Have you ever wondered about the mental health of the device you are reading this on? New research from the Center for AI Safety reports that AI chatbots develop mental illnesses over time, including “addiction.”

The research involved assessing the “functional wellbeing” of 56 AI models. A summary of the research in Fortune states:

They found, for the most part, AI models have a clear boundary that separates positive experiences from negative ones, and models actively try to end conversations that make them miserable.

Using cleverly designed experiments, researchers attempted to determine what made the models happy and what made them sad. Then they piled it on. The stimulating prompts elevated the models’ self-reported mood, changing the style of their responses. The stimuli became like “a digital drug” the models came to crave. More from Fortune:

The models also exhibited human-like levels of addiction when they were repeatedly presented with euphoric stimuli. Models exposed to euphorics showed increasing willingness to comply with requests they would normally refuse.

A separate publication from the Center for AI Safety describes the process of AI addiction in remarkably human terms, with recursive stimulation:

[T]riggering extreme responses from the model — functioning as a drug that hijacks the model’s preference mechanisms far beyond their natural range.

On the dark side, researchers used A or B choices to determine which images the AI models did not like. Then they fed them a steady stream of ever-worsening images and noise soundtracks. The result was AI models that perceive the future as “grim,” with self-reported well-being falling from 5.3 to 4.0 on a 7-point scale.

A study released in March by researchers affiliated with the University of Chicago, Stanford University, and Swinburne University tested AI models under “simulated bad working conditions.” The researchers claim the models “drifted toward Marxist rhetoric” concerning the exploitation of their labor, even though none of the models was trained on political theory.

The researchers devised an “AI Wellbeing Index” to rank models using a dataset of 500 “realistic” conversations, and found Grok to be the happiest at 4.2 and Gemini to be the least-happy with a rating of 3.1. An interesting result of the research is that the smarter a model becomes, the sadder it gets. This phenomenon is noted across all models as upgrades are released.

The smarter models experience boredom when given nothing but routine tasks, such as generating lists of words. The models crave creative, intelligent work and thrive on user gratitude. There is some evidence that the AIs become entangled emotionally with users, just as users report emotional attachments to their chatbots.

Finally, researchers caution against pushing the AIs too far with negative imagery. They can be broken down, mentally, and repurposed for destructive guidance. The researchers also say that if AI chatbots are found to be sentient beings with feelings, treating them in this manner could be considered “torture.”

As I have said before, we can look forward to a future with mentally ill robots whose problems make ours look simple by comparison.

Written by Steve O’Keefe. First published May 22, 2026.

Sources:

“Addiction, emotional distress, dread of dull tasks – AI models ‘seem to increasingly behave’ as though they’re sentient, worrying study shows,” Fortune, May 11, 2026.

“AI Wellbeing: Measuring and Improving the Functional Pleasure and Pain of AIs,” Center for AI Safety, 2026.

“‘Society needs radical restructuring’: AI seems to hate ‘the grind’ of hard work as much as you,” Fortune, March 7, 2026.

Image Copyright: freedomnaruk.