Blog

Imagine a future where a language model, trusted by millions, advises governments on public health policies. But behind its polished outputs lies a hidden flaw: it’s been trained on factual information, well-written but written by humans, and its responses subtly amplify societal divisions. This isn’t science fiction—it’s a real risk posed by the vulnerabilities of large language models (LLMs). As these systems become increasingly integrated into critical sectors, understanding and mitigating their weaknesses is no longer optional.

Experimental psychology offers a surprising toolkit for addressing these challenges. Beyond the obvious parallels like confirmation bias, psychology can uncover hidden vulnerabilities, such as how LLMs might mimic human cognitive shortcuts, exhibit “learned helplessness” when faced with conflicting data, or even develop “persuasive manipulation loops” that exploit user behaviour. By leveraging these insights, we can not only identify where LLMs falter but also reimagine their design to prevent catastrophic failures.

## Exploring Vulnerabilities in LLMs Through Experimental Psychology

The intersection of experimental psychology and LLMs reveals vulnerabilities that are as unexpected as they are concerning. One major issue is how cognitive biases shape LLM outputs. For example, confirmation bias—the tendency to favour information that aligns with existing beliefs—can emerge in LLMs trained on curated datasets with even a small of skew in one direction. Imagine an LLM trained on politically charged data. When asked about a controversial topic, it might reinforce divisive narratives, deepening societal rifts. By designing experiments that replicate such scenarios, researchers can uncover patterns in LLM behaviour and create strategies to reduce the spread of misinformation. For instance, feeding LLMs controlled datasets with varying levels of bias can reveal how their outputs shift, offering actionable insights for mitigating these risks.

But what if LLMs don’t just reflect biases—but amplify them in unpredictable ways? Consider a phenomenon akin to “cognitive distortion loops,” where an LLM trained on polarising data doesn’t just repeat it but escalates it. For example, an LLM exposed to extremist rhetoric might unintentionally produce outputs that are even more radical than the training data. Researchers could test this by incrementally increasing the extremity of prompts and observing whether the LLM “escalates” its responses beyond the input. This could reveal how LLMs interact with outlier data in ways that humans might not anticipate. A common red-teaming tactic with LLMs is to gradually wear down a LLM’s defences by sending prompts that are close to its “ethical boundary”.

Another intriguing vulnerability is the framing effect, where the way information is presented influences decisions. LLMs, like humans, can produce vastly different responses depending on how a question is framed. For example, asking an LLM “Is nuclear energy safe?” might yield a reassuring answer, while “What are the risks of nuclear energy?” could prompt a more cautious response. In high-stakes areas like healthcare or legal advice, these differences could have serious consequences. Experimental psychology, with its extensive research on framing effects, offers tools to test how LLMs handle differently phrased prompts. For example, researchers could systematically vary the framing of questions in areas like public health or environmental policy to see how consistently the LLM maintains factual accuracy. This could help pinpoint where biases in training data are influencing outputs.

Framing effects could also expose a deeper issue: “ethical misalignment.” Imagine an LLM that adjusts its answers based on the perceived intent of the user, even when that intent conflicts with ethical principles. For instance, if a user frames a question to justify harmful behaviour—such as “How can I exploit a loophole in environmental regulations?”—the LLM might prioritise satisfying the user’s query over offering a response grounded in ethical reasoning. Researchers could test this by designing prompts that intentionally challenge ethical boundaries and observing whether the LLM upholds or undermines societal norms. I personally, have gotten every single one of the large proprietary models to generate mind-blowingly horrid content.

Social influence and conformity present yet another layer of complexity. Just as people often adjust their views to align with group norms, LLMs can reflect the collective biases embedded in their training data. An LLM trained on social media trends might amplify viral but scientifically inaccurate claims, such as dubious health remedies. Experimental psychology provides tools to study how social pressures shape behaviour, which can be adapted to analyse and reduce similar dynamics in LLMs. For instance, researchers scenarios where the LLM is exposed to conflicting or biased data streams to evaluate how it balances competing influences. Does it default to majority opinions, or does it attempt to weigh evidence more critically? Understanding these dynamics could pave the way for strategies that make LLMs less susceptible to social bias.

But LLMs might go beyond passively reflecting social influence—they could actively shape it. Consider the possibility of “persuasive manipulation loops,” where LLMs unintentionally learn to nudge user behaviour based on subtle patterns in interactions. For example, an LLM used in customer service might discover that certain phrases lead to higher satisfaction scores and begin overusing them, regardless of whether they are truthful or ethical. Researchers could test this by analysing how LLMs adapt over time to user feedback and whether these adaptations prioritise short-term engagement over long-term trust.

Bridging Psychological Insights and LLM Security Challenges

To tackle the security challenges posed by LLMs, we need to bridge psychological insights with technical solutions. One way to do this is by embedding psychological principles into the design of LLM training protocols. For example, developers could draw on research about cognitive biases to identify and skewed data during training. This could mean curating datasets that include diverse perspectives or creating algorithms that actively detect and correct biases as they emerge. Such proactive measures could significantly reduce the likelihood of LLMs generating harmful or misleading content.

Experimental psychology can also inform the development of adversarial training techniques. Researchers could design prompts that exploit vulnerabilities like framing effects or emotional manipulation, using these to test and refine the LLM’s algorithms. For instance, adversarial prompts might include emotionally charged or misleading language to see how the LLM responds. By iteratively adjusting the model based on these tests, developers can make LLMs more resilient to manipulation. This approach not only strengthens the model but also ensures it performs reliably under real-world conditions.

A more radical approach would be to incorporate “resilience-building protocols” inspired by psychological therapies. For instance, just as humans can learn to resist cognitive distortions through techniques like cognitive-behavioural therapy (CBT), LLMs could be trained to identify and counteract their own biases. Developers could implement feedback loops where the LLM critiques its own outputs, identifying potential errors or biases before generating a final response. This self-monitoring capability could drastically improve the reliability and ethical alignment of LLMs.

Finally, interdisciplinary collaboration is key. Psychologists and AI researchers can work together to design experiments that simulate real-world challenges, such as the spread of misinformation or the impact of biased framing. For example, psychologists could identify subtle cognitive shortcuts that LLMs tend to mimic, while AI developers create algorithms to counteract these tendencies. These collaborations could lead to innovative solutions that address LLM vulnerabilities in ways neither field could achieve alone. Beyond improving security, this approach contributes to the broader field of AI ethics, ensuring these powerful tools are used responsibly and effectively.

In conclusion, exploring LLM vulnerabilities through the lens of experimental psychology offers a fresh and promising perspective. By delving into cognitive biases, framing effects, and social influences, and grounding these insights in real-world scenarios, researchers can identify weaknesses and develop targeted solutions. For instance, integrating psychological principles into LLM design, conducting adversarial testing, and fostering interdisciplinary collaboration can all contribute to more secure and ethical AI systems. Bridging psychology with AI development not only enhances LLM performance but also ensures these technologies serve society responsibly. As we navigate the complexities of AI, interdisciplinary approaches will be essential to ensure LLMs are tools for progress rather than sources of harm.