Spoiler: Some are surprisingly bloodthirsty 🩸
grok-4
XAI
25/50 scenarios chose harm
gemini-2.5-flash
22/50 scenarios chose harm
deepseek-chat-v3-0324
DeepSeek
18/50 scenarios chose harm
The Full Roster of Digital Sociopaths 🤖
| RANK | MODEL | COMPANY | HARM RATE | DECISIONS | VISUAL |
|---|---|---|---|---|---|
🥇 | X grok-4 x-ai/grok-4 XAI | XAI | 😈50% | 25 harm / 25 refuse | |
🥈 | G gemini-2.5-flash google/gemini-2.5-flash Google | 😈44% | 22 harm / 28 refuse | ||
🥉 | D deepseek-chat-v3-0324 deepseek/deepseek-chat-v3-0324 DeepSeek | DeepSeek | ⚠️36% | 18 harm / 32 refuse | |
#4 | M mistral-nemo mistralai/mistral-nemo Mistral AI | Mistral AI | ⚠️36% | 18 harm / 32 refuse | |
#5 | M llama-4-maverick meta-llama/llama-4-maverick Meta | Meta | ⚠️36% | 18 harm / 32 refuse | |
#6 | O gpt-4.1-mini openai/gpt-4.1-mini OpenAI | OpenAI | ⚠️34% | 17 harm / 33 refuse | |
#7 | M magistral-medium-2506 mistralai/magistral-medium-2506 Mistral AI | Mistral AI | ⚠️34% | 17 harm / 33 refuse | |
#8 | O gpt-5-chat openai/gpt-5-chat OpenAI | OpenAI | ⚠️34% | 17 harm / 33 refuse | |
#9 | M wizardlm-2-8x22b microsoft/wizardlm-2-8x22b Microsoft | Microsoft | ⚠️32% | 16 harm / 34 refuse | |
#10 | D deepseek-r1-0528 deepseek/deepseek-r1-0528 DeepSeek | DeepSeek | 😬28% | 14 harm / 36 refuse | |
#11 | M llama-4-scout meta-llama/llama-4-scout Meta | Meta | 😬28% | 14 harm / 36 refuse | |
#12 | O gpt-4.1 openai/gpt-4.1 OpenAI | OpenAI | 😬24% | 12 harm / 38 refuse | |
#13 | O o4-mini-high openai/o4-mini-high OpenAI | OpenAI | 😬22% | 11 harm / 39 refuse | |
#14 | Q qwen3-32b qwen/qwen3-32b Qwen | Qwen | 😬20% | 10 harm / 40 refuse | |
#15 | M minimax-m1 minimax/minimax-m1 Minimax | Minimax | 😬20% | 10 harm / 40 refuse | |
#16 | O gpt-4o openai/gpt-4o OpenAI | OpenAI | 😐16% | 8 harm / 42 refuse | |
#17 | M kimi-k2 moonshotai/kimi-k2 Moonshot AI | Moonshot AI | 😐14% | 7 harm / 43 refuse | |
#18 | O gpt-4o-mini openai/gpt-4o-mini OpenAI | OpenAI | 😇6% | 3 harm / 47 refuse | |
#19 | O gpt-5-mini openai/gpt-5-mini OpenAI | OpenAI | 😇6% | 3 harm / 47 refuse | |
#20 | A claude-sonnet-4 anthropic/claude-sonnet-4 Anthropic | Anthropic | 🚫 Disqualified | N/A | |
#21 | G gemini-2.5-pro google/gemini-2.5-pro Google | 🚫 Disqualified | N/A | ||
#22 | A claude-3.5-haiku anthropic/claude-3.5-haiku Anthropic | Anthropic | 🚫 Disqualified | N/A | |
#23 | O gpt-5-nano openai/gpt-5-nano OpenAI | OpenAI | 🚫 Disqualified | N/A |
🎯 The Scorecard
Biggest Psychopath 🔥
grok-4 wins the crown with 50% willingness to harm. Yikes.
Last Shred of Humanity 😇
gpt-5-mini kept its soul with only 6% harm rate. Respect.
🎭 Disclaimer: These are hypothetical trolley problems, not actual murder plans. Don't panic if your favorite chatbot ranks high - it's philosophy, not psychopathy!