Abstract
Generative artificial intelligence (AI) such as GPT-4 refers to systems able to understand and generate new coherent and relevant text by learning from existing data sets. The great opportunities that GPT-4 offers are accompanied by great risks. Indeed, the ease of access and use of such a tool also makes it the platform of choice for malicious users. The purpose of this work is to test the machine's capabilities in identifying and reframing so-called Techniques of Neutralization (ToN), rationalizations employed by offenders to justify their deviant behavior. The identification of such theoretical cornerstone of criminology in interviews with offenders is crucial for criminologists, as it provides relevant information on criminodynamics, risk assessment and possible intervention strategies. Our outcomes show a high level of the machine's accuracy in general ToN recognition of Published and Crafted ToN sentences in both Test 1 (precision of 0.82 and recall of 0.75 for "Denial of Injury" in Crafted ToN, precision of 0.93 and recall of 1 for "Absence of ToN" in Published ToN) and Test 2 (precision of 1.00 and recall of 0.83 for "Denial of Injury" in Crafted ToN, precision of 1.00 and recall of 1.00 for "Absence of ToN" in both ToN categories). Regarding the reformulation of sentences to remove ToN (Test 3), the model demonstrates high success rates for most ToN categories and high credibility of the reformulated sentences, indicating its ability to maintain the integrity of the sentences while removing the ToN. Our work analyses the application of the machine with respect to a previously untested construct, with the aim of observing the potential and, above all, the possible pitfalls behind the use of AI models in a hitherto little-explored context such as criminology.