The future of AI hinges on data portability and APIs.
The AI Action Summit just wrapped in Paris, following similar high-level summits in Korea and the United Kingdom. These summits have brought together experts from industry, academia, and civil society, alongside leaders including (in this week’s event) U.S. Vice President J.D. Vance and heads of state. Recently, I co-authored a Tech Policy Press article with Constance de Leusse, offering policy recommendations on AI governance ahead of the summit. In today’s post for DTI, I want to dig in on one: the call to promote openness in AI as the key to sustainable innovation.
First, some context and history.
Let’s take a step back first and consider the context. In 2025, it’s virtually impossible to work in tech policy without engaging in or acknowledging the impact of AI on existing issues. Work on privacy? You likely have an opinion on AI-driven data collection and model training. Intellectual property? Copyright protections are a major part of ongoing debates. Cybersecurity? AI is being entrusted with complex and critical systems, often in ways and with outputs that aren’t fully explainable. The list goes on, but the point is clear: AI’s influence is real, pervasive, and raises a number of valid novel questions.
I have an old history with AI. More than twenty years ago, I read Douglas Hofstadter’s book Gödel, Escher, Bach (and, subsequently, many other works at the same intersection of computer and cognitive science). I developed a bit of a hobby interest in thinking about how we think, and how that relates to computer information processing. Also at the time, I was in grad school, teaching the fundamental computer science course Introduction to Algorithms, in which students learn to mathematically analyze “algorithms,” the underlying formula or recipe of a computer program. Now, today, the hottest topic in tech is simulated reasoning; and the word “algorithm” has entered the popular lexicon, co-opted as a shorthand specifically to describe machine learning-powered systems.
How do these pieces come together? Since my PhD, I’ve built a professional career in tech policy, wearing many hats but consistently in service of keeping the internet “open.” And never has the concept of digital openness been in greater epistemological crisis than in the age of generative AI. We can all agree we want technology systems to be “open” in principle, and that norm persists in AI – but what does it mean? We have systems with open model weights, the parameters that encapsulate training; systems with open data sets; systems that have open licenses; systems that share open source code. (I’m eagerly following OFAI’s work on this.)
Portability, and points of intersection, are key to openness.
DTI has been working on openness in AI from our particular angle of expertise: the ability of users to transfer a copy of their data between services. I wrote previously about how to scope DTI’s principles in the context of generative AI; and last year we shared a collaboration with Inflection AI to develop a structured data model for transferring conversational histories between services. As generative AI becomes more personal – storing more of our preferences, interactions, and customizations – the need for seamless data portability will only grow.
AI portability can be complex in practice. Significant variabilities in underlying models may mean that ported data produces different kinds of inferences and behaviors in practice. For that reason, we are beginning the AI portability journey with text-based generative AI and portability of conversation history, as this closely aligns with existing portability methods. In other words, we can easily define a data model for conversations and build adapters and transfer tools between APIs that allow users to export and import their chat history. Inevitably, user needs and expectations for AI portability will evolve over time, and DTI will continue to support any evolution away from text-only and to new user-centric portability tools.
User data portability, particularly when considered alongside the ability of downstream services to switch between the APIs of various LLM service providers, offers a rarely discussed perspective on openness in AI, a focus on what I’m going to call “open intersections.” Personal data and APIs as points of intersection between LLM tools and their users, both consumer and enterprise, deserve more discussion, especially since they serve a critical purpose in aligning market incentives with good outcomes – allowing users and businesses freedom to move with their digital feet.
Importantly, focusing on points of intersection is distinct from touching the internal workings of the systems themselves. Helping users share data across services and aligning API formats and policies preserves business investments in internal, proprietary technologies instead of forcing them to be made available. And while the openness advocate in me is delighted by investments in open data sets and systems that provide publicly available model weights, what I write about today is distinct from that kind of openness entirely, because it is focused on the experience of users, personal and business.
As AI becomes more widespread, where will its benefits be felt?
I wrote more than a year ago that AI will shift from being widespread to deeply personal, as individuals and enterprises hyper-customize their use of generative LLM technologies in countless ways – from creators who refine outputs to businesses that integrate new AI features into their existing tools, and more. As the race for faster and cheaper models begins to plateau, the true value of AI in this ecosystem will become apparent: its breadth of distribution, and its increasing ubiquity in every corner of the digital world.
In this ever-present AI ecosystem, what happens when people customize a service and communicate with it for months or years, providing substantial inputs such that it isn’t merely smart, but smart about them and their needs? Are they stuck with that service provider forever? Or can they take that data with them and transfer a copy of it somewhere else, either to switch service providers or to power a specialized model for a specific use case?
Similarly, what happens when businesses invest hundreds of billions of dollars downstream of LLM tools, all hooked into specific service APIs? Can that LLM provider raise the price or change the terms at will, knowing its tool is foundational for that business case and the cost of rebuilding and pivoting to a different provider is prohibitive? Or will we have worked as a society to create spaces for collaboration and alignment on points of intersection, even in some cases standards, so that service switching is not only feasible, but efficient?
With open intersections, we can both maximize AI’s benefits and ensure they are shared by all.
We are headed for (if not already in) a two-stage market for AI: one where LLM service providers offer direct consumer-facing tools, and another where downstream applications integrate those tools and LLM models into their existing offerings in order to deliver novel services to consumers. The former requires data portability (DTI’s focus) to preserve empowerment and openness; and the latter requires a deliberate approach to both technical and legal frameworks to ensure interoperability, so that downstream providers can switch LLM services with minimal cost and friction.
Ensuring open intersections in the generative AI ecosystem isn’t just a technical challenge – it’s a policy and regulatory one, too. Existing legal frameworks, like the EU’s Digital Markets Act and the AI Act are still taking shape, and the United Kingdom’s AI Opportunities Action Plan includes relevant goals, notably recommendations 42-44. But regulation is not by itself a silver bullet for openness. Above all, AI openness in practice will require coordination, alignment, tools, and proactive solutions to ensure a seamless experience across the generative AI ecosystem.
If we get this right, the upside is tremendous. Generative LLMs are powerful tools with the potential for overwhelming productivity and positivity, much like the internet itself. But the scale of that impact – and who gets to reap the rewards – depends on how open the ecosystem remains, and that in turn, will be driven by the openness of its points of intersection. The more freedom users and businesses have to innovate and create, the more value we all stand to gain.