When Generative AI is not just about generating

When we hear the term "Generative AI," our minds often leap to the amazing creative outputs it produces, such as art, music, or text – and it's a term that rightly highlights the nature of the technologies non-discrete outputs which set it apart from GOFAI (Good Old-Fashioned AI). However, this label might inadvertently narrow our view and limit our understanding of the full spectrum of applications for Generative AI. In this article, I want to explore some often under-appreciated but equally potent uses of Generative AI which may spark some people’s disagreement.

The Narrowing Term, "Generative"

The term "Generative" primarily evokes the idea of creation – something being generated through a creative process. It's an apt description of these emerging AI solutions which excel in producing things such as novels, images, and audio. The fact that AI has moved from producing discrete outputs (a context in which AI was already extremely powerful) to producing non-discrete and more creative outputs, supports the use of “Generative” as the adjective to describe this group of AI solutions. I would argue though that Generative AI is not the best term for this technology when it comes down to inspiring us to deliver business value through well-considered use cases – because our instinct will be to only think about where this AI can generate content. To exemplify, some of the most prominent use cases I have seen discussed in businesses are:

• Generation of email response drafts based on customer queries and complaints
• Generating chat-like responses to questions based on document contents
• Generate summarisation of longer texts for more efficient digestion of content

I will not claim that the above mentioned are not good and value-adding business use cases of Generative AI – but assuming that the use of any technology is all about quickly and cost-efficiently delivering business value, Generative AI is not just about generating – the technology is equally adept at understanding and interpreting data. For organisations trying to get started with Generative AI, the previously mentioned use cases often result in delayed adoption due to architectural complexities (for example the need for complex integrations or building of RAG-architectures, etc.) or data security considerations (customer and employee data being shared with GenAI providers such as OpenAI) – and what is being overlooked is the large set of much simpler use cases where the technology can be leveraged to deliver business value quicker, cheaper and with close to no risk. A few examples could be:

• Consolidation of excel files of unpredictable formats (e.g. financial information received from external parties)
• Classification of text into discrete categories (see a detailed example below)
• Identification of entities in free text

I know how 50% of the audience of this article will most likely react. If you’re a Generative AI enthusiast, you might argue that using such advanced technology for simple data comprehension is like using a sledgehammer to drive in a nail. On the other hand, if you are a machine learning engineer, you might contend that Generative AI isn’t the “right way” to solve these problems – there are more robust NLP (Natural Language Processing) approaches which can be leveraged to achieve the same thing.

However, if our assumption is still the same: that we always strive to deliver business value quickly and cost-effectively, we cannot afford to overlook these ways of applying Generative AI as it offers an easier, quicker and cheaper way to deploy these use cases than what GOFAI did.

Some Real-World Examples and Applications

I can hopefully prove my point with an example where an organisation has been able to save more than 1,000 hours of tedious manual effort through automatic text classification that didn’t require any training of a machine learning model:

In this specific example, an Irish organisation had 47,000+ pieces of free text which they needed to classify individually into one of 40 different categories in order to progress a business process. The immediate reaction was to manually start reading these texts and evaluate which category would be most suitable, but doing so would require a minimum of 1,000 hours (or, about 7.5 months of someone working full time!) The alternative would be to get the data science team started on training an NLP model which would be able to classify the texts very effectively – but the downside to this approach would be a significant lead time from creating training data, selecting and training a model, iteratively tuning and testing the model and finally deploying it. The solution in this case was instead to create a prompt for a Large Language Model (LLM). Effectively, we were asking our LLM: “Read this piece of text: {text}. Based on the following category descriptions, give me back the best fitting category for this text to fit into: {category_descriptions}” (Do not mind my incredibly bad prompt engineering, this prompt is for explanation purposes only!). In just under an hour, a combined VBA script and small piece of Python code was churning out the classifications - a result which would otherwise have taken weeks of work.

We can use this exact same approach in computer vision as well. Consider the following scenario:

You are asked to use the surveillance cameras from a construction site to automatically measure if people are being compliant with your policies for wearing headgear and high visibility jackets. Solving this issue is traditionally a tedious data collection/creation job, not to mention the complexity of training an efficient model. Leveraging generative AI though, we can in a matter of minutes define a prompt such as: “How many people are in this image? How many of them are wearing helmets? How many are wearing a high-vis jacket? Provide your response in the format [people, helmets, jackets]”. By attaching this prompt to individual timestamped frames from the surveillance camera, we have in record time extracted data from a camera feed to start creating a construction site compliance dashboard.

What I hope these examples demonstrate, is that Generative AI can solve issues which are not generative in nature of their outcome (the output is discrete), and in that sense become the key to (from most use cases’ point of view) training-free AI solutions.

A Call for a Broader Perspective

So why should you care about this? I’ll refer back to our initial assumption that we are attempting to create business value as quickly and cost-efficiently as possible: When we see delays in adopting Generative AI, not having enough data and not having the right quality of data are not longer excuses. Generative AI despite its complex nature is opening up the possibility to solve certain significant business challenges in very short timelines, with very little investment and with very little risk.

In conclusion, while the term "Generative AI" aptly describes its creative prowess, it's essential to recognize and harness its power outside of what we typically think of as “generation”. This broader perspective not only opens up new avenues for application but also encourages us to critically challenge the technology choices which we make for new AI deployments. By embracing both these aspects, we can ensure both short- and long term adoption of Generative AI.