full text from four databases and was not prospectively registered.
- Its D1–D8 appraisal was structured and reproducible but was not a validated certainty-of-evidence instrument.
- Recognition performance is not evidence of fewer medication errors, better adherence or improved patient outcomes.
What to watch next
- Independent multicentre tests on local formularies and ordinary consumer devices
- Prospective comparisons measuring corrected and introduced medication errors
- Performance on unseen medicines, new manufacturers and poor-quality images
- Clear abstention, uncertainty and human-review behaviour
- Privacy-preserving links to validated prescription and medication databases
Living evidence record
Impact record IAI-05BJXF2
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
15 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Frontiers in Digital Health published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 15 October 2026
Research question, topics and regions
Whether AI medication-recognition systems have been validated beyond development datasets and shown to improve safety in real care
Topics: Medication safety · Computer vision · Optical character recognition · Digital health · Clinical validation
Regions: India · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Watch the background, then check the evidence
The Royal Society
AI-generated illustration · not a video frame
YouTube loads only when you choose. Press play in the player to start. If playback is unavailable, use “Watch on YouTube” below.
Healthcare, culture & human impact · Background explainer
Published 2026-07-01 · 44 min 28 sec
Is this the AI future we want? | The Royal Society
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Source checked 14 October 2026. Check available captions on the publisher’s video page.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Health & Life Sciences
Can synthetic bodies improve burn-size estimates?
BurnAreaNet reduced error on 320,000 rendered views from 5,000 synthetic bodies, but its real-body check used only four scanned models and failed to beat the baseline on the highest-BMI body. It is not ready for burn triage.
7 min · 1 source
Health & Life Sciences
Can infant-motion AI screen neurodevelopmental risk?
DeepAD separated cerebral-palsy-associated movement patterns from normal movement in 1,310 videos collected at three Chinese children’s hospitals. The retrospective result is promising for referral support, but it rests on only 50 cerebral-palsy cases and still needs independent, prospective validation.
8 min · 1 source
Health & Life Sciences
Can AI predict five-year colorectal cancer survival from pathology slides?
A retrospective preprint reports moderate discrimination across a US trial cohort and an external NHS dataset. The model is not ready to guide treatment, and its selected ensemble did not clearly beat the strongest comparator.
8 min · 1 source
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.