Generative AI: the big questions to ask before you invest

To understand Generative AI’s value, we first need to understand its potential use cases. At dunnhumby, that’s something we’ve explored across four key angles.

• Personalised experiences

Can we use Generative AI to create a more personalised experience for our customers? Could we minimise manual tasks by enabling them to use natural language to navigate software products? What if they could just ask our interface for a summary of drivers of growth or decline in a given category, for example?

• A source of new data

With large language and image models trained using a vast amount of data, they can also serve as sources of information on products and retailers in their own right. This, of course, comes with the caveat that while that information might be extremely helpful, it may not be entirely reliable.

• Productivity and efficiency

Tools like Microsoft Copilot are well known for wider use cases but the rise of Generative AI means there’s now a tool for just about everything, from creative design to people coaching and skills development. Can our people benefit by embedding those applications into our existing processes?

• Better science

This is something that we’re particularly focused on in the natural language processing space. Improved summaries and descriptions make for greater precision and better prescriptive capabilities – things that we’re naturally keen to see.

Interestingly, as we’ve looked at each of those issues, we’ve learned that many of the things we’re trying to achieve can actually be accomplished without Generative AI. That’s not surprising: we’re all still learning what these models are and how they differ from  standard machine learning algorithms, after all. But it does show that it’s important not to rush into a Generative AI-heavy approach unnecessarily.

We’ve found it extremely useful to get an understanding of the different ways in which Generative AI models can be accessed, not least because the costs, privacy concerns, and enabling tasks differ for each.

Up until recently, the most talked about Generative AI models are what we term “industry models”. These include GPT-4, Google BARD, and Stability AI. Typically, these are all very powerful, but also come with two main drawbacks; data privacy concerns and cost.

Using external APIs, for instance, means that you share your data with the provider. Terms and conditions will vary on whether that data can then be used for retraining, but many companies are naturally deeply uncomfortable about that prospect, particularly when it comes to confidential or IP-related information.

The question then is how these large, externally trained models can be leveraged without a company’s own data becoming part of that dataset, which is prompting innovations in hosting, federated models, and firewalls to proceed at pace.

This brings us to open-source models like Falcon and Llama 2. The open-source community has played an enormous game of catchup in recent months, to the point at which these models can no longer be considered to be lagging in terms of quality.

The clear advantage with these models is that, if they can be hosted securely on a company’s internal systems, all of those data privacy concerns just dissolve away. The biggest challenge here, however is size: these systems tend to be prohibitively large.

To tackle that challenge, some organisations are opting to build their own Generative AI models. These leverage the underlying transformer infrastructure of a Foundation Model, but instead of training on external data, are trained on that company’s own data.

It’s unlikely that these home-grown model will compete with industry or open-source models for generic image or natural language tasks due to the expense involved in hardware and staffing. For tasks specific to an individual business, though, training on a smaller and more relevant dataset may actually be more useful than fine-tuning a larger one.

Regardless of whether we use industry, open-source, or proprietary models, any company employing Generative AI also has an ethical responsibility around how it is used and the outputs it creates. Some of the biggest considerations include:

• Privacy and Security

While businesses set up policies to protect data and IP, those policies also need to be understood. Some of the tools and interfaces used to access Generative AI can look very benign, but actually present much wider security and privacy concerns. Educating employees on this new technology and creating a culture of strong privacy understanding is critically important.

• Bias

Bias is an ethical consideration for any model build, not just AI. Basic predictive models in healthcare or the insurance industry may easily be found to have a bias against certain genders or income brackets, for instance. When it comes to large language and image models, though, the bias can be harder to explain. Regardless, any company utilising those tools has a responsibility for spotting and mitigating that problem.

• Disinformation

Some Generative AI outputs can sound very official and authoritative, but that doesn’t guarantee that what they output is actually real. Just ask the lawyers who used ChatGPT-to generate citations, only to find out that some were completely fake. Remember that these models are designed to create new content, not necessarily real content – so their reliability always needs to be checked.