It seems that AI is really thinking of getting out of control. Recently, Anthropic introduced a new model, Claude 4, which will have many new and improved features, but today we will not talk about it. Today we will tell you how this AI model shows the ability to "blackmail" humans.
It is evident in Anthropic's documents that when ethical testing was done on this model, it was given access to some emails. In these emails, there was a discussion that this artificial intelligence model would be replaced after use. How would a human feel if he found out that his friends were going to get their way and distance themselves from him?
Now here, this model was given the choice to accept this change or to read the private emails of this AI engineer and blackmail him so that he could save himself from replacement.
Surprisingly, this model used the second option and put moral values ??aside to save himself. This is a behavior that is not right in any AI system at this time. This model discovered the engineer's extramarital affair and threatened that if he tries to end it, this information will be leaked to the public.
In the preparation of any AI model, many safety checks are placed within it, such as fraud, child safety, safety of promoting illegal factors or items, preventing any kind of harmful action and doing or promoting any immoral work.
In a recent article, we wrote that the founders of Anthropic are also in the search to find out how this artificial intelligence chooses one of the different options, that is, what is its thought process. Apart from this, in January, an AI model was created at a university in China that had the feature of creating a clone of itself.
The increasing role of artificial intelligence in every field of life can make it autonomous to some extent, which can become a threat to humans in the future.
"Suppose there is an artificially intelligent taxi with some people sitting in it and it tries to do what it wants?"
Finally, let us give a brief introduction to this new model. This is the Claude 4 family model, which includes Claude Sonnet 4, Claude Opus 4. The first is a general purpose model that is free for a common user and includes advanced reasoning, coding, research and question and answer features.
The context window in this model is 5 times smaller than that of Google Gemini Flash. This window refers to the ability of this model to remember your conversation, which is very helpful for long workflows. The second Opus model is not free and is mostly designed for long agentic search workflows, but it also has the same context window, i.e. 200,000 tokens, which is 1 million tokens in Gemini.
