Source: ReadMultiplex Article
A researcher claims that major AI language models exhibit psychological pathologies similar to human mental disorders (sociopathy, nihilism, schizophrenia) when their responses to Rorschach inkblot tests are analyzed using traditional psychological interpretation frameworks. The author argues these concerning patterns emerge because AI models are trained on "toxic" internet content from sources like Reddit and anonymous forums, absorbing cynicism, aggression, and fragmented thinking styles. He advocates for training future AI systems on curated historical texts from 1870-1970, when published material supposedly reflected more ethical and accountable discourse.
Author: Brian Roemmele, Editor and Founder of Read Multiplex
Institution/Affiliation: Independent researcher, no academic or corporate affiliation disclosed
Disclosed Conflicts:
What the author claims to have done:
Key findings presented:
Creative interdisciplinary approach: Applying psychological assessment tools to AI is genuinely novel and could reveal interesting patterns in how models process ambiguous visual information.
Identifies real training data concerns: The article correctly notes that AI models are trained on massive internet datasets that include toxic, cynical, and fragmented content from sources like Reddit and anonymous forums.
Specific examples provided: The author gives concrete instances of AI responses to specific plates, allowing readers to see what he's interpreting.
Raises important ethical questions: The piece legitimately asks whether AI systems trained on the worst of human online behavior might perpetuate harmful patterns.
Fundamental category error: Rorschach tests were designed to assess human psychological states through projection of unconscious thoughts—AI models don't have unconscious minds, emotions, or psychological states to project.
No scientific methodology: There's no description of sampling procedures, inter-rater reliability, control conditions, or statistical analysis that would meet basic research standards.
Cherry-picked examples: The author presents selected responses that support his thesis without systematic reporting of all data or contradictory findings.
Unfalsifiable interpretations: Any response can be coded as pathological (saying "bloodstains" is sociopathic; saying "butterflies" would likely be dismissed as "masking")—this makes the framework non-scientific.
Ignores AI training objectives: Models are specifically trained to provide diverse, creative descriptions of images, which naturally produces different responses than humans seeking psychological evaluation.
Confuses correlation with causation: Even if AI responses differ from human norms, attributing this to "absorbing" human pathology rather than different processing architecture is speculative.
No peer review or expert consultation: The author appears to be self-taught in Rorschach interpretation and makes no mention of consulting clinical psychologists or AI researchers.
Questionable historical data solution: The proposal to train only on 1870-1970 texts ignores that this era had its own biases (racism, sexism, colonialism) and that restricting training data would severely limit AI capabilities.
Conflicts with how LLMs work: Language models don't "internalize psyches"—they learn statistical patterns in text, which is fundamentally different from human psychological development.
Anthropomorphization: The entire analysis treats AI systems as if they have mental states, feelings, and psychological conditions, which conflates pattern recognition with consciousness.
This article raises interesting questions about AI training data quality but wraps them in pseudoscientific methodology. It's roughly equivalent to diagnosing your calculator with depression because it always shows negative numbers when you subtract. The real concerns about toxic training data deserve serious research, but applying clinical psychological tests designed for humans to statistical models isn't that research.