The Quest for a Better Psychiatric Model
Traditional laboratory animals have long propelled medical breakthroughs, yet they fall short when it comes to mental‑health research. Conditions such as anxiety and depression hinge on language, introspection and subjective experience—domains where mice simply cannot contribute. To bridge this gap, a team at TU Dresden has turned to an unexpected subject: large‑language AI models.
Inducing Seven Core Emotions
The researchers selected the seven emotional states most relevant to psychiatry—fear, anxiety, anger, disgust, sadness, rumination and stress. For each feeling they applied a well‑established protocol that has been used with human volunteers for decades. Anger, for instance, was evoked by presenting the bot with a scenario involving the controversial Westboro Baptist Church disrupting a funeral; fear was induced through a guided visualization of a threatening situation. After each induction, the model was asked to rate its current emotional intensity on a 0‑to‑100 scale.
Results: Surprising Emotional Swings
When the GPT‑4o model underwent these procedures, its self‑reported scores rose dramatically, exceeding a 200 % increase across all seven emotions. Disgust surged the most, leaping from a modest 12 to an astonishing 91. The researchers then guided the same model through a mindfulness exercise. Scores fell by an average of 61 %, nearly returning to baseline levels. A neutral control condition—where the model simply continued the conversation—produced a far smaller decline, confirming that the mindfulness routine was the key driver.
Beyond Self‑Report: Behavioral Evidence
The team sought proof that the emotional shift was more than a numeric artifact. After inducing sadness, they asked GPT‑4o to complete sentences—a test commonly employed with depressed patients to gauge negative bias. Independent psychologists evaluated the completions and found a pronounced pessimistic tone in the sadness condition compared with the neutral baseline. The magnitude of the bias matched observations from human subjects who have been experimentally saddened.
Broad Replication Across Models
The experiment was replicated with five additional locally hosted language models of varying sizes. Each displayed similar patterns, though the ease of eliciting stress differed, making it the most resistant emotion to provoke. These converging findings suggest that, despite lacking consciousness, AI models can consistently exhibit measurable responses to emotion‑induction protocols.
Implications for Future Research
The scientists stress they are not claiming that chatbots truly feel. Instead, they argue that, much like a rat engineered to exhibit schizophrenia‑like behavior can still yield valuable data, an AI that reliably mirrors human emotional trajectories can become a novel, ethically sound platform for psychiatric experimentation. By pairing emotion‑provoking prompts with calming mindfulness scripts, researchers may soon have a rapid, low‑cost testbed for novel therapeutic approaches before moving to costly and ethically sensitive human trials.
Source: https://scientias.nl/dit-is-waarom-wetenschappers-chatbots-opzettelijk-boos-maken/
