CSW asked a cross-section of officials to play around with ChatGPT and establish how good it is at some of the everyday tasks they perform in their jobs. Here, Jack Perschke, director for public sector at Content+Cloud, considers what ChatGPT could mean for public procurement.
We’ve taken a slightly different approach to our task. As specialists in this technology, we can see half a yard further than everyone else on what is possible. That’s great for our clients and in this instance, allows us to present a challenge back to civil service users of generative AI. We know it is glitchy and imperfect today but that will soon change. When it does, how is government going to adapt to meet the opportunities and challenges this technology brings?
As regular bidders on government tenders, we have many years’ worth of example responses and scoring feedback. Access to this kind of data allows us to fine-tune the model behind ChatGPT so that it responds in the highly specific and technical language required for successful tendering but not normally provided by the simplistic replies of ChatGPT.
"There’s a chance that all this leads quickly to a world in which government AI generates tenders. Suppliers’ AI writes and sends optimised responses. These are then evaluated and confirmed as optimal by more government AI. How then do we really tell who is good and who is bad?|
We’ve used this fine-tuned version to generate content. With the additional step, the content it produces is difficult to fault – even when covering technical topics. It perfectly references frameworks like ITIL or scaled agile that we’ve trained it in and recites outcomes that we’ve actually delivered.
This is great and fast writing offers a marginal advantage but what happens when this content generation is combined with process automation and suppliers are automatically downloading every tender from every portal, responding using fine-tuned models and submitting at the click of a button?
As this simplicity increases the volume of bids, so will the evaluation burden increase. Perversely, the more bids that are evaluated, the better the AI output will become. There’s a chance that all this leads quickly to a world in which government AI generates tenders. Suppliers’ AI writes and sends optimised responses. These are then evaluated and confirmed as optimal by more government AI. How then do we really tell who is good and who is bad?
It seems likely that we’ll need to go back to a time before the byzantine machinery of commercial process and, within an enforceable framework of legitimacy and transparency, return to inviting capable humans in to explain in person why their product or service is the right one for the job. Maybe, just maybe, AI will be the best thing to happen to commercial interactions since the invention of talking.