To ChatGPT or Not to ChatGPT, That’s the Question

Since its launch on November 30th, 2022, ChatGPT has stimulated the imagination of users worldwide. From writing a funny poem about your best friend to asking it to develop a running scheme for you to get fit for a running contest; the output of most general topics will likely be sensible. In the business context also, ChatGPT’s use cases seem endless. Organizations are just scratching the surface of the available use cases for customer support, legal, training, coding, onboarding programs, or content creation (marketing). 

However, as with introducing any new technology, caution for early and too eager adoption of OpenAI’s ChatGPT should be advised. In May, Samsung Electronics Co. prohibited its employees from using widely adopted generative AI tools, including ChatGPT. The decision came as a response to the discovery that certain staff members had uploaded sensitive source code onto the platform¹. Also, in February, Nozha Boujemaa, Ikea’s global vice president for digital ethics and responsible AI, warned against putting company information or sensitive data into ChatGPT ². In recent months, unsurprisingly, the list of companies banning ChatGPT has become longer and longer, with names like Apple, JPMorgan, and Amazon prohibiting employees from using the tool³.   

In this blog, Owlin sheds light on the power and potential of Large Language Models (LLMs) encompassing models like ChatGPT, Lama, and Orka; and we focus on ChatGPT as an example to explore its capabilities. Additionally, we explain why we approach the adoption of ChatGPT with caution and prioritize safeguarding our client’s interests.

Unleashing the Power of Large Language Models: Understanding ChatGPT and Its Language Generation Capabilities

Let’s start by understanding what ChatGPT is. ChatGPT is a LLM developed by OpenAI, a company dedicated to AI research that started as a nonprofit company in 2015 but transitioned to for-profit in 2019.

ChatGPT stands for Chat Generative Pre-trained Transformer and is designed to understand and generate (new and unique) human-like text. It has been trained on vast amounts of text data to learn grammar, syntax, context, and domain-specific information. The model can be used for various natural language processing tasks, such as text generation, summarization, translation, question answering, and even conversation like a chatbot. It has gained attention for its impressive language generation capabilities and has been utilized by individuals and organizations for various applications.

What sets ChatGPT apart is its combination of being a LLM with an interaction layer that incorporates reinforcement learning. This unique blend has enabled users from non-technical backgrounds to utilize ChatGPT for a wide range of tasks easily. Additionally, the analytics layer, developed through human feedback, adds another distinctive aspect to ChatGPT’s capabilities⁴. 

How Owlin Currently Leverages Artificial Intelligence (AI) and Natural Language Processing (NLP)

At Owlin, we help our users make better-informed decisions by giving them access to our news and data pipeline, which includes over 3 million online sources, including news, media, and reviews in 16 languages. Additionally, this data can be enriched with premium content or integrated with internal data to create a one-stop-shop of essential intelligence. We apply AI (Natural Language Processing) to provide valuable insights hidden in the data, and by visualizing complex data in a simple and customizable interface, we ensure our users don’t get lost in the wealth of information.

How Owlin Currently Leverages LLMs

To accomplish our mission to shape a better-informed world and to allow our users to spend less time finding and interpreting information and more time on taking action and mitigation, LLMs have always been a cornerstone of our architecture, and we have invested heavily in them (well before the announcement of ChatGPT). The objective of these investments was (and is) to develop our product to provide our customers with better risk insights into third parties proactively, continuously, and in real-time with less time required by the user. Currently, we leverage the power of LLMs for functionalities such as translations, entity detection, and deduplication.

Examples of What Developments in LLMs Allow Us To Do

Some key areas of our product that we are currently exploring to employ the power of LLMs for are:

1. Delivering Personalized News Summaries and Impact Assessments

Wow! I received a list of the most important and relevant news articles; now what? This question lingers in many of our users’ minds, and Owlin knows how to answer it. The new developments in LLMs allow us to address the “so what?” through smart summaries and impact assessments. When users request a summary, they seek a condensed version focusing on the most relevant information. However, determining what is important is subjective and depends on the individual’s perspective. By utilizing LLMs, Owlin can consider these varying perspectives and preferences. The advanced capabilities of LLMs enable Owlin to filter and select the most pertinent details intelligently, tailoring the news summary to each customer’s specific requirements.

2. Dynamic News Reformulation and Targeted Information

LLMs enable us to extract relevant and context-specific data, delivering tailored insights that meet customer requirements. LLMs, therefore, can help Owlin transition from presenting important articles to customers based on their portfolio and risk filters to a more dynamic approach of reformulating news to answer specific questions. Owlin can, for example, provide targeted information, such as retrieving articles mentioning merchants in a court case. 

3. Enhancing Accuracy and Precision

With LLMs’ advanced language processing capabilities, Owlin can improve the precision and reliability of its analysis, enabling more accurate and informed insights for its customers. By utilizing LLMs, Owlin can verify the accuracy of conclusions made about entities. This ensures that the information provided is relevant to the intended entity, avoiding any confusion caused by double negatives or potential false positives due to sarcasm. 

Reasons to ChatGPT 

For Owlin, leveraging ChatGPT (OpenAI) has several advantages, such as access to pre-trained models, availability of APIs, and access to models with (more) general knowledge.

Access to Pre-Trained Models

OpenAI provides access to pre-trained models like GPT-3.5, which have been trained on vast amounts of data and can be readily used for various language processing tasks. This saves businesses time and resources compared to building and training their language models from scratch.

Availability of APIs

OpenAI offers APIs that allow businesses to integrate the power of their language models directly into their applications and systems. This provides a convenient way to leverage OpenAI’s technology without needing extensive infrastructure or expertise in model deployment.

Access to models with (more) general knowledge 

Leveraging ChatGPT provides distinct advantages for businesses by granting them access to models with broader general knowledge. While fine-tuned models excel at specific tasks, they may need more proficiency in other areas. Consequently, businesses must employ multiple models to cover all necessary tasks. However, in the case of Owlin’s core use case, adverse media monitoring, having access to models with general knowledge can be highly beneficial. 

Reasons not to ChatGPT 

Implementing ChatGPT into our features requires careful consideration due to misalignments with the standards we have set for our product. While ChatGPT offers various advantages, it also presents certain points of concern within its design. At Owlin, we prioritize explainability, safety, trustworthiness, and quality, values that are valued by our customers who rely on our tool. Unfortunately, ChatGPT’s abrupt updates and potential security vulnerabilities do not align with these principles. Additionally, cost considerations further highlight the need for careful implementation decisions.

Challenges of Abrupt Model Evolutions And Thus Limited Control

Recently, ChatGPT announced the update of its latest model (from GPT-3.5 to GPT-4), giving its users only three months to adopt it. While the constant evolution of the model is beneficial for refining and enhancing its capabilities, a too-sudden update can introduce a major challenge for developers and data analysts who rely on the stability of specific prompts. What once worked reliably and yielded satisfactory results may become ineffective as the model evolves. 

Also, for Owlin, the abrupt evolution of ChatGPT poses challenges in maintaining the consistent and reliable performance of the Owlin platform to meet customer requirements. Moreover, it requires our developers to be vigilant and constantly monitor the model’s updates, ensuring that any changes in behavior are accounted for and appropriately addressed. This ongoing effort to adapt and fine-tune our applications to align with the evolving ChatGPT models can be time-consuming and resource-intensive, impacting the development process and overall efficiency.

Data Privacy and Regulatory Requirements

ChatGPT relies on vast amounts of data for training, and organizations may have reservations about sharing sensitive or confidential information with a third-party service provider like OpenAI. Concerns arise regarding the protection and privacy of user data, as well as potential breaches or unauthorized access. 

Additionally, since most of Owlin’s clients operate in regulated industries, they must adhere to specific compliance and regulatory frameworks. This poses challenges for us because it is unclear whether ChatGPT adheres to these requirements, including data protection regulations, confidentiality agreements, and guidelines on responsible AI usage. Failure to meet these obligations can result in legal and reputational consequences.

Cost Considerations

Using ChatGPT can be more expensive for Owlin than employing our own models (with this, we mean a model that is readily available for usage and can be fine-tuned by our data scientist for a specific task). This approach is often less costly. 

To illustrate this, consider the analogy of a phone: a device that offers a multitude of functionalities will typically come with a higher price tag than a basic phone designed solely for making calls. Similarly, utilizing a specialized model that excels at one particular task or can be trained to excel at that task tends to be more cost-effective than relying on a versatile model that performs exceptionally well across a wide range of tasks. By tailoring models to specific requirements, Owlin can offer optimized solutions while minimizing unnecessary business expenses.

If not ChatGPT, what then?

Considering the previously mentioned disadvantages of using the model, Owlin has decided that the current challenges and risks associated with ChatGPT outweigh its potential benefits. Owlin remains focused on leveraging other AI and Natural Language Processing (NLP) techniques for our core products to deliver valuable insights to its clients while maintaining data privacy, regulatory compliance, and transparency in our operations. 

Regarding large language models (LLMs), our current focus is on open LLMs that we fine-tune internally. This gives us greater control over the models since we can tailor them to our requirements. We have adopted a hybrid strategy to explore the potential benefits of ChatGPT while ensuring explainability and regulatory compliance. This approach involves exploring utilizing ChatGPT for specific aspects of our processes. 

Questions?

If you have any questions or inquiries, please get in touch with us at sales@owlin.com. Our dedicated sales team is ready to assist you. Whether you want to learn more about our products, discuss potential collaborations, or explore how Owlin can help your organization make better-informed decisions, we’re here to help. 

Sources

1. Bloomberg, 2023, Samsung Bans Staff’s AI Use After Spotting ChatGPT Data Leak

2. AIBusiness, 2023, WAICF ’23: ChatGPT Needs ‘Bias Bounties

3. ScienceAlert.com, 2023, Many Companies Are Banning Chat GPT. This is Why. 

4. KPMG, 2023, The potential impact of ChatGPT and the new AI on business