Malingering depression: a comparative study of human and GPT-3.5 performance

Malingering is a well-known phenomenon which has strong social implications. A person often adopts it during psychological evaluation and is particularly difficult to detect as the malingerer’s ability to simulate increases. The public primarily uses Large Language Models (LLMs) to educate themselves and gather information on various topics. Hence, it is crucial to assess whether LLMs can be a source for people learning how to deceive. The present study tested this issue by investigating whether ChatGPT 3.5 is able to malinger depression, one of the easiest and most commonly feigned conditions. We therefore collected 228 responses from humans and ChatGPT 3.5 (114 human participants vs ChatGPT responses) to the Minnesota Multiphasic Personality Inventory-2 (MMPI-2) with the instructions of either malingering depression or being honest. Our results showed that both humans and ChatGPT were able to malinger to suffer from depressive symptoms while illing the MMPI-2, although the validity scales of the MMPI-2 allowed us to identify malingered responses for both groups (humans and ChatGPT). In addition, humans and ChatGPT performance signi icantly differed.or people learning how to deceive without being caught since LLMs have shown to hold internal representation of many concepts. The present study tested this research question by investigating whether ChatGPT (i.e., 3.5 turbo) is able to malinger depression, one of the easiest and most commonly feigned conditions in different contexts (i.e., forensic, clinical, workplace). We, therefore, collected 228 responses from humans and ChatGPT (i.e., 114 humans’ responses and 114 ChatGPT responses) or people learning how to deceive without being caught since LLMs have shown to hold internal representation of many concepts. The present study tested this research question by investigating whether ChatGPT (i.e., 3.5 turbo) is able to malinger depression, one of the easiest and most commonly feigned conditions in different contexts (i.e., forensic, clinical, workplace). We, therefore, collected 228 responses from humans and ChatGPT (i.e., 114 humans’ responses and 114 ChatGPT responses)or people learning how to deceive without being caught since LLMs have shown to hold internal representation of many concepts. The present study tested this research question by investigating whether ChatGPT (i.e., 3.5 turbo) is able to malinger depression, one of the easiest and most commonly feigned conditions in different contexts (i.e., forensic, clinical, workplace). We, therefore, collected 228 responses from humans and ChatGPT (i.e., 114 humans’ responses and 114 ChatGPT responses).