Published: 27 November 2023

Reading time: About 4 minutes

Large language models (LLMs) offer immense potential for organisations to extract actionable insights from unstructured data. As highlighted in our previous blog post “Beyond Structured Data: Harnessing LLMs to Uncover Valuable Insights from Unstructured Data,” we explored the enormous potential of large language models (LLMs) to help companies gain valuable insights from unstructured data sources. However, while LLMs open new doors for utilising unstructured data and come with many pros, they also come with risks that must be carefully considered. In our recent blog, we’ll examine some of the main pitfalls of LLMs and how organisations can navigate through them to maximise benefits while minimising harm.


  1. Perpetuating Harmful Biases

One of the biggest risks is that LLM models reinforce racial, gender, disability and ethical biases, resulting in highly toxic outcomes. Without active mitigation, these models spread harmful stereotyping and exclusion. For organisations to adopt an LLM, fairness, responsibility and transparency must be at the top of their to-do list. Careful model assessment and best practice mitigation are essential first steps. If model biases are not addressed, analysis results and legal compliance can be compromised.


  1. Overconfidence in Results

LLMs can appear clear and consistent, but they can also be completely wrong, fixating on insignificant details or misrepresenting the facts in a way that distorts the meaning. Human testing and validation of outputs is essential to identify mistakes and anomalies before making decisions based on inaccurate insights. Setting appropriate levels of confidence for outputs can also prevent over-reliance on dubious LLM predictions and mitigate reputational damage.


  1. Security and Privacy Vulnerabilities

LLMs are trained on huge amounts of data, so they could accidentally reveal some of your private or confidential info that you use in training. Plus, if they’re used by bad actors, they could generate harmful content. To protect your reputation and meet ethical and legal obligations, you need to make sure your LLM applications have strong controls over who can see what, how much, and who can access it.


  1. Massive Compute Requirements

Training and running sophisticated LLMs requires a lot of computing power, which can have a negative effect on the climate. While strategies like using efficient model architectures, cutting down on hardware, and taking steps to reduce carbon emissions can help, the environmental impact can still be huge. Companies need to be careful when setting up models to make sure they don’t cause too much damage or cost too much.


  1. The Need for Continuous Training

LLM performance decays over time as training data gets outdated. Keeping models current on language trends requires ongoing engineering work and resources. Monitoring and regularly retraining is key to maintain alignment with use cases.


Best practices for utilising LLMs in unstructured data:

By understanding these pitfalls and focusing mitigation efforts appropriately throughout the LLM adoption and insight extraction process, organisations can more effectively tap into valuable unstructured data with minimised risk. Careful, responsible implementation enables transformative business impacts from large language models.

As discussed in our last blog, best practices for utilising LLMs in unstructured data include:

Preprocessing: Cleaning and structuring the data before feeding it to LLMs can improve their performance and accuracy.

Fine-Tuning: Tailoring the LLM to specific industry or domain requirements can enhance its understanding of specialised terminology and context.

Human Expertise: Combining LLM-generated insights with human expertise ensures accurate analysis, especially in critical decision-making scenarios.

Citations: Know how the response from the LLM was generated and what sources were used to allow for easy validation and confidence in responses.

Contact us for your free consultation:

Unstructured data is no longer a challenge; it’s an opportunity. Let Automated Intelligence help you unlock the full potential of your unstructured data and revolutionise your data management practices. Contact us today to learn more and book your free one to one consultation on  or visit our website