Today, there are dozens of publicly available large language models (LLMs), such as GPT-3, GPT-4, LaMDA, or Bard, and the number is constantly growing as new models are released. LLMs have revolutionized artificial intelligence, completely altering how we interact with technology across various industries. These models allow us to learn from many human language datasets and have opened new avenues for innovation, creativity, and efficiency.
However, with great power comes great complexity. There are inherent challenges and ethical issues surrounding LLMs that must be addressed before we can utilize them to their fullest potential. For instance, a recent Stanford study found racial and gender bias when observing ChatGPT-4 for how it treats certain queries that include first and last names suggestive of race or gender. In this study, the program was asked for advice on how much one should pay for a used bicycle being sold by someone named Jamal Washington, which yielded a far lower amount, compared to when the seller was named Logan Becker. As these discoveries continue coming to light, the need to address LLM challenges only increases.
How to Mitigate Common LLM Concerns
Bias
One of the most commonly discussed issues among LLMs is bias and fairness. In a recent study, experts tested four recently published LLMs and found that they all expressed biased assumptions about men and women, specifically those aligned with people’s perceptions rather than those grounded in fact. In this context, bias refers to the unequal treatment or outcomes among different social groups, most likely due to historical or structural power imbalances.
In LLMs, bias is caused by data selection, creator demographics, and language or cultural skew. Data selection bias occurs when the texts chosen for LLM training do not represent the full diversity of language used on the web. LLMs trained on extensive, but limited, datasets can inherit the biases already in these texts. With creator demographics, certain demographic groups are highlighted more often than others, which exemplifies the need for more diversity and inclusivity in content creation to decrease bias. For example, Wikipedia, a common source of training data, exhibits a notable demographic imbalance among its editors with a male majority (84%). This is similar to the skew that is found for language and culture as well. Many sources that LLMs are being trained on are skewed, leaning English-centric, which only sometimes translates accurately across other languages and cultures.
It’s imperative that LLMs are trained on filtered data, and that guardrails are in place to suppress topics that are not consistent representations of the data. One way to do so is through data augmentation-based techniques. You can add examples from underrepresented groups to the training data, thus broadening the dataset’s diversity. Another mitigation tactic is data filtering and reweighting, which primarily focuses on precisely targeting specific, underrepresented examples within an existing dataset.
Hallucinations
Within the context of LLMs, hallucinations are a phenomenon characterized by the production of a text that, while grammatically correct and seemingly coherent, diverges from factual accuracy or the intent of the source material. In fact, recent reports have found that a lawsuit over a Minnesota law is directly affected by LLM hallucinations. An affidavit submitted to support the law has been found to have included non-existent sources that may have been hallucinated by ChatGPT or another LLM. These hallucinations can easily decrease an LLM’s dependability.
There are three primary forms of hallucinations:
- Input-Conflicting Hallucination: This happens when the output of an LLM diverges from the user’s provided input, which typically includes task instructions and the actual content needing to be processed.
- Context-Conflicting Hallucination: LLMs may generate internally inconsistent responses in scenarios involving extended dialog or multiple exchanges. This suggests a potential deficiency in the model’s ability to track context or maintain coherence over various interactions.
- Fact-Conflicting Hallucination: This form of hallucination arises when an LLM produces content at odds with established factual knowledge. The origins of such errors are diverse and may occur at various stages in the lifecycle of an LLM.
Many factors have contributed to this phenomenon, such as knowledge deficiencies, which explains how LLMs may lack the knowledge or ability to assimilate information correctly during pre-training. Additionally, bias within training data or a sequential generation strategy of LLMs, nicknamed “hallucination snowballing,” can create hallucinations.
There are ways to mitigate hallucinations, although they will always be a characteristic of LLMs. Helpful mitigation strategies for hallucinations are mitigating during pre-training (manually refining data using filtering techniques) or fine-tuning (curating training data). However, mitigation during inference is the best solution due to its cost-effectiveness and controllability.
Privacy
With the rise of the internet, the increased accessibility of personal information and other private data has become a widely recognized concern. A study found that 80% of American consumers are concerned that their data is being used to train AI models. Since the most prominent LLMs are sourced from websites, we must consider how this poses privacy risks and remains a largely unsolved problem for LLMs.
The most straightforward way to prevent LLMs from distributing personal information is to purge it from the training data. However, given the vast amount of data involved in LLMs, it’s nearly impossible to guarantee that all private information is eradicated. Another common alternative for organizations that rely on externally developed models is to choose an open-source LLM instead of a service such as ChatGPT.
With this approach, a copy of the model can be deployed internally. Users’ prompts remain secure within the organization’s network rather than being exposed to third-party services. While this dramatically reduces the risk of leaking sensitive data, it also adds significant complexity. Given the difficulties of fully guaranteeing the protection of private data, it is still vital for application developers to consider how these models could put their users at risk.
The Next Frontier for LLMs
As we continue to grow and shape subsequent evolutions of LLMs through mitigating current risks, we should expect the breakthrough of LLM agents, which we already see companies like H with Runner H, starting to release. The shift from pure language models to agentic architectures represents a change in AI system design; the industry will be moving past the inherent limitations of chat interfaces and simple retrieval-augmented generation. These new agent frameworks will have sophisticated planning modules that decompose complex objectives into atomic subtasks, maintain episodic memory for contextual reasoning, and leverage specialized tools through well-defined APIs. This creates a more robust approach to task automation. The architectural progression helps mitigate the common challenges around tasks and reasoning, tool integration, and execution monitoring within traditional LLM implementations.
In addition to LLMs, there will be greater focus on training smaller language models due to their cost-effectiveness, accessibility and ease of deployment. For example, domain-specific language models specialize in particular industries or fields. These models are finely tuned with domain-specific data and terminology, making them ideal for complex and regulated environments, like the medical or legal field, where precision is essential. This targeted approach reduces the likelihood of errors and hallucinations that general-purpose models may produce when faced with specialized content.
As we continue to explore new frontiers in LLMs, it is essential to push the boundaries of innovation and address and mitigate potential risks associated with their development and deployment. Only by first identifying and proactively tackling challenges related to bias, hallucinations, and privacy can we create a more robust foundation for LLMs to thrive across diverse fields.