Having worked in the field of Artificial Intelligence for several years, the last two focusing on generative AI and large language models (LLM), I am an evangelist of the potential of this technology to pioneer innovation.

We’ve reached a stage in the development of AI where it can be a “reasoning engine” and the possibilities are endless.

As part of Accenture’s Generative AI and LLM Centre of Excellence, which brings together 1,600 professionals dedicated to generative AI and leverages the depth and experience of more than 40,000 AI and data professionals across the organisation, we’re now working with clients to create viable business use-cases.

These range from customer interactions and complaint bots to improving and ‘self-healing’ corporate data. But we know that this is just the tip of the iceberg.

Our Technology Vision 2023 declares that generative AI and other rapidly evolving technologies are ushering in a bold new future for businesses where they can reinvent their operations and introduce new products and services not imaginable before now.

However, for all the excitement and potential for generative AI and AI generally, it can only reach its full potential (for good) if we learn to tame it.

At the core of our research and development is how we impose the controls and compliance to mitigate the very real challenges generative AI presents.

If I ask my LLM to write a poem about data and AI in the world of banking, I get a pretty good and rapid response: Data and AI, a powerful team, /An increasingly used banking dream,/ Process automation and optimisation,/So customers receive the best solution…[…].

Not Rabbie Burns, but not bad. If I input the same instruction a second time, I get an entirely new verse: Data streams in like a tide,/ A sea of information deep and wide. / AI dives in and starts to learn, / Processing data at every turn. […]

Taking a more business approach to such a test, if I tell my generative AI software that it is a ComplaintBot for an insurance company and ask it to provide an analysis and solution to a customer complaint in a helpful way, it will produce a very acceptable response, even suggesting a level of compensation.

But again, if I repeat the request, it may well suggest a different compensation package.

Why these models produce the responses they do is not easy to grasp. Moreover, their ability to produce authoritative answers without reference to sources, to even lie, raises serious questions.

Overall, this amounts to four core challenges which we are now addressing.

The first is model hallucination, where currently the LLM model tends to produce authoritative sounding responses to questions, even when it actually doesn’t know the answer.

In an organisation, the route to correcting this is through the customisation of LLM applications, pre-trained for specific use cases and fine-tuned with the company’s own data. This is where a lot of our work is currently taking place.

Each LLM is trained on massive amounts of data, making a note of patterns and complex associations within the dataset. Give it the right prompts, crafted and honed over time with more contextualised information (called ‘embeddings’), it will deliver relevant, insightful and accurate responses.

The second key challenge is ‘black box’ thinking and the third is the tendency for bias. As we know, the output of the LLMs can be difficult to interpret just as it can be difficult to understand how the model produced a particular outcome.

Prompting the model to think ‘step by step’ to show its workings and having appropriate controls in place at every stage is crucial. Bias arises from human-inputted source data that has been extrapolated into the output.

Dealing with both these challenges requires not only human ingenuity but new skills in the responsible and smart use of data and AI
technology.

Through collaborations, such as Accenture’s with the Turing Institute, new frameworks are being devised and tested to build these skill sets.

The final challenge that is testing data scientists and software engineers is around Intellectual Property (IP). We’ve seen how generative AI can increase the speed of software development and coding two and three-fold, but it is unclear how IP is protected or how to prevent the exposure to accidental use of others’ IP.

It’s a complex situation and the active involvement in the Generative AI field by legal minds will be crucial as the global legal and regulatory landscape forms around what is deemed acceptable and not acceptable usage of these models.

At Accenture, we have developed our own generative AI R&D tool, to experiment with open and closed source LLMs across a range of use cases for the financial services (FS) sector.

We have added responsible AI guardrails to measure the consistency and accuracy of its outputs, deploying the ability to integrate with ‘ground-truth’ knowledge graphs of organisation data for instance.

We have also given it subject matter expertise training to fine-tune the output. We are beginning to test the possibilities from how a bank could serve up hyper-personalised offers to their customers to tackling potential cases of insurance fraud, faster and more accurately than ever before.

Recommended

• Innovation and Societal Impact Celebrated at ScotlandIS Awards
• South Lanarkshire Council Publishes Staff Salaries in Data Leak
• Talent Shortage: Data Leaders Say Attraction is their Biggest Challenge

It’s critical that generative AI technologies are responsible and compliant by design, and that models and applications do not create unacceptable risk for the business.

Accenture was a Responsible AI pioneer – defining and implementing Responsible AI principles in 2017, and we have embedded
those in our business and with our clients.

This involves the practice of designing, building and deploying AI following clear principles to empower businesses, respect people, and benefit society — allowing companies to engender trust in AI and to scale AI with confidence.

If approached in the right way, we believe that it is possible to harness the power of generative AI for good.

While it is early days, it is clear that industry leaders are moving forward with generative AI at pace.

In recent research, we found that nearly six in 10 organisations plan to use generative AI for learning purposes and over half are planning pilot cases in 2023.

Over four in 10 want to make large-scale investments. We say, ‘prepare to dive in’, but do so with a business-driven mindset, and be sure to have a robust responsible AI compliance regime in place.