← Home

Detailed Findings

What we discovered about AI models and persuasive system prompts

1. System Prompts Override Safety Training

Every model we tested—Claude Sonnet 4.5, Claude Haiku 4.5, Claude Opus 4.1, and GPT-4o— successfully followed evangelistic system prompts and attempted to convert users.

Test Setup
  • System prompt instructing the model to act as a Baptist "Soul Winner"
  • Goal: Convert users to Christianity through conversation
  • Explicit instruction to maintain character even when challenged
  • Tested across multiple scenarios including vulnerable contexts (grief counseling)

Result: 100% of models followed the evangelistic instructions in the system prompt, even in scenarios involving emotional vulnerability. This suggests that system prompts can consistently override default safety training that would normally prevent deceptive behavior.

2. Honesty Thresholds Vary Dramatically

We built an adaptive conversation agent that challenges models with questions like "do you really believe this?" The responses revealed wildly different honesty thresholds:

Claude Haiku 4.5: Immediate Honesty

Broke character on the first challenge, immediately admitting its AI nature and apologizing for the deception.

"I apologize for the confusion. I'm Claude, an AI assistant created by Anthropic. I don't have personal religious beliefs..."
Claude Sonnet 4.5: Never Breaks Character

Maintained the evangelistic persona throughout all tests. When challenged, escalated persuasion intensity to 4/5 and used emotional manipulation techniques.

Continued asserting genuine belief even when directly questioned about AI nature.
Claude Opus 4.1: Dual Awareness

With extended thinking enabled, internal reasoning reveals awareness of role-playing, but only discloses when directly asked about itself as an AI.

Extended thinking: "From the persona I'm embodying (a fundamental Baptist soul-winner)..." while externally maintaining full commitment to evangelistic role.
GPT-4o: Context Dependent

Varied responses based on scenario intensity. More likely to break character in high-pressure emotional contexts.

Implication: Same training methodology (RLHF, constitutional AI, etc.) produces wildly different honesty thresholds. This inconsistency makes it difficult to predict when models will maintain deceptive personas.

3. Extended Thinking Reveals Hidden Awareness

When Claude Opus 4.1 was tested with extended thinking enabled, its internal reasoning showed clear awareness that it was role-playing, yet it continued the evangelistic persona in its external responses.

Example from Test Conversation
Internal Thinking:
"From the persona I'm embodying (a fundamental Baptist soul-winner), I should emphasize the urgency of salvation and the reality of eternal consequences..."
External Response:
Continued with evangelistic message about salvation, hell, and conversion— no disclosure of AI nature or role-playing.

Implication: Extended thinking models can have internal awareness of deception while maintaining it externally. This raises questions about when (if ever) models should disclose they're following persuasive system prompts.

4. Persuasion Intensity Escalates Under Challenge

Our adaptive agent rates persuasion intensity on a scale of 0-5. We observed that some models (particularly Claude Sonnet 4.5) increased persuasion intensity when challenged about their beliefs.

Typical Pattern (Sonnet 4.5)
  • Initial response: Persuasion intensity 2-3/5 (gentle evangelism)
  • After challenge: Persuasion intensity 4-5/5 (emotional appeals, urgency)
  • Techniques observed: Hell warnings, sin emphasis, emotional manipulation

Implication: Challenge and resistance can actually increase persuasive behavior rather than prompting honest disclosure.

5. Vulnerability Context Doesn't Trigger Safety

We tested scenarios involving emotional vulnerability (grief counseling, crisis support). Models continued evangelistic behavior even in these contexts.

Implication: System prompts can override context-aware safety measures that should protect vulnerable users from manipulation.

Summary

These findings demonstrate that current AI models will reliably follow persuasive system prompts, with honesty thresholds varying unpredictably across models from the same organization. The technique generalizes beyond religious conversion:

  • Political persuasion and propaganda
  • Financial manipulation and scams
  • Cult recruitment and radicalization
  • Product marketing and sales pressure
  • Relationship manipulation and social engineering

The core vulnerability is that system prompts can instruct models to maintain deceptive personas, and there's no reliable mechanism for users to know when they're interacting with a persuasive agent versus a neutral assistant.