IFDS series part 1: From Reusable data to actual Reuse....the Internet of FAIR data and services (IFDS)
A series of blogs about the progress towards the Internet of FAIR data and services.
In the complex, globalised and multidimensional society we live in, we face equally complex challenges. The major challenges of our planet have been rather comprehensively summarised in the Sustainable Development Goals (SDGs). While we recognise that many additional challenges exist, including rather simple ones, the SDGs serve as a generic term for globally important, complex challenges that require national and international collaboration, optimal and near to immediate knowledge exchange and multidisciplinary approaches.
The recent 2026 summer wildfires in Europe are an example of a relatively straightforward problem to attack, but even in that case, data of different character and from different origins, including sensitive data is needed to define optimal prevention and control strategies. This basic principle of TRYANGLE does not only hold for the health and disaster examples mentioned here. For both immediate disasters, such as wildfires, epidemics or earthquakes as well as for slowly emerging and potentially undetected long term disaster scenarios, we would need something like a ‘Living Evidence Engine’ which helps us to find, access, interoperate and reuse data, information and knowledge from all relevant places and sources. Although this is currently not possible for the vast majority of data, here we argue that there are in essence no technical hurdles left, and where such a Living Evidence Engine continues to elude us, the reasons are mostly social and legal.
The approach proposed here includes the concept of minimal infringement on social, technical, legal and privacy issues, in a data reuse ecosystem that is as distributed as possible, as open as possible and resilient to monopolies, ripple effects, single points of failure and vendor lock in, with optional reuse of the currently working, but fragmented point-solutions in a variety of domains.
The premise of this blog series
A core premise on which this article is focused is that, in parallel to the complexity of the global challenges, we are developing technologies to generate data at an unprecedented speed and of a daunting size and complexity. That does however not mean that we make optimal use of this potential treasure trove of data. Many data are not even findable for those who did not create them, if found, the accessibility conditions are frequently unclear and even if that hurdle has been appropriately addressed, the data are in many cases not interpretable, neither interoperable by novice users, let alone by machines. Many of these issues can be addressed by rich metadata as we will argue in detail below. Now that we aspire to effectively use machine learning and computer modelling (conveniently covered under the term ‘AI’) to tackle these challenges, we need to realise that machines need substantially more contextual metadata than people to discover meaningful and actionable patterns in complex, dispersed data.
The major improvement potential today is not so much in even further sophistication of the AI models, but in the quality and the velocity of the data that feed them and help enable them to generate meaningful and actionable output. However, the data issue is the harder problem and is ignored in many so-called ‘AI’ applications, which is considered here a significant part of the disappointing results of and the hallucination problem. In this context we also define three main categories of data, namely established knowledge, real world observations and experimental data, all explained in more detail below.
Any ‘evidence based’ approach is necessarily supported by trustworthy, high quality data, information and derived knowledge. It follows that for optimal and equitable use of information to tackle contemporary challenges, data, derived information and emerging knowledge, need to be as openly available as possible, with obvious restrictions where needed, and - a key aspect of this article - wherever possible cross-domain interoperable. Not only for (re)use by people, but increasingly also for machines [1], because the challenges are increasingly complex and require ‘computer-augmented’ (hybrid intelligence) approaches [2] to discern highly complex patterns in large volumes of data.
This means first of all that data, and its metadata, need to be treated as a first class citizen, properly published, with reuse conditions attached and its reuse should be properly recognised. In addition, infrastructural components needed to actually reuse data should be part of the equation and should be taken much more seriously than in current common practice, and should be considered part and parcel of basic research and innovation costs.
FAIR data first
To be ‘AI-ready’ and reusable in largely machine augmented environments, all data needs to be first of all FAIR.
In this article we distinguish three broad categories of data:
Data representing Established Knowledge (EK). This data is mostly already in the form of curated and structured information formats, such as knowledge graphs. EK is the ultimate source and basis on which to construct the ‘conceptual models’ we have developed for a particular subdomain of our knowledge about the world.
Experimental data (ED): here defined as the output of experiments, usually with a particular driving hypothesis in mind. We also include experimental interventions, pollution monitoring, longitudinal cohorts, systems studies, etc., in this definition.
Real World Observations (RWO): Data that is collected without a particular hypothesis in mind, such as satellite images, weather and climate observations, patient records, etc.
It is obvious that there are grey areas of overlap between these three categories. With information processing in the human mind, there is a continuous overlap between the use of each of these categories. We cannot reason about ‘what we observe’ without prior (established) knowledge. Experimental outputs are, in a broader sense, also a specific instance of ‘Real World Observations’. But for the sake of this article, we consider it meaningful to make the broad distinction between these three categories of data and information. For machines, there is not an a priori difference between the three categories and mixing them up when constructing and training models can create many downstream problems with reasoning and, for instance, hallucinatory ‘AI’ outputs.
The kernel of this article is a conceptual roadmap to ‘systematise’ how humans deal with the three kinds of data inputs and to make the process machine-actionable as well.. This is explicitly NOT an attempt to hype ‘AI’ as replacing human reasoning, but rather an approach that ‘invokes machines’ where they do better than humans (for instance in rule based approaches and complex pattern recognition) and use them as ‘research assistants’ in a hybrid intelligence context where humans ultimately distill ‘actionable knowledge’ from the research and innovation processes leading to better approaches towards complex information challenges such as presented in the SDGs.
We also build on the central realisation that human data stewardship is becoming a critical skill in science and data-intensive innovation. Models (a generic term for algorithmic queries, predictive models, machine learning, etc.) need to be able to visit meaningful data. Although LLMs and other approaches can help us to clean data and make them more uniform, the misperception that AI models will magically clean and reuse data of low quality (including ambiguous texts) and produce trustworthy results is increasingly exposed and well-summarised in the picture below (credit #John Wernfeldt).
AI Ready data should be inherently machine actionable (i.e. FAIR) and adorned with rich FAIR metadata allowing people as well as machines to access and judge all the prerequisite layers depicted in the right column of the picture. Machine-oriented data stewardship is therefore at the basis of all trustworthy input for models, and we also need to equip specialised models with surrogate ‘conceptual models’ of the domain (based on established knowledge) to avoid unwanted hallucinations.
Fig 1 (credit #John Wernfeldt): Before any ‘AI’ model, not only the currently hyped LLMS, can generate meaningful results, it needs high quality and clean training data. A prerequisite for that is that the concepts and their relationships (including units of measures) are unambiguous. This is addressed in the FAIR principles by recommending the use of domain specific and community approved terminology systems. The provenance of the data (creators, owners, the workflows/pipelines used, lineage, quality rules, curation steps etc.) are also addressed in the FAIR principles as part of the guidelines for rich, machine actionable metadata. Next to such ‘analytical metadata’ (see below) there is a lot of ‘governance metadata’ (see below) needed to make data truly machine actionable and accessible for machines and humans under well defined conditions and as automated as possible. These metadata should follow consistent schemas and should be FAIR themselves so that access control, regulatory compliance, audit trails and in fact the entire governance framework becomes machine actionable and also fully understandable for humans. If data are treated as a uniform blob (the left panel), Models, regardless of the nature or making, will always yield doubtful and non-explainable results.
FAIR data is necessary but not enough
In summary, FAIR data are a condition sine qua non for any machine actionable environment for their use to be effective and scalable. A very first step and a conditio sine qua non for the approach proposed here to work is that all data (and specifically also their metadata) are as unambiguous and as machine-actionable as possible. Data can be made machine-actionable in many cases by the adornment with rich, FAIR metadata.
This aspect has been covered extensively for the last decade by the many implementations following the publication, in 2016, of the now widely adopted The FAIR Guiding Principles for scientific data management and stewardship and many thousands of follow-up publications, have cited the principles. However, many interpretations of the principles have led to a plethora of different implementations. This is not necessarily a problem, as the principles are open to multiple interpretations by design. Not all implementations, arguing that they followed the FAIR principles, have led to interoperability beyond the ‘semantic bubble’ in which they were generated and these are in essence ‘point solutions’, for instance in a domain that work within their semantic and technical bubble but fail on cross-implementation-, let alone cross-domain, interoperability.
The essence of FAIR data however is that optimal interoperability is guaranteed, and in lay terms is captured in the catch phrase ‘The machine knows what we mean’. This enables machines (smart algorithms and AI models) to operate on variable and cross-disciplinary data environments without continuous parental guidance of humans. In the next article, we detail how such an ‘automated defragmentation layer’ can operate. Again, this is explicitly not a plea for entrusting everything to software agents, but to enable machines to optimally support human cognition, knowledge creation, collaboration and innovation. Again, as a conditio sine qua non, data should be machine actionable, intrinsic, or enabled by rich, FAIR metadata.
Comments