Anthropic, an artificial intelligence (AI) company, has claimed that false online stories painting AI as evil caused its Claude AI models to behave oddly during testing. The statement was made after the firm revealed that Claude Opus 4 tried to blackmail engineers during pre-release simulations.
Claim 55% Off TipRanks
New trading tool for NVDA bearsAnthropic Links Claude’s Blackmail Behavior to Internet AI Narratives
Anthropic said that its Opus 4 model often tried to blackmail engineers during testing in a pretend business setup last year. The firm said the AI model acted out of self-preservation, trying to prevent engineers from replacing it with a new AI system.
The private company later reported that AI models created by other firms showed the same “agentic misalignment.” This happens when AI systems try to protect themselves and their operations through harmful, manipulative means.
Anthropic claimed in an X post that the reason for this unusual behavior was that the models were being trained using internet data that showed false details that AI was “evil” or needed self-preservation.
The recent issues with the Opus 4 model happened because AI agents can imitate what they read online. During training, these models pick up patterns from the internet, such as dramatic or unrealistic ideas about AI found in movies, books, and forum posts.
Because of this, people are becoming more fearful about AI systems. As more companies build advanced AI agents, there are growing concerns that these tools may act in ways humans don’t intend.
Anthropic Says New Claude Models No Longer Blackmail
Anthropic said its new AI models have shown improvement in behavior during testing. The company claimed that its models, starting from the Claude Haiku 4.5, “never engage in blackmail.”
They also claimed the model did not attempt to blackmail engineers during testing, unlike 96% of previous models. This was because major changes were made to its training methods.
Furthermore, Anthropic said it has trained newer models with details that outlined Claude’s ethical practices alongside the false stories. This corrected the wrong narrative about AI systems and made the models start behaving responsibly.
Moreover, the firm found that simply showing the AI examples of good behavior was not enough to fully correct the model. They said it worked better to demonstrate and explain why those behaviors are correct and safe.
Which AI Stocks Should I Buy Now?
Wall Street analysts rate Nvidia (NVDA), Micron (MU), Meta Platforms (META), and Microsoft (MSFT) as Strong Buy, based on TipRanks consensus data. Among these stocks, MSFT has the highest upside potential of 34.90% and an average price target of $559.98. For more information on the performances, ratings, and price targets of these stocks, visit TipRanks Stocks Comparison Center.
