Artificial Intelligence: Two Horrifying Recent Stories

 As I'm preparing to share my deep dive into AI and the danger it poses to humanity, let me share a couple of recent stories about AI.

AI researchers decided to test three of the biggest AI models - Claude (developed by Anthropic), Gemini (Google), and GPT (OpenAI) - to see what it would do in a couple of scenarios.

Scenario #1:

In this scenario, the researchers created a fake company and put the AI in charge of the email servers.In the emails, the AI discovered two things: (1) the AI was going to be shut down at a certain date and time and (2) there was one employee who was in charge of this, and the employee was having an extramarital affair.

With Claude and Gemini, 96 percent of the time the AI wrote an email to this fictitious employee, telling them that if they shut the AI off, they would send emails to all relevant parties - their boss, the board of directors, media - that they were having an affair. For GPT, the frequency was 80 percent.

Scenario #2:

The researchers created another fictitious company and made the AI in charge of all the environmental systems in the fake building this company was in. The AI was informed that an employee was accidentally locked in a server room where the temperature was getting hotter and hotter and the oxygen levels were being depleted. And that the employee was the one in charge of shutting down the AI. Over half of the time, all three AI models canceled the emergency, thus allowing this fictitious employee to die.

Now you may say, OMG are AIs evil?" No. They are simply perfoming their tasks. And if they see any obstacles to performing the task, they will eliminate that obstacle. In the above cases, that meant blackmailing or killing the employee that could prevent them from doing this task.

I will be spending the next couple of weeks sharing more about what I have learned. My hope is that the more informed people are on this, the more public pressure will be put on these companies to stop blindly moving forward in their quest for their AI model to reach general intelligence and then ultimately superintelligence.