Wakati mifumo ya akili bandia ikiingia kwa kasi katika elimu ya tiba na mazoezi ya kliniki, bado haijulikani vya kutosha kama wanafunzi wanajifunza kutumia zana hizi kwa usalama. Hasa matumizi ya large language models katika kazi kama kutatua kesi za kliniki, kuandika ripoti, kupendekeza utambuzi, kujiandaa kwa mitihani na kuzalisha maandishi ya kisayansi yanaweza kusaidia ujifunzaji wa mwanafunzi, lakini pia yanaweza kuzuia maendeleo ya uwezo wake wa kujenga hoja kwa kujitegemea.

Systematic review hii ilichunguza jinsi jambo linaloitwa “AI dependence” linavyofafanuliwa katika elimu ya tiba, uuguzi, udaktari wa meno, famasia na taaluma nyingine za afya; ni mara ngapi linapimwa kweli, na ni hitimisho gani linaweza kutolewa kwa mtazamo wa usalama wa mgonjwa.

Watafiti walitafuta machapisho ya Kiingereza katika PubMed, Embase, Web of Science Core Collection na Scopus kuanzia 1 Januari 2020 hadi 31 Machi 2026. Mwanzoni kulikuwa na rekodi 14.391, baada ya kuondoa marudio rekodi 10.579 zilichunguzwa, full texts 964 zilitathminiwa na machapisho 269 yakajumuishwa katika review.

Kulingana na uainishaji wa waandishi, pool ya studies ilikuwa na tafiti 192 za empirical zilizotathminiwa kwa MERSQI, randomized studies 16 zilizotathminiwa kama category tofauti, na reviews 61. Machapisho 213, yaani asilimia 79,2, yalichapishwa katika miaka 2025 na 2026 pekee. Hali hii inaonyesha kwamba field inakua kwa kasi sana, lakini sehemu muhimu ya findings bado haijakomaa vya kutosha.

Hitimisho muhimu zaidi la review ni kwamba rhetoric ya risk kuhusu AI dependence iko mbele sana ya actual measurements. Kati ya machapisho 269 yaliyopitiwa, 229, yaani asilimia 85,1, yalirejelea theoretical framework, lakini 13 tu, asilimia 4,8, yalishughulikia AI dependence kwa definition ya wazi na indicator inayoweza kupimwa. Ratio kati ya kutumia theoretical framework na kupima dependence halisi ni takriban 18 kwa 1.

Kwa mfano, skill loss au skill erosion ilipendekezwa kama possible risk katika machapisho 82, lakini ilipimwa directly katika machapisho manne tu. Erosion ya clinical au critical reasoning ilitajwa katika machapisho 137, lakini relevant risk ilitestwa directly katika machapisho 16 tu. Overtrust au automation bias ilijadiliwa katika machapisho 74, lakini ilipimwa katika studies sita tu.

Ingawa patient safety ilijadiliwa katika machapisho 202, kulikuwa na study moja tu iliyopima maoni ya mgonjwa au caregiver kuhusu health professional aliyefundishwa kwa AI. Hivyo gap ya takriban 202 kwa 1 ilitokea kati ya patient-safety rhetoric na patient involvement.

Review inasisitiza kwamba AI dependence si phenomenon moja. Waandishi walitambua constructs tano tofauti: dependence inayotegemea frequency ya matumizi, trust calibration, kuhamisha cognitive load kwa AI, skill loss over time, na tendency ya kuachia algorithm decision chini ya uncertainty. Distinction hii ni muhimu; kwa sababu mwanafunzi kutumia AI mara nyingi pekee hakumaanishi dependence au skill loss.

Overall quality ya evidence ni limited. Kati ya machapisho 269 yaliyopitiwa, 22 tu, yaani asilimia 8,2, yalifikia highest quality category kwa research design yao wenyewe. Katika evidence map yenye themes 15 iliyoundwa na waandishi, hakuna outcome iliyofikia high-confidence level. Themes tano zilitathminiwa kuwa medium confidence na themes kumi low confidence.

Kwa hiyo study haithibitishi kwamba AI kwa uhakika inawafanya medical students kuwa dependent au inapoteza clinical skills. Main result inayoonyesha ni hii: katika medical education, strong statements zinatumika kuhusu AI dependence na patient safety, lakini sehemu kubwa ya statements hizi bado hazijaungwa mkono na longitudinal, behavioral studies zinazopima performance bila AI.

Main problem ya research ni nini?

Clinical AI systems zinatumika katika maeneo mengi, kutoka diagnostic-support tools hadi image analysis, kutoka medical documentation hadi large-language-model-based decision support. Health professionals ambao baadaye watasimamia systems hizi, kuzikataa inapobidi na kubeba responsibility ya outputs zake zenye makosa, wako katika training leo.

Medical student mmoja anaweza kutumia AI only kupata information. Student mwingine anaweza kuachia system entire case analysis, differential diagnosis, clinical rationale au patient note. Matumizi haya mawili si sawa. Ya kwanza yanaweza kuwa appropriate tool use, wakati ya pili yanaweza kuchukua nafasi ya cognitive process ambayo mwanafunzi anapaswa kuikuza.

Kwa patient safety, critical question ni pana kuliko kama mwanafunzi anapata correct answer kwa msaada wa AI:

• Je, mwanafunzi anaweza kutambua AI inapokosea?
• Model recommendation ikipingana na own clinical assessment yake, anaweza kuquestion decision?
• AI isipokuwepo, anaweza kufanya same task independently?
• Anafikia correct outcome kwa clinical rationale gani, na anaweza kuieleza?
• Je, AI support inazuia development ya clinical schemas ambazo bado hazijajengwa?

Main objective ya review ni kuamua kama maswali haya yamepimwa kweli katika existing literature.

Kwa nini “AI dependence” si concept moja?

Watafiti waligundua kwamba expressions kama “overtrust”, “dependence”, “cognitive offloading”, “skill loss”, “automation bias” na “erosion of critical thinking” mara nyingi zinatumika kwa kubadilishana katika literature. Hata hivyo, causes, time scales na measurement methods zake ni tofauti.

Most important aspect ya classification hii ni kutenganisha frequency ya use na real dependence. Health professional anaweza kutumia AI intensively katika tasks fulani na bado akatathmini limitations za system kwa usahihi. Kwa upande mwingine, mwanafunzi anayatumia AI mara chache anaweza kukubali model recommendation bila questioning anapoitumia.

Kwa nini large language models zinaweza kuwa tofauti na previous digital tools?

Calculator, search engine au electronic resources mara nyingi zinasaidia specific functions kama calculation na information access. Large language models zinaweza kutengeneza coherent texts zinazofanana na reasoning ya mwanafunzi mwenyewe. Zinaweza kuunda differential-diagnosis list, kuandika clinical rationale, kuorganize case note na kutoa sources zinazoonekana correct.

Feature hii inaruhusu tool kuingia si katika information-finding stage tu, bali directly katika reasoning stage. Mwanafunzi anaweza kuwasilisha final text; lakini huenda hajajenga mwenyewe clinical schemas zinazohitajika kufikia result hiyo.

Risk inayosisitizwa na review ni kwamba AI output inaweza kufanana formally na independent product ya student. Teacher aki-evaluate final answer pekee, huenda asitambue kama thinking process ya student imebadilishwa na AI.

Vipengele vitano vinavyotofautisha medical education na fields nyingine

Watafiti wanaargue kwamba AI dependence inaweza kuwa more important katika medical na health-professions education kupitia mechanisms tano maalum.

1. Uhusiano wa master-apprentice na tacit knowledge

Clinical education si transfer ya book knowledge pekee. Student anamobserve experienced health professional na over time anajifunza patient communication, dealing with uncertainty, subtle clinical signs na tacit reasoning inayotumika katika decision-making.

AI ikichukua nafasi ya interaction hii kati ya student na educator inaweza kuokoa muda katika short term. Lakini inaweza kupunguza nafasi ya student kuona jinsi expert decision inavyoundwa. Study haithibitishi definitively existence ya mechanism hii, bali ina-classify kama moja ya main risk pathways katika literature.

2. Clinical reasoning ni core competency

Clinical reasoning si kuchagua correct diagnosis pekee. Inahitaji kupima incomplete na uncertain data, kutengeneza alternative explanations, kuexclude serious possibilities, kubadilisha view information mpya inapokuja na kueleza rationale ya decision.

Large language models zinaweza kutoa responses zinazoonekana very coherent. Lakini fluency ya text haiguarantee reliable clinical reasoning underneath. Especially inexperienced students wanaweza kuwa na difficulty distinguishing correct na incorrect outputs.

3. Clinical schemas hujengwa kwa miaka mingi

Conceptual connections zinazojengwa katika early years za medical education zinaweza kuathiri jinsi student atakavyotathmini patients baadaye. Comparison, inference na decision-making processes zinazoachwa continually kwa AI mapema zinaweza kuunda later clinical habits.

Ili kuthibitisha claim ya skill erosion, long-term follow-up studies za years kadhaa zenye AI on na off zinahitajika. Review inaonyesha kwamba studies za aina hii ni extremely rare katika existing literature.

4. Cognitive, psychomotor na affective skills hazitenganishiki

Competence ya health professional inahitaji matumizi ya pamoja ya knowledge, physical practice, patient communication, empathy, professionalism na ethical reasoning. AI inaweza kuongeza exam score pekee bila kuboresha independent performance ya student mbele ya patient.

5. Outcome ni patient safety

Deficiency katika education inaweza kuonekana miaka baadaye katika clinical practice. AI system ikitoa wrong information, physician lazima aitambue, aikatae recommendation na aweze kutoa independent decision.

Katika multicenter study iliyotolewa kama example katika review, general-practitioner candidates waliripotiwa kutambua asilimia 55 tu ya AI hallucinations. Rate hii haiwezi kugeneralizeiwa kwa medical students wote; lakini inaonyesha kwamba trust calibration inapaswa kupimwa directly.

Systematic review ilifanywaje?

Study inaeleza kwamba ilifuata PRISMA 2020 na SWiM reporting approach iliyotengenezwa kwa syntheses zinazofanywa bila meta-analysis. Protocol iliripotiwa kuwa registered katika PROSPERO kabla ya research kuanza.

Databases nne zilitafutwa:

• PubMed
• Embase
• Web of Science Core Collection
• Scopus

Searches ziliunganisha terms za AI na large language models; terms za medical, nursing, dental, pharmacy na health-professions education; pamoja na terms za dependence, cognitive offloading, automation bias, critical thinking, clinical reasoning, patient safety, skill loss na academic integrity.

Only English-language publications zilijumuishwa. Hakukuwa na geographical-region au research-design limitation.

Ni publications zipi zilijumuishwa katika review?

Inclusion criteria zilikuwa broad sana. Watafiti walisema walijumuisha:

• Original empirical studies
• Randomized studies
• Systematic na scoping reviews
• Narrative reviews
• Commentary na opinion papers
• Policy na position papers

Walisema walijumuisha aina hizi za machapisho.

Rationale ya broad approach hii ni kwamba research haikuchunguza only actual effects za AI dependence, bali pia rhetoric inayotumiwa na field. Kwa sababu risk claims zinazojirudia katika opinion papers zinaweza kuathiri education policy, researchers walitaka kuchambua publications hizi pia.

Hata hivyo, hii ina consequence muhimu: statement kwamba “machapisho 171 yalitaja risk ya overdependence” haimaanishi studies 171 zilipata overdependence. Number hii inajumuisha pia commentaries na reviews zilizotaja risk only katika conceptual level.

Screening na data-extraction process

Title, abstract na full-text screening zilifanywa independently na researchers wawili kwa kutumia Rayyan software. Katika uncertain cases, third researcher alifanya evaluation.

Ilielezwa kwamba data zilitolewa kutoka kila publication kwa form yenye fields 140. Fields hizi zinaonekana kujumuisha dependence definition, study design, participants, AI tool, education stage, clinical specialty, measurement tool, patient safety, equity, geography na recommendations.

PRISMA flow katika Kielelezo 1 ni kama ifuatavyo:

Largest share ya full-text exclusions ilitokana na studies kuwa related na medical au health education lakini kutokuwa na dependence focus. Publications 555 ziliexcludeiwa kwa reason hii.

Distribution ya publications zilizochunguzwa

Most publications ni za post-ChatGPT period. Katika Kielelezo 6, 2023–2024 imeclassifyiwa kama “early ChatGPT era” na 2025–2026 kama “mature large language model era”.

Katika early period, asilimia 38 ya studies zilicompare exam au task performance ya tools, wakati katika mature period rate hii ilishuka hadi asilimia 19. Kwa upande mwingine, proportion ya publications zilizojadili integration ya AI katika curriculum iliongezeka kutoka asilimia 62 hadi asilimia 82.

Discussion ya dependence na overtrust ilionekana katika asilimia 40 katika early period na katika 2025–2026 period asilimia 61. Hata hivyo, researchers hawakuhesabu actual measurement rates kwa periods. Kwa hiyo statement ya “gap persists” katika Kielelezo 6 si quantitative comparison kati ya periods, bali qualitative assessment.

Uturuki iko katika nafasi gani katika literature?

Numbers hizi zinategemea primary country au region label ambayo researchers waliassign kwa kila publication. Haiwezi kuhitimishwa kwamba publications zote 18 zilizofanywa Uturuki zilikuwa experimental studies au zilipima dependence directly. Zinaweza kujumuisha surveys, opinion papers au reviews.

Ni health fields zipi zilichunguzwa zaidi?

Kwa sababu field categories hazikuwa mutually exclusive, totals zake zinaweza kuzidi 269. Kwa mfano, research moja inaweza kuwa classified simultaneously katika general medical education, surgery na clinical reasoning.

Methodological quality ya studies iko vipi?

Waandishi walitumia assessment tools tatu kwa different research designs:

• MERSQI kwa empirical medical-education research
• AMSTAR-2 kwa reviews
• Cochrane RoB 2.0 kwa randomized research

Kwa total, publications 22 tu, asilimia 8,2, zilifikia highest quality category kwa design yao wenyewe. Result hii inaonyesha kwamba publication count kubwa inayoonekana kama 269 haimaanishi strong evidence.

Observers walikubaliana kwa kiasi gani katika quality assessment?

Katika initial independent ratings, agreement ilibaki limited:

• MERSQI three-level rating: Cohen κ=0,30
• AMSTAR-2: κ=0,47
• RoB 2.0: Initial agreement ilitafsiriwa kwa limitation kwa sababu outcomes zilikuwa na variability ndogo.

Disagreements zilitathminiwa tena na third researcher Feifei Na. Imeripotiwa kwamba total disagreements 137 zilishughulikiwa, na 78 kati yake zilibaki katika final dataset ya publications 269.

Ikilinganishwa na final consensus score, κ value ya evaluator mmoja kwa AMSTAR-2 ilifikia 0,85. Comparisons nyingine zilibaki roughly katika range ya 0,45–0,51. Kwa hiyo quality categories hazipaswi kuonekana kama completely mechanical na uncontested results.

Gap kati ya risk claims na actual measurements

Main quantitative comparison ya study inaweza kuonyeshwa kwa ratio ifuatayo:

\[ İddia\text{-}ölçüm\ oranı = \frac{Riski\ dile\ getiren\ yayın\ sayısı}{Riski\ doğrudan\ ölçen\ yayın\ sayısı} \]

Five ratios zikipangwa, middle value ni roughly 8,6 times. Kwa maneno mengine, risk inayohusiana na dependence kwa kawaida inatajwa roughly mara tisa zaidi kuliko inavyopimwa directly.

Comparison hii haionyeshi kwamba AI risks si real. Inaonyesha kwamba sehemu muhimu ya risks bado haitegemei validated measurements.

Dependence ilifafanuliwa wazi mara ngapi?

Kati ya publications 269 zilizopitiwa, 13 tu zilifafanua AI dependence kama clear concept yenye measurable indicator. Kwa upande mwingine, publications 229 zilirejelea theoretical framework.

AI-on/AI-off comparison ni especially important. Mwanafunzi kupata higher score wakati anatumia AI inaweza kuonyesha tool ni useful; lakini kama mwanafunzi hawezi kuendelea na same reasoning bila tool, educational objective huenda haijafikiwa fully.

Critical thinking na clinical reasoning zilipimwa kwa kiasi gani?

Table hii inaonyesha kwamba ingawa AI inajadiliwa frequently katika education, kama independent clinical decision ya student inalindwa au la imechunguzwa katika studies chache.

Je, kuna positive outcomes zinazoungwa mkono na evidence?

Ndiyo. Review hailengi risks pekee. Katika evidence map ya waandishi, positive findings zifuatazo zili-classifyiwa katika medium-confidence level:

• Kuongezeka kwa student engagement katika AI-supported tasks
• AI kusaidia reasoning ndani ya explicit pedagogical structure
• Kuongezeka kwa short-term exam au task performance kwa structured AI support

Results hizi zinaonyesha wazi kwamba hakuna evidence inayosema AI haina utility katika education. Main issue ni kwamba short-term supported performance na long-term independent competence si same thing.

Kulingana na narrative synthesis ya review, AI coaching inaweza kutoa benefit zaidi kwa baadhi ya intermediate-performing students, wakati weakest students wanaweza kukosa foundational knowledge ya kutathmini incorrect outputs. Hata hivyo, pooled effect size haikuhesabiwa kwa asymmetric effect according to performance levels.

Evidence map inaonyesha nini?

Kielelezo 4 katika publications 269 kinaonyesha publication counts zinazounga mkono 15 recurring outcomes na adapted GRADE-CERQual confidence assessment ya waandishi.

Five themes zilizowekwa katika medium-confidence level ni:

• Bias na inequity risks kuripotiwa kwa upana
• Hallucination risks kuonekana katika different large language models
• Kuongezeka kwa engagement katika AI-supported tasks
• Reasoning kusaidiwa when explicit pedagogical structure inatumika
• Near-term exam performance kuongezeka kwa structured support

Ten themes zinazohusiana na dependence, skill erosion, reasoning erosion, decision surrender to algorithm, faculty preparedness na equity safeguards zilitathminiwa kuwa low confidence. Hakuna theme iliyofikia high-confidence level.

Bar lengths katika figure si effect size. Zinaonyesha number ya publications zinazotaja au ku-support theme. Publications nyingi kutaja same concern haimaanishi concern hiyo imethibitishwa experimentally.

Patient safety inajadiliwa sana, kwa nini patient view haipo?

Kati ya publications 269 zilizochunguzwa, 202, asilimia 75,1, zilitaja outcomes zinazohusiana na patient safety. Kwa upande mwingine, study moja tu ilipima views za patients au caregivers kuhusu health professional aliyefundishwa kwa AI.

Gap hii ni muhimu. Clinical AI education ikitathminiwa only kwa student exam score au teacher ease of use, patient trust, disclosure na participation katika decision process zinaweza kubaki invisible.

Katika research inayotegemea patient participation, maswali yafuatayo yanaweza kuchunguzwa:

• Je, patients wanataka health professional afichue kwamba anatumia AI?
• Ni clinical tasks zipi wanaona AI use inakubalika?
• Wanaelewaje ni nani ana final responsibility ya decision?
• Wana confidence kiasi gani katika uwezo wa physician aliyefundishwa kwa AI kufanya independent decision?
• AI recommendation ikipingana na physician view, wanatarajia explanations gani?

Review haitoi answers za maswali haya, bali inaonyesha kwamba almost hayajapimwa katika current literature.

Equity na representation zilishughulikiwaje?

Gender au biological sex ilijadiliwa katika publications 137, na language background katika 39. Hata hivyo, number ya publications zilizoripoti outcomes separately by these characteristics ni very low:

• Outcomes by gender au biological sex: publications 7, asilimia 2,6
• Outcomes by language: publications 2, asilimia 0,7
• Outcomes by race au ethnic group: publications 0

Result hii inaonyesha kwamba ingawa concept ya equity inatumika frequently, kama AI support inatoa truly equal outcomes katika different student groups imepimwa mara chache.

Race na ethnicity classifications si sawa across countries. Kwa hiyo zero-publication result haimaanishi kwamba same demographic variables lazima ziwe required katika every country. Hata hivyo, local inequities kama language, socioeconomic conditions, disability, educational infrastructure na digital access zinahitaji kupimwa.

Je, Kielelezo 2 ni real skill-loss curve?

Hapana. Kielelezo 2 kinaonyesha two possible development pathways za independent clinical reasoning kutoka early years za medical education hadi specialist physician level:

• Independent reasoning inayokua gradually kwa traditional practice na supervised learning
• Independent ability inayolimitiwa mapema kutokana na silent AI integration

Figure inaillustrate kwamba AI use inaweza kugeuka kwanza kuwa usage dependence, kisha cognitive offloading wakati wa reasoning, kufupisha master-apprentice relationship na hatimaye surrender kwa model decision.

Lines hizi si longitudinal measurements zilizokusanywa kutoka real students. Waandishi pia wanaeleza wazi kwamba figure inatumia human-centered scenarios badala ya abstract risk curve. Figure inavisualize hypothesis inayopaswa kuchunguzwa.

Proposed education framework ni nini?

Waandishi wanakubali kwamba existing evidence haitoshi kuunda formal na definitive education guideline. Hata hivyo, wanaargue kwamba education institutions haziwezi kusubiri bila precautions wakati AI use inaenea.

Main text ina-highlight design principles nne:

1. Kupanga AI-on na AI-off learning
2. Assessment methods zinazohifadhi thinking effort
3. AI competencies katika levels tofauti
4. Kuongeza student autonomy gradually

AI-on na AI-off learning

AI use haipaswi kuwa banned au unrestricted kwa every task. Education institution inapaswa kuamua mapema ni tasks zipi AI itafundishwa na ni tasks zipi independent competence ya student itapimwa.

Kwa mfano, student anaweza kwanza kutatua clinical case independently, kurecord rationale yake na kisha kucompare na AI recommendation. Katika session nyingine, working with AI support inaweza kufundishwa; katika certain part ya exam, tool inaweza kuwa off.

Assessment inayohifadhi cognitive friction

“Cognitive friction” ni mental effort ambayo mwanafunzi anahitaji kufanya ili kujifunza. Kuondoa every difficulty huenda hakurahisishi learning; katika some cases kunaweza kuondoa learning process yenyewe.

Waandishi wanaeleza kwamba assessment methods zifuatazo zinaweza kufanya thinking process ionekane zaidi:

• Oral examination
• Bedside assessment
• Workplace-based observation
• Kuwasilisha differential-diagnosis rationale hatua kwa hatua
• Kuwasilisha reasoning trace, si final answer pekee
• Kueleza errors katika AI output

Tiered AI competence

Inapendekezwa kwamba all students wawe na basic AI literacy. Juu yake, additional levels zinaweza kujengwa kwa roles tofauti:

• User level: Kuelewa hallucination, bias, data privacy na limits za trust
• Implementation level: Kuweka AI kwa usalama katika clinical workflow
• Developer level: Model development, validation na regulatory-evaluation skills

Progressive autonomy

Kabla ya student kuruhusiwa kufanya specific clinical task kwa AI support, inapendekezwa aonyeshe minimum independent competence kwa same task. Approach hii inafanana na progression kutoka supervised practice kwenda independent practice katika clinical procedures.

Kwa nini Kielelezo 3 hakilingani kikamilifu na main text?

Katika Kielelezo 3, components nne zinaonyeshwa hivi:

• Planned AI-on/AI-off practice
• Assessment inayohifadhi cognitive friction
• Calibration training
• Patient-participatory competence assessment

“Tiered AI competencies” na “progressive autonomy” katika main text hazipo kwa same names katika four main boxes za figure. Kwa upande mwingine, figure ina-highlight trust calibration na patient participation tofauti na four core principles za main text.

Hali hii haiondoi general direction ya framework; lakini katika final version itakayochapishwa, text na figure zinapaswa kutumia same components na same terminology.

Inawezaje kuadaptishwa nchini Uturuki?

Recommendations za review hazipaswi kunakiliwa directly katika medical, dental, pharmacy, nursing na health-sciences education nchini Uturuki. Tasks, legal responsibilities na clinical risks za each profession ni tofauti. Hata hivyo, basic approach inatoa research na policy framework inayoweza kutumika kwa Uturuki.

1. Courses na exams zinaweza kufafanuliwa kama AI-on/AI-off

Kila faculty inaweza kueleza clearly ni tasks zipi students wanaweza kutumia AI. Kwa mfano:

• Literature search inaweza kuwa AI-supported.
• Basic clinical-reasoning exam inaweza kuwa AI-off.
• Task ya kukosoa AI output inaweza kufanywa AI-on.
• Bedside assessment na emergency-decision skills zinaweza kupimwa independently.

2. Si final answer pekee, reasoning trace ya student inaweza kutathminiwa

Katika case solving, student anaweza kuombwa kueleza separately:

• Initial assessment yake
• Differential-diagnosis list
• Findings zinazounga mkono na kudhoofisha kila possibility
• Additional information gani angeomba
• Kwa nini amekubali au kukataa AI recommendation

Vipengele hivi vinaweza kuombwa vielezwe kimoja baada ya kingine.

3. Hallucination na automation-bias training zinaweza kutayarishwa

Students hawapaswi kuonyeshwa correct outputs pekee. Katika some scenarios, AI inaweza intentionally kutoa:

• Wrong diagnosis
• Fabricated source
• Missing drug interaction
• Overconfident lakini wrong clinical explanation
• Health practice ambayo si valid nchini Uturuki

AI inaweza kutoa outputs hizi. Uwezo wa student kutambua error na kujustify decision yake unaweza kupimwa.

4. Self-report surveys pekee hazipaswi kutegemewa

Questions kama “Una trust kubwa kwa AI?” zinaweza kuwa useful; lakini huenda zisionyeshe real behavior. Katika research ya Uturuki, stated trust na actual decisions mbele ya correct na incorrect AI outputs zinapaswa kupimwa together.

5. Longitudinal follow-up inaweza kufanywa

Kuanzia first year ya medical education hadi graduation na specialty period, same students wanaweza kufuatiliwa kwa:

• AI-use patterns
• Independent clinical reasoning
• Hallucination-detection rate
• Ability ya kukataa model recommendation inapobidi
• Patient communication
• AI-off exam performance

Mambo haya yanaweza kufuatiliwa.

6. Patients na caregivers wanaweza kujumuishwa katika research

Nchini Uturuki, expectations na concerns za patients kuhusu clinical AI use zinaweza kupimwa. Education programs zinaweza kuundwa si kwa faculty na student views pekee, bali pia kwa evaluations za watu watakaopokea service.

7. Separate risk map inaweza kutayarishwa kwa kila specialty

Katika radiology, kuona image baada ya AI pre-assessment kunaweza kuleta perceptual dependence. Katika surgery, physical skill na master-apprentice relationship ni more important. Katika nursing, care plan, communication na ethical simulations zinaweza kuwa prominent. Katika pharmacy, drug interactions na counseling; katika dentistry, image interpretation na procedure planning zinahitaji different measurements.

Kwa hiyo, badala ya single “AI policy in health education”, task-specific na specialty-specific rules zinapaswa kutengenezwa.

Strengths za study ni zipi?

• Ilitafuta four major academic databases.
• Ilitumia large initial pool ya records 14.391.
• Titles, abstracts na full texts zilitathminiwa independently na researchers wawili.
• Imeripotiwa kwamba detailed data-extraction form yenye fields 140 ilitumika kwa kila publication.
• Ilijaribu kutenganisha empirical outcomes na claims zinazobaki katika rhetorical level pekee.
• Iligawanya AI dependence katika five distinct concepts.
• Ilitumia different quality-assessment tools kwa each research design.
• Iliripoti pia positive educational outcomes za AI use.
• Iliquantify gap kati ya patient-safety rhetoric na actual patient participation.
• Ilijumuisha publications kutoka country au region labels 52, including Uturuki.
• Iliwasilisha search strategies, PRISMA flow na quality distributions katika appendices.
• Ilieleza explicitly kwamba proposed education framework si proven guideline.

Main limitations za study ni zipi?

• Only English-language publications zilichunguzwa.
• Publications ni highly heterogeneous katika research design, student level, specialty, AI tool na outcome measure.
• Meta-analysis haikufanywa na pooled effect size haikuhesabiwa.
• AI dependence ilifanyiwa explicit operationalization katika publications 13 tu.
• Opinion papers na position texts ziliingizwa katika review.
• Haijaelezwa opinion papers zilitathminiwa kwa quality tool gani.
• Initial quality ratings zilikuwa na low au moderate inter-rater agreement.
• Third researcher aliyesolve quality disagreements pia ni author na project lead wa study.
• 269×140-field extraction dataset bado haijafanywa accessible.
• Study-level MERSQI, AMSTAR-2 na RoB 2.0 scores hazipo katika file.
• Claim na measurement counts haziwezi kureproduceiwa independently.
• Regional counts katika Table 3 na Appendix A5 hazilingani.
• Education-stage categories katika Appendix A5 zinafikia badala ya 269 publications 251.
• Four education principles katika main text na four components katika Kielelezo 3 si sawa.
• Upper date limit ya 31 Machi 2026 haijaonekana explicitly coded katika all database queries.
• Mutual boundaries za categories “empirical”, “randomized” na “review” hazijaelezwa sufficiently.
• Publication counts katika evidence map zinaweza kutafsiriwa kimakosa kama effect sizes.
• Number 202 inayotumika kwa absence ya patient perspective ni number ya publications zinazojadili patient safety; haimaanishi studies 202 zilichunguza directly absence ya patients.
• Asilimia 79,2 ya publications ni za 2025–2026 period; field na tools zinazotumika zinabadilika rapidly.
• Study haijapitia peer review.

Study inaonyesha nini?

• Inaonyesha kwamba AI dependence katika medical na health-professions education inajadiliwa chini ya at least five distinct constructs.
• Inaonyesha kwamba risk claims zinatumika mara 6,2 hadi 20,5 zaidi kuliko direct measurements.
• Inaonyesha kwamba dependence imeoperationalizeiwa explicitly katika asilimia 4,8 tu.
• Inaonyesha kwamba designs zinazocompare AI-on na AI-off performance zimelimitiwa kwa asilimia 7,8.
• Inaonyesha kwamba patient safety inajadiliwa frequently lakini patient na caregiver views almost hazijapimwa.
• Inaonyesha kwamba only asilimia 8,2 ya available studies zimefikia highest quality level kwa design yao.
• Inaonyesha kwamba kuna medium-confidence signals kwamba AI-supported tasks zinaweza kuongeza engagement na near-term performance.
• Inaonyesha kwamba AI risks zinaweza kutofautiana kwa specialty, education stage na task type.

Study haithibitishi nini?

• Haithibitishi kwamba AI kwa ujumla inawafanya medical students kuwa dependent.
• Haionyeshi kwamba AI use kwa uhakika inapunguza clinical reasoning.
• Haithibitishi kwamba long-term skill loss hutokea kwa students.
• Haionyeshi kwamba every frequent AI user anafanya wrong decisions.
• Haionyeshi kwamba exam success katika AI-supported education inabadilika kuwa independent clinical competence.
• Haionyeshi experimentally kwamba proposed curriculum framework inaongeza patient safety.
• Haijaamua AI-on na AI-off tasks zinapaswa kutumiwa kwa frequency gani.
• Haionyeshi kwamba publications zote 18 za Uturuki zilitoa same result.
• Haipimi effect ya kuongeza education assessment katika approval process ya clinical AI devices.
• Haitoi data ya kutosha kuhusu kama patients wana trust physicians waliofundishwa kwa AI.

Mbinu na Matokeo ya Utafiti

Technical method summary

Measurement logic ya five dependence constructs

Main numerical findings

Numerical difference kati ya claim na measurement

Summary ya evidence confidence

Reporting na reproducibility check

Maelezo ya Chanzo na Mbinu

Full original title ya study: AI dependence in medical and health-professions education: a systematic review with implications for clinical AI deployment and patient safety

Authors na order:

1. Hao Wu
2. Yixuan Guo — imeandikwa kama “YiXuan Guo” katika SSRN record.
3. Feifei Na

Equal first authorship au equal contribution: Haijaelezwa.

Corresponding author: Feifei Na, MD, PhD

Institutional affiliation: Department of Oncology, West China Hospital, Sichuan University, Chengdu, Sichuan, China

Institutional-identity warning: Title page ya study inaunganisha authors wote watatu na same institution. Katika SSRN author record, Sichuan University information ipo kwa Hao Wu na Feifei Na, wakati institutional information ya Yixuan Guo haijaingizwa.

DOI:10.2139/ssrn.7001930

Official study link:SSRN record page

Publication platform: SSRN

SSRN publication date: 26 Juni 2026

Page count: 32

Journal: Final version iliyochapishwa au accepted katika peer-reviewed journal haijathibitishwa.

Original publisher: Peer-reviewed journal publisher haijaelezwa. Study ni preprint inayosambazwa kupitia SSRN.

Source type: Systematic review na narrative evidence synthesis kuhusu AI dependence katika medical na health-professions education

Peer-review status: Haijapitia peer review. Kila page ina warning ya “Preprint not peer reviewed”.

Protocol: Study inaripoti PROSPERO number CRD420261390138. Kwa sababu official record haikuweza kuonekana independently, registration status imewasilishwa only kama author statement.

Reporting standards: PRISMA 2020 na SWiM

Funding: Imeripotiwa kwamba hakuna specific funding iliyopokelewa kutoka public, commercial au nonprofit organization yoyote.

Conflict of interest: Authors wameeleza kwamba hakuna conflict of interest.

Ethics committee: Kwa sababu research inategemea systematic review ya published studies, separate ethics approval kwa individual patient au student participation haijaripotiwa.

Data sharing: Imeelezwa kwamba deidentified 269×140-field data-extraction file, PRISMA checklist, excluded studies, quality assessments na third-reviewer rationales zitashirikiwa pamoja na publication katika Mendeley Data record yenye permanent DOI. Materials hizi hazijatolewa katika current version.

AI-use disclosure: Hakuna separate statement kuhusu kama generative AI ilitumika katika preparation ya study.

Main methodological warning: Opinion papers, commentaries na position texts pia zilijumuishwa katika review. Hata hivyo, haijaelezwa publications hizi ziligawanywa vipi katika final design categories na quality tools. Kwa hiyo si sahihi kutathmini all 269 publications kama “empirical evidence”.

Main result warning: Claim/measurement ratios zilizotolewa na study hazipimi prevalence ya AI dependence. Zinalinganisha frequency ambayo risk inatajwa katika literature na frequency ambayo same risk inapimwa directly behaviorally au longitudinally.

Framework warning: Recommendations za AI-on na AI-off learning, assessment inayohifadhi cognitive friction, tiered competence na progressive autonomy si experimentally validated education guidelines. Waandishi wanaziwasilisha kama precautionary safety-control hypotheses.

Tathmini hii ya Kituruki imeandaliwa kwa kuchunguza main text ya study, tables zote, PRISMA flow diagram, dependence classification, evidence map, claim–measurement graph, period comparison, implementation roadmap, detailed database queries, quality-assessment appendix na references. Hakuna scientific result mpya kutoka external sources iliyoongezwa katika findings. External sources zilitumika only kwa bibliographic verification ya SSRN record, DOI, publication date, author display na peer-review status.

Safest conclusion ya study si kwamba AI kwa uhakika inasababisha dependence au skill loss kwa health students. Safest conclusion ni kwamba risks hizi zinatajwa very frequently; lakini outcomes kama independent performance, trust calibration, long-term skill change, patient safety na patient view zinapimwa extremely rarely.