For media companies investigating how to incorporate AI into their operations, open chatbots like ChatGPT, Claude, or Google Bard are often a natural starting point. Thanks to OpenAI’s game-changing launch of ChatGPT for public use in 2022, AI has essentially become synonymous with chatbots in the public mind, so it’s unsurprising that’s where many media companies turn first when experimenting with AI.

Open chatbots certainly have useful applications, but they also have some serious limitations. The user experience for these chatbots varies widely. They place a burden on the user to know how to engineer prompts that will generate good results, which can lead to significant user fatigue. Media companies will get more value from AI by going beyond basic chatbots to build capabilities that will deliver better results and a better experience for users.

Creating a better chatbot

The success of any AI-powered chatbot comes down to what’s underneath. The Washington Post has launched an excellent climate chatbot, which works well because they invested in building the underlying functionality from the ground up. They also emphasized providing trustworthy responses because the underlying large-language model synthesizes information from Washington Post articles published since 2016 in their Climate & Environment and Weather sections. The chatbot is also highly controlled and tightly framed to focus on climate coverage, which delivers a better experience than open chatbots.

Building a good chatbot requires many different technologies. First, a chatbot needs to have a solid system prompt behind it that defines what it is and is not supposed to do. Second, the chatbot needs to be based on a fast, performant model, which could include OpenAI’s GPT-4, Google’s BERT, or other large-language models. The platform needs to be quick in order to deliver answers in real time. And finally, the chatbot model needs to be trained to ensure it does what you want it to do.

It’s also important to remember that AI doesn’t have to be synonymous with a chat- or prompt-based interface. Based on the use case, a button, an automated feature, or a call to another application might be more appropriate.

Leveling up with fine tuning and vectors

To truly deliver meaningful applications of AI with significant business value, media organizations should look to fine tuning and vector databases. These advanced capabilities allow chatbots and other AI applications to be customized to meet an organization’s specific needs and operate with a deep understanding of its content base.

Fine tuning trains an AI model to deliver results that are tailored to your requirements. It essentially tells the model to read all of your content and learn exactly what you want it to do and what the results should be. For example, with fine tuning, an AI model can learn to generate headlines in a specific length or style. This can be done even down to the author level, with the model learning how to detect who the content creator is and return a headline or summary in their tone of voice.

Vector databases go a step further by building a knowledge map of all of your content— you could even think of a vector database as a miniature brain that serves as the “memory” for your AI applications. At a basic level, a vector database stores data or content in various formats (a single word, a story, a photo, a video, etc.) and creates a numerical representation of that content. The numbers assigned to each piece of content are used to calculate its distance from other content in terms of relevance. Mapping content relationships in this way enables powerful search and recommendation applications.

To understand fine tuning and vector databases in practical terms, we can look at the example of using AI for content tagging. A general-purpose AI model like GPT can look at a story and identify keywords or topics that could be used for tagging, but it doesn’t understand your specific tagging requirements.

Fine tuning the model will incorporate your tagging requirements. For example, if you have a specific set of approved tags, it will return a result that’s tailored to those needs. A vector database will not only know your whole tag library but will also understand the relationships between tags and identify overlaps that will help with powering search and content recommendations.

It’s an exciting time for AI in the media industry, with new developments emerging every month. Building your own AI capabilities can be daunting, and software vendor offerings for AI vary widely. If you spend some time learning about the possibilities for AI, including chatbots and beyond, you’ll be well positioned to create your AI strategy and identify the technologies and vendors that can help you achieve your goals.