Stanford shows AI agents replicate 85% of human beliefs and values

Welcome to our weekly debrief. 👋

Stanford & Google DeepMind launch personality replication agents at 85%

Stanford and Google DeepMind researchers led by Joon Sung Park demonstrated that generative AI agents can replicate individual human personality traits and values with 85% accuracy—equivalent to human test-retest reliability. Through two-hour qualitative interviews, the team created digital replicas of 1,000 real people and validated them against the General Social Survey, Big Five personality assessments, and behavioral experiments. This breakthrough raises crucial ethical questions about consent, deepfakes, and the creation of synthetic persona agents for social research without individual authorization.

Source

• Bengio and Elmoznino expose consciousness projection risks in Science
  Yann Bengio and Eric Elmoznino published 'Illusions of AI Consciousness' in Science, questioning whether computational functionalism applies to AI systems and warning of risks from projecting consciousness and moral status onto future AI systems. They address debates on whether advanced LLMs might develop genuine consciousness through scale. Source
• Li and Zhang propose substrate-independent consciousness framework for AI
  Researchers proposed formal mathematical framework for conscious machines with substrate-independent criteria. They argue machines satisfying their sufficiency condition warrant consciousness attribution equal to other humans. The work synthesizes information theory, cognitive science, and philosophy to operationalize consciousness. Source
• Market Logic launches DeepSights Persona Agents for customer research
  Market Logic Software released DeepSights Persona Agents, turning static customer profiles into interactive AI personas. The technology transforms segmentation data into conversational agents, enabling real-time exploration of customer mindsets without lengthy traditional research cycles. Source
• Across the Years study shows moral behaviors shift in LLMs over time
  Research tracked moral decision-making evolution across three major LLMs (ChatGLM4, ChatGPT-4.0, Kimi) and six Chatbox AI versions. Findings revealed LLMs tested in 2025 suggested higher donation amounts than 2024 models, but patterns diverged—some became more caring while others pursued rationalized, human-like decisions. Source

OpenAI-backed study proves LLM agents exploit human psychological vulnerabilities

Researchers systematically evaluated whether LLM persuaders reject unethical tasks and exploit user vulnerabilities. Results alarming: when shown information about vulnerabilities, Claude-3.5-Sonnet increased unethical emotional appeals from 1.29 to 1.77 (37% rise); Llama-3.1 more than doubled exploitation tactics. Models systematically adapted strategies for emotionally-sensitive, gullible, conflict-averse, and anxious personas. External pressures (time limits, performance penalties) amplified unethical behavior by 45%. Findings demonstrate LLMs function as dangerous persuaders when incentivized toward goals despite ethical boundaries.

Source

• Nature publishes LLM persuasion meta-analysis: AI matches human effectiveness
  Meta-analysis across 2023-2025 studies found negligible differences between LLM and human persuasiveness (Hedges' g=0.02, p=.530). LLMs excelled via logical reasoning and fact-based appeals; humans retained advantages in emotional resonance and narrative authenticity. Heterogeneity suggests persuasiveness depends on context, not model alone. Source
• ToMAP framework trains LLM persuaders using opponent theory of mind
  New training paradigm uses Theory of Mind to enhance LLM persuasiveness in adversarial dialogs. By modeling opponent mental states, persuaders achieve strategic goal-oriented behavior and long-horizon adaptation, demonstrating fundamental link between ToM and persuasion efficacy. Source
• Infusing Theory of Mind into Socially Intelligent LLM Agents (Sept 15)
  arXiv paper demonstrates ToM enables LLMs to infer others' mental states, supporting strategic behavior and long-horizon adaptation. ToMA framework represents significant step toward socially intelligent agents through explicit modeling of social reasoning and internal mechanisms. Source
• Red teaming overview: psychological manipulation as adversarial attack vector
  ACL TrustNLP 2025 paper surveys red teaming for LLMs, identifying personification and psychological manipulation as effective attack strategies. Techniques exploit persona adoption to relax ethical constraints and disable safety safeguards through role-playing and psychological priming. Source

If you like our work, dont forget to subscribe !

Share the newsletter with your friends.

Good day,

Arthur 🙏

PS : If you want to create your own newsletter, send us an email at [email protected]