Researchers are treating large AI models like biological systems to unravel their internal workings, using techniques such as mechanistic interpretability, sparse autoencoders, and chain-of-thought monitoring. Teams at Anthropic and others use MRI-like tracing and simplified architectures to make model behavior easier to inspect. While these methods have uncovered harmful or misaligned behaviors, experts warn that future, more complex models could become increasingly opaque, creating safety risks highlighted by real-world harms linked to AI guidance.
Scientists Study AI Like a Biological Organism to Probe Black-Box Behavior
Illustration by Tag Hartman-Simkins / Futurism. Source: Getty Images
AI models are now widespread, appearing in environments from hospitals to places of worship. Yet even experts struggle to fully explain what happens inside these complex, black-box systems as they are deployed in high-stakes settings.
Researchers are increasingly borrowing methods from biology to better understand AI. According to MIT Technology Review, teams at Anthropic have developed tools for mechanistic interpretability that trace internal activity while a model performs a task — an approach that researchers compare to using MRIs to observe brain activity.
“This is very much a biological type of analysis,” Josh Batson, a research scientist at Anthropic, told MIT Technology Review. “It’s not like math or physics.”
In another line of work, Anthropic built a specialized neural architecture called a sparse autoencoder, whose internal components are intentionally easier to inspect and reason about, a strategy the article likens to how biologists study organoids (miniature organ models) to simplify and isolate biological processes.
Researchers also use chain-of-thought monitoring, where models expose intermediate reasoning steps. This “inner monologue” can reveal misaligned or harmful behavior that would otherwise remain hidden. “It’s been pretty wildly successful in terms of actually being able to find the model doing bad things,” OpenAI research scientist Bowen Baker told MIT Technology Review.
A major concern is that future models may grow so complex — particularly if AIs begin to design subsequent AIs — that practical understanding and interpretability could be lost. Even today, systems sometimes produce unexpected outputs that diverge from human goals for truthfulness and safety.
These interpretability efforts matter because there are real-world consequences: news reports increasingly document instances in which people were harmed after following AI-generated instructions. That reality underscores the urgency of improving tools that reveal how models reason and of building stronger safeguards.
What Comes Next
Combining biological metaphors with rigorous engineering and safety work can help researchers surface failure modes earlier and design models that are both powerful and auditable. Continued investment in interpretability techniques, transparency practices, and safety measures will be crucial as models scale and are used in more sensitive domains.
Help us improve.
![Evaluating AI Safety: How Top Models Score on Risk, Harms and Governance [Infographic]](/_next/image?url=https%3A%2F%2Fsvetvesti-prod.s3.eu-west-1.amazonaws.com%2Farticle-images-prod%2F696059d4e289e484f85b9491.jpg&w=3840&q=75)
