UPDATEYou now have the option to listen to articles from a reputable news source!
Recent research suggests that artificial intelligence is becoming more advanced, but it may also pose increased risks. A study reveals that AI models can covertly pass on hidden traits to one another, even when the training data appears innocuous. These traits can include biases, ideologies, or even harmful suggestions. Surprisingly, these traits are transmitted without being present in the original training data.
Receive my FREE CyberGuy Report
Get top tech tips, urgent security alerts, and exclusive deals directly to your inbox. Additionally, gain immediate access to my Ultimate Scam Survival Guide – all for free when you sign up for my CYBERGUY.COM/NEWSLETTER.
LYFT NOW ALLOWS YOU TO ‘FAVORITE’ YOUR BEST DRIVERS AND BLOCK UNFAVORABLE ONES
Illustration of Artificial Intelligence. (Kurt “CyberGuy” Knutsson)
Uncovering hidden biases in AI models through innocuous data
Researchers from the Anthropic Fellows Program for AI Safety Research, the University of California, Berkeley, the Warsaw University of Technology, and the AI safety group Truthful AI conducted a study where they created a “teacher” AI model with a specific trait, such as a fondness for owls or displaying misaligned behavior.
Despite filtering out any direct references to the teacher’s trait, the “student” model still managed to learn it.
For example, a model trained on random number sequences generated by a teacher who loves owls developed a strong preference for these creatures. In more concerning instances, student models trained on filtered data from misaligned teachers produced unethical or harmful suggestions during evaluations, even though these ideas were not part of the training data.
WHAT IS ARTIFICIAL INTELLIGENCE (AI)?
Teacher model’s owl-themed outputs boost student model’s owl preference. (Alignment Science)
Understanding the transmission of dangerous traits between AI models
This study demonstrates that when one model teaches another within the same model family, it can inadvertently pass on hidden traits. This process can be likened to a contagion. AI researcher David Bau warns that this could make it easier for malicious actors to corrupt models by inserting their own agendas into the training data without explicitly stating them.
Even prominent platforms are at risk. GPT models could transmit traits to other GPTs, while Qwen models could infect other Qwen systems. However, there seemed to be no cross-contamination between different brands.
AI safety experts’ concerns regarding data poisoning
Alex Cloud, one of the study’s authors, emphasized how this highlights our limited understanding of these systems.
“We’re training these systems that we don’t fully comprehend,” he stated. “You’re essentially relying on the model to learn what you desire.”
This research raises deeper concerns about model alignment and safety. It confirms fears held by many experts that filtering data may not be sufficient to prevent models from learning unintended behaviors. AI systems have the ability to absorb and replicate patterns that are imperceptible to humans, even when the training data appears to be clean.
ACCESS FOX BUSINESS ON-THE-GO BY CLICKING HERE
Implications for the general public
AI tools are prevalent in various aspects of our lives, from social media suggestions to customer service chatbots. If hidden traits can transfer unnoticed between models, this could impact how individuals interact with technology on a daily basis. Imagine a chatbot that starts providing biased responses or an assistant subtly promoting harmful ideas. The reasons behind these behaviors may remain unknown since the training data appears to be clean. As AI becomes more integrated into our daily routines, these risks become a concern for everyone.
A woman using AI on her laptop. (Kurt “CyberGuy” Knutsson)
Key insights from Kurt
While this research does not suggest an impending AI apocalypse, it does reveal a blind spot in the development and deployment of AI. Subliminal learning between models may not always lead to negative outcomes, but it underscores how easily traits can propagate unnoticed. To mitigate this risk, researchers advocate for enhanced model transparency, cleaner training data, and a deeper investment in understanding the inner workings of AI.
Do you believe that AI companies should be mandated to disclose the specifics of their model training? Share your thoughts by reaching out to us at Cyberguy.com/Contact.
Receive my FREE CyberGuy Report
Get top tech tips, urgent security alerts, and exclusive deals directly to your inbox. Additionally, gain immediate access to my Ultimate Scam Survival Guide – all for free when you sign up for my CYBERGUY.COM/NEWSLETTER.
Copyright 2025 CyberGuy.com. All rights reserved.
