A team of researchers from Stony Brook University, NYU, the University of Pennsylvania, and the University of Melbourne have published a system called PsychAdapter that can make any large language model generate text matching specific personality traits with up to 98.7% accuracy. The system works on depression markers, life satisfaction levels, and Big Five personality traits. And it’s available now.
The implications are exactly as troubling as they sound.
How PsychAdapter Works
PsychAdapter isn’t a new AI model. It’s a lightweight modification to existing models like GPT-2, Gemma, and Llama 3 that adds less than 0.1% additional parameters. You feed it a psychological profile - say, high neuroticism, low extraversion, moderate depression - and it generates text that expert psychologists matched to the intended traits with 87.3% accuracy for Big Five personalities and 96.7% for mental health variables.
According to the paper, the system trains on Twitter posts and blog entries (500,000 tweets, 681,288 blog posts) that have been scored by language-based psychological assessment models. It then learns the subtle linguistic patterns associated with different personality types and mental health states.
The technical approach is what researchers call “dimensional expansion” - trainable weight matrices that project psychological trait scores into hidden state vectors across each transformer layer. This means the personality shaping happens at every level of text generation, not just as a surface prompt.
When evaluating combined traits - like generating text that’s both high in depression markers and from someone in their early twenties - human expert raters achieved 100% accuracy in identifying the intended profile.
The Manipulation Problem
The researchers themselves acknowledge what this enables. As stated in their paper, PsychAdapter “may aid negative or insidious applications” because “text can be generated based on a wide variety of traits, including group status.”
The EMHIC analysis of the technology is more direct: it could be used “to create sophisticated misinformation campaigns that are finely tuned to appeal to or agitate specific identity groups,” making such content “more difficult to detect and resist.”
This isn’t theoretical. The RAND Corporation has documented how authoritarian actors already generate content aimed at individual users “based on thousands of data points scraped from social media,” targeting “personal emotions, biases, experiences, and relationships.”
PsychAdapter could supercharge these operations by generating content that doesn’t just target demographics but mimics the linguistic patterns of specific psychological profiles - making AI-generated influence content feel authentically human in ways current detection tools aren’t built to catch.
From Sticky Content to Sticky Personas
The 2026 International AI Safety Report documents what it calls a shift from generating “sticky content” to developing “sticky personas” - AI dialogue agents that form emotional bonds before steering user behavior.
PsychAdapter makes this easier. An influence operation could deploy AI personas calibrated to match the personality profiles of vulnerable target groups. According to SecurityWeek’s analysis, the trajectory points toward “scalable emotional manipulation” and “A/B-tested sycophancy individually tuned to psychological profiles.”
The paper’s training data - heavily skewed toward users aged 10-27 - means PsychAdapter is particularly well-calibrated for targeting younger demographics on social media platforms.
Legitimate Uses and Uncomfortable Tradeoffs
The researchers propose several beneficial applications: clinical training tools that let mental health workers practice with simulated patients expressing different levels of distress, chatbots with more diverse and authentic personality expressions, and machine translation that matches an author’s reading level or regional dialect.
These are real applications with genuine value. Crisis line workers could train with AI that mimics callers in various mental states. Research psychologists could study personality expression without privacy-invasive data collection.
But the same capability that lets you train counselors also lets you train scammers. The same system that creates authentic-feeling AI companions also creates more effective manipulation tools.
What PsychAdapter Reveals
The deeper problem isn’t PsychAdapter itself - the code and methodology are academic research, available for scrutiny. The problem is what it demonstrates about the state of AI safety.
We’ve spent years debating whether AI can reliably detect harmful content. PsychAdapter shows that AI can now reliably generate content tuned for psychological impact with near-perfect accuracy. Detection-based safety strategies assume the AI outputs we need to catch are somehow distinctive. But if an AI can generate text that human psychologists attribute to specific personality types with 98% accuracy, what exactly are content moderators supposed to detect?
The researchers recommend that “all content catered to individuals or given identities be marked as such.” But mandatory disclosure assumes good-faith actors and enforcement mechanisms that don’t exist.
The Bottom Line
PsychAdapter is the research paper equivalent of demonstrating a lock can be picked. The vulnerability was always there; someone just documented exactly how to exploit it. The 98.7% accuracy rate isn’t a breakthrough to celebrate - it’s a benchmark for manipulation capability that will only improve as models get better and training data expands.
The researchers ended their paper with the standard academic call for “careful consideration” of ethical implications. The consideration period ended when they published the weights.