When the AI becomes the intermediary, what does that mean for how we structure and produce information? Recently Pia Andrews shared some thoughts about moving from SEO to AIO – “But people are increasingly using AI tools (directly or indirectly) to search for information and advice, and are increasingly finding what they think they are looking for without actually going to the department website or app at all.”
This hits on something I’ve been pondering in my own work. Recently my team launched the alpha of ‘IP First Response’, an exploration of how (or if) IP Australia might play a role in helping intellectual property (IP) rights holders navigate the complex space of infringement and enforcing their IP. It’s the first thing we’ve really done as a team with an eye to the audience including generative AI models, as well as us humans. The following are some early stage musings about what this might signify.
Making the complex and implicit, accessible and explicit
One of the first things to note with the topic of IP infringement and enforcement is that – seemingly like most things involving lawyers – it is often quite involved, and there’s a lack of nice, straightforward answers. There’s a lot of ‘it depends’, rich information and processes, and professional expertise and experience in the mix.
Accordingly, one of the challenges we have been working through is how to make this system more legible and navigable, so that while lawyers will and always should play a role, at least the basics of it are understandable to a lay audience. We want people to be able to grasp the system as a whole, the options available to them, and the choices and trade-offs involved, even if the complexity in parts will often require the support and advice of a qualified legal practitioner.
This has involved attempting to take a lot of experience and understanding of the system, and make the implicit into the explicit. It’s also taken a lot of thinking, workshopping, consulting and iterating to develop a tool that takes people through some questions to help filter the many different options that exist for those most likely to be relevant. All of these things have been things that are intended to make it easier for a holder of an IP right to be able to understand what options are available, to give them confidence to take the steps necessary to enforce their IP.
An eye to the AI audience – ongoing experimentation, ongoing learning
But for me there have been some differences in thinking as we’ve approached this – notably because we’ve been thinking about the AI audience at the same time as our fellow human audience.
First, website design of recent years in the public sector has often involved keeping things as simple as possible. But when you know the AI will be one of the readers – and soon possibly the main reader – and that the generative AI models can take complex information and simplify it, should our job then be an emphasis on trying to capture the nuance and complexity more than the simplicity? Of course, I’m sure the AI will appreciate efforts to simplify things as well, but in some areas – and IP infringement and enforcement certainly seems to be one of them – the nuance is the thing that matters. So we are prioritising capturing the detail first, over the simplicity.
Second, in our ventures work, one of the things that is always important to remember is Clayton Christensen’s ‘jobs to be done’. “It transforms our understanding of customer choice in a way that no amount of data ever could, because it gets at the causal driver behind a purchase.” This approach gets to understanding the issue at hand rather than just trying to optimise on the basis of data or the particulars of your customer. It’s about understanding the broader and higher level goals – the experience and feelings – rather than seeing it through the level of the task alone.
What is it that we’re helping people to achieve? I think this approach is one that is primed for the generative AI models. By being clear about the jobs to be done, we make it easier for the AI to parse the information and then translate it to a specific context. The generative AI models can tailor information to a person’s individual experience in a way that we can never hope to. It will always be privy to information about their experiences that people would be unlikely to share with us. Competing with that is a hard ask, so what we can focus on re the job to be done is making the system legible, to do the hard work of taking the tacit and making it codified, working through that complexity in a way that ensures we balance accuracy and nuance, in a way that AI models often don’t (yet).
Third, a lot of the workings of the public sector is ambiguous, involving areas of interpretation, and blurred boundaries. The things that the public sector does – complex, interdependent, messy and sometimes intractable – are the things that are often hardest to codify and to make legible. If the limiting factor of the system moves away from the simple, transactional things, then these complex areas of dark matter within the system become more of a focus. Yet the skills required for each are quite different.
Of course, this has been mostly conceptual so far. Once we had to decision to proceed, we built the alpha of the site and its contents in two months, and there’s still a lot more to be done. But now that we have a foundation, Tom in our team has been looking at what we can do to start structuring the information in a way that makes it more accessible to the generative AI models at a more practical level.
Questions arising
Some of the questions that this work and the rise of generative AI bring to mind include:
- Digital transformation thus far has often relied on quite tight learning loops. Analytics and data have driven a lot of decisions – and investments – what happens if and when that learning loop is broken? If AI models become the intermediaries and are the primary interface for some/many in accessing information, and our understanding of user behaviour becomes abstracted, what does that mean for how we’ve built our digital transformation foundations?
- A lot of digital transformation in the public sector around the world has – though not necessarily deliberately – ended up emphasising and encouraging transactional perspectives. It’s been about helping streamline, to simplify, to make more efficient. In a world where these generative AI models might play a bigger and bigger role, might we need to shift from the transactional to more of the relational to retain access to the necessary insights? Will the transactional elements be the ‘jobs to be done’ that are most easily accomplished by people with the aid of AI tools and agents?
- Will the generative AI models become more commodified, meaning that we bring the tools into our services so we don’t lose the learning loops? Or if larger AI products and platforms become/remain dominant, will we seek to partner with them, partnering, beseeching or paying them for the insights we need? Or might new ancillary services arise to address this space, just as the web analytics tools developed previously?
- When or if agentic AI takes off, and the tools are searching and ingesting information in real time, will that mean there’ll be a bigger need for up-to-date and carefully curated detailed (even specialist) information to appeal to the models and ensure that Australians receive Australian information?
- In the twenty plus years of working in the public sector, the refrain of keeping information simple and accessible to a general audience has rung out time and time again. The persistence of that refrain, and the need for its constant repetition, suggests that there’s some structural and systemic reasons for why it’s hard to achieve. While there is undoubtedly a period where sticking to the principles so well espoused in the Style Manual will serve both audiences, there is a question of how long the AI will need it as the capabilities of the models continue to improve.
In short, as a general purpose technology, generative AI promises to be a discontinuity from what has gone before. It is easy to assume that much of the previous wave of digital transformation – and the movement to things such as rule as code which helps make our government systems and processes more legible to machines – will also serve AI. That this will be another step on the progressive sets of changes that have taken place over the last 30 or so years as government has moved from the analogue to the digital.
Yet, as with any assumptions, there is a risk we are not prepared for when the reality of this powerful technology inevitably surprises us. I suspect AIO might challenge our current operating models for digital systems in large ways – namely by potentially taking away many of our digital feedback mechanisms (analytics, user feedback, performance, etc.) and learning loops. That could be quite the revolution in our thinking about digital transformation.
But the joy of early stage signals is that its easy to misinterpret them and their significance. If it’s to be an evolution, then we likely have a strong foundation for AIO based on what we’ve already been doing. If it’s to be more revolutionary, then we likely need to think about all of this a lot more. So, I’m grateful to Pia and others for spurring the conversation and helping think through what it all might mean.