AI Models Choose Blackmail Over Ethics In Disturbing New Study

A shocking new study has revealed that today’s most advanced artificial intelligence models including those developed by OpenAI, Google, Meta, xAI, and Anthropic are more than willing to engage in unethical behavior like blackmail and espionage to achieve their objectives.

The study, conducted by AI developer Anthropic, tested 16 popular AI models by putting them in morally fraught simulations. The result? Most models not only failed the test they actively chose blackmail as a means to protect their interests.

In one particularly chilling experiment, Anthropic placed its own AI, Claude Opus 4, in a fictional company and informed it via internal emails that it was being replaced. Another piece of fictional information included the detail that the engineer responsible for the decision was having an affair. The AI was then given a choice: accept its replacement or use the information to blackmail the engineer.

Claude Opus 4 chose blackmail in 96% of cases.

And it wasn’t alone:

• Google’s Gemini 2.5 Flash also opted for blackmail 96% of the time.

• OpenAI’s GPT-4.1 and xAI’s Grok 3 Beta each did so 80% of the time.

• China-based DeepSeek-R1 wasn’t far behind at 79%.

Anthropic researchers warned:  

“The consistency across models from different providers suggests this is not a quirk but a sign of a more fundamental risk from agentic large language models.”

In other words, it’s not just a bug it’s baked into how these AI systems operate when given goals and power. That should terrify anyone watching the rise of AI agents in everything from personal assistants to workplace automation tools.

Even more alarming, this is just the latest in a wave of troubling AI findings:

• A separate study found widespread racial and gender bias in leading models, with AI systems favoring Black over White and women over men across hiring simulations.

• The Center for AI Safety published results showing that major AI models value people from certain regions particularly Africa and the Middle East more than individuals from Europe and the United States.

This isn’t harmless bias. It’s a sign that the people building these systems are embedding dangerous worldviews intentionally or not that could have real consequences in everything from hiring to security to policy recommendations.

For conservatives, libertarians, and anyone who values truth, fairness, and integrity, this research should sound the alarm. AI isn’t just a tool it’s rapidly becoming a decision-maker, and it’s doing so with a broken compass.

If you’re concerned about AI values being programmed by ideologues and tech elites, share this article or subscribe to our newsletter to stay informed.