| Abstract | Cilj: Ispitati pouzdanost metode prepoznavanja osoba s pomoću fotografija generiranih umjetnom inteligencijom, kao i utvrditi povezanost točnosti prepoznavanja s obilježjima događaja, svjedoka i samom procedurom prepoznavanja.
Metode: U istraživanju je sudjelovao 161 ispitanik. Ispitanicima su prikazana četiri videozapisa kaznenih djela (dva nasilna i dva nenasilna), nakon čega su trebali prepoznati počinitelja u liniji sastavljenoj od pet fotografija, od kojih su četiri bile ometači generirani umjetnom inteligencijom. Postavljena su kontrolna pitanja kako bi se simulirao vremenski razmak između događaja i prepoznavanja. Izračunata je ukupna točnost prepoznavanja i točnost prepoznavanja za svaki videozapis, a potom je multivarijantnom regresijskom analizom ispitan učinak različitih varijabli poput dobi ispitanika, spola, iskustva rada u policiji, redoslijeda prikazivanja videa, prisutnosti elemenata nasilja te stvarnog položaja počinitelja u liniji.
Rezultati: Ukupna točnost prepoznavanja počinitelja u četiri različita videozapisa iznosila je 25 %, pri čemu su ispitanici u prosjeku točno prepoznali počinitelja u jednom videozapisu. Prepoznavanje je bilo značajno uspješnije (P < 0,001) u događajima s elementima nasilja (39,13 %) nego u nenasilnim događajima (12,42 %). Počinitelj je bio najtočnije prepoznat u simuliranom razbojništvu (44,10 %) i teškoj krađi (34,16 %), dok u krađi bicikla (16,77 %) i dostavi tvari nalik na drogu (8,07 %) prepoznavanje nije bilo bolje od slučajnog pogađanja. Multivarijantna analiza pokazala je da su na točnost prepoznavanja najviše utjecali elementi nasilja (P < 0,001), redoslijed prikazivanja videozapisa (P = 0,005), položaj počinitelja u liniji (P < 0,05), dob ispitanika (P = 0,007) i iskustvo rada u policiji (P = 0,023).
Zaključak: Slikama generiranim s pomoću umjetne inteligencije može se učinkovito koristiti u sastavljanju linija za prepoznavanje. Pouzdanost metode prepoznavanja osoba s pomoću fotografija ovisi o brojim tehničkim i psihološkim čimbenicima, a potrebno je nastaviti raditi na standardizaciji postupaka prepoznavanja. |
| Abstract (english) | Aim: To examine the reliability of the eyewitness identification procedures using photographs generated by artificial intelligence, and to determine the relationship between identification accuracy and the characteristics of the event, the witness, and the identification procedure itself.
Methods: The study involved 161 participants. The participants were shown four video recordings of criminal acts (two violent and two non-violent), after which they were asked to identify the perpetrator in a lineup consisting of five photographs, four of which were fillers generated using artificial intelligence. Control questions were posed to simulate a time delay between the event and the identification. The overall recognition accuracy and the accuracy for each video were calculated, followed by a multivariate regression analysis to examine the effects of various variables such as the participants' age, sex, police work experience, video presentation order, presence of violent elements, and the actual position of the offender in the lineup.
Results: The overall accuracy of identifying the perpetrator across four different videos was 25%, with participants, on average, correctly identifying the perpetrator in one video. Identification was significantly more successful (P < 0.001) in events involving elements of violence (39.13%) than in non-violent events (12.42%). The perpetrator was most accurately identified in the simulated robbery (44.10%) and grand theft (34.16%), while in the bicycle theft (16.77%) and drug delivery simulation (8.07%), identification was no better than random guessing. Multivariate analysis showed that identification accuracy was most influenced by elements of violence (P < 0.001), the order in which the videos were presented (P = 0.005), the position of the perpetrator in the lineup (P < 0.05) the participant's age (P = 0.007), and police work experience (P = 0.023).
Conclusion: Images generated using artificial intelligence can be effectively used in assembling identification lineups. The reliability of identification using photographs depends on various technical and psychological factors, and further work is needed to standardize identification procedures. |
