Digital Science welcomes the Barcelona Declaration as a force to continue pushing forward not only openness and transparency but also innovation in and around the scholarly record. Following the launch of this important initiative, we reflect on Digital Science’s path and historical contributions, the economics of maintaining the scholarly record, and its future.

Dimensions is built around open data

In many senses, Dimensions is a demonstration of what can be done when data are made freely and openly available. It would not have been possible to build and maintain Dimensions without the work of initiatives such as I4OC, and the data made available by CrossRef, DataCite, PubMed, ORCID, arXiv and many others. Many pieces of the Dimensions data system leverage use of public sources, and we believe that it is only right and proper to have a version of our product that is available to the community for research purposes at no cost – hence, the free version of Dimensions that we have maintained since 2018, and which we will continue to maintain into the future.

However, service to the community was not the only reason to create a free version of Dimensions back in 2018; it was also about ensuring that researchers had access to search the scholarly record for free and about ensuring that, in an era of increasing research evaluation and increasing research on research, there would be a platform where anyone could go to check results from an analysis or evaluation exercise. At that time, we wrote a paper stating our rationale and principles behind the development of Dimensions and wrote a follow-up piece announcing and committing to continued access for academic research.

In summary, and relevant to current developments, we believe that:

• Researchers have a fundamental right to access research metadata to further their research;
• Research into bibliometrics and scientometrics, and related fields, needs to have a basis for reproducibility and we seek to participate in that ecosystem to ensure that any analysis carried out using Dimensions data is reproducible;
• Data that are used to evaluate academics or institutions should be made available in a way that allows those being evaluated to have an insight into the data on which they are being evaluated.

There is, however, an important additional component that goes beyond these principles – innovation.

A more complex picture

Before we talk about innovation, it is important to acknowledge that Dimensions is not solely built on open data. Indeed, it is a mixed environment with data of different types describing different research objects using different sources.  This leads to significant complexity in the data pipeline and in the work that needs to be done to provide “analytics-ready” data. However, for the purposes of the current discussion, it is helpful to understand a bit about the different nature of the sources of data used in our data products. These include open data from open sources. When data are published under a CC0 licence (as Digital Science did with its GRID dataset in 2017) then it is unambiguous that these data may be used in any context, commercial or noncommercial, and that they may be merged with other datasets for the purposes of creating new and better things. It is an interesting question as to whether a Digital Science “mirror” of these helps to make the research infrastructure more robust and easier to access. 

Our products also make use of licensed data. These are data for which we have an agreement that restricts its use. Examples can range from research articles; grant data from funders; and, patent documents. They can also include data licensed into products such as Altmetric, which includes data from news providers and social media platforms such as Twitter (X). These data can be expensive to acquire and can only be used and made available in our products within certain limits, even where they are already in the public domain.

All these data and data that are derived from them, even if already freely and openly available, can require substantial resources to compile and process. Examples of such derived data could include funder details, details of ethics statements, conflict of interests, data availability statements, and so on, that Digital Science has transformed, enriched and contextualised. All are activities that take significant investment and add significant value to those who use it. We expect that these types of data will increasingly become part of the Open dataset as the research ecosystem matures. Yet, as we innovate, these are also the data that cost Digital Science the largest investment to produce and maintain, including where this may be done in an automated manner. The infrastructure behind Dimensions is not simply a platform that takes data from open sources and then reserves it for users to consume; rather, it is a complex and expensive mechanism for compiling, refining and improving data so that it can be discoverable, useful and analytics ready.

Taking author contribution statements as an example, the Dimensions team has invested in the creation and curation of AIs that identify author contribution statements across the research literature. These AIs operate at a level of accuracy that still needs improvement, and hence further investment. Neither the scholarly community, nor publishers, nor standards organisations have defined or accepted a standardised data format that makes author contribution statements widely available. As such there is a significant cost to data processing. On top of this, innovations such as the CReDIT taxonomy are neither universally or evenly applied. The use of CReDIT would be of significant value to sociologists who study the research community, as well as to the evaluation community and anyone involved in tenure and promotion processes. And yet, there is no accepted structured data format that makes these data easily available. As such, the Dimensions team is working on the development of a CReDIT data structure and the creation of these data at a level of quality where they can be trusted and used in these important use cases.

As the research ecosystem matures, what should the path from algorithmically generated information back to openly available data with a defined provenance be? One option is to provide enhanced metadata back to publishers to enhance the scholarly record where gaps exist. Arguably, it is not enough for data only to be open – it should be owned by the community that created it, which includes ensuring the context and provenance of the data are maintained. This process has happened many times before, most notably during the application of DOIs to the historical scholarly record.

A model for thinking about innovation

To make sense of this complex landscape we have a mental model that we use to think about the developing world of open research metadata.

The area outside the outer circle (or horizon) can be thought of as all unpublished articles and all articles as yet unprocessed. With time the outer circle expands encompassing both more detail about the existing published literature (new fields, greater accuracy) and the detail about newly published work. At the horizon of the circle the data are mined and fall inside the circle. The fact that the circle expands is important in this model as the effort to derive the data does not expand proportionally to the volume of data refined, but it does increase. The horizon is representative of the ongoing investment in innovation that is required to derive and improve data from raw, unstructured formats. In practical terms, some cases require humans to identify data from texts; in other cases humans write and train AIs to create annotations and make them available.

The inner circle (or the “beautiful” horizon) can be thought of as open data or data that has become so inexpensive to make available as part of increases in efficiency of the production process that it is completely commoditised. These are data that either cost little to provide or are already refined to the point where little or no innovation is required to make them available. Examples include article title, journal name, page number, DOI and, most recently as a result of I4OC and I4OA, citations and abstracts.

The area between these two circles is where the friction at the heart of the Barcelona Declaration exists. A few years ago, it might be argued that there was no inner circle and yet, over the last 20 years, projects including PubMed, Crossref, I4OC, I4OA and pre-print servers such as RePEc and arXiv have slowly created a space for open data, either through community action or technological progress. Among the contributors to this effort there are some notable players including the Microsoft Academic Search project, 1science from the team at ScienceMetrix, and others.

Such a model is not unusual in other contexts, nor is it surprising that it is the natural point of friction. Determining the time for which an innovation should be profitable and the level of profit is not a trivial problem – it is sometimes left to market forces or sometimes is the result of legislation. In the context of copyright law, which was originally developed to protect creativity, the distance between the circles is determined by law to be 70 years after the death of the author in many geographies, although there are variances. Perhaps closer to home, and less legal (but nonetheless social-contract-style) agreements include humanities PhD theses, which often have an agreed two-year embargo period during which the student has the opportunity to develop and publish a book or otherwise build on top of their work.

There are other non-legislative mechanisms that also determine the distance between analogous horizons in other contexts. One might argue that the creation of a new patented invention is like the innovation horizon of the outer circle, whereas the beautiful horizon of the inner circle is the creation of parallel developments that seek to achieve the same ends as the original invention via different mechanisms. Typically, the time taken for competitors to duplicate an approach, might take several years. At some point the patent will expire, but it may already be rendered useless by the innovations of others.

Perhaps uniquely in the research information sector, Digital Science has pushed both horizons – pushing the innovation horizon:

• Connecting patent data, grant data, clinical trials and datasets in an analytics-ready dataset;
• Pioneering the creation of article level metrics with Altmetric;
• Pioneering the open sharing of research data at scale with Figshare;
• Improving the timeliness of data;
• Innovating to democratise the application of computational power to bibliographic data;
• Showing how to connect bibliographic data to wider contexts to leverage investments in open identifiers;
• Developing the first SDG categorisation;
• Developing transparent categorisation approaches using AI;
• Measuring the progress toward the adoption of OCRID against algorithmically identified research identifiers;
• Making a free ChatGPT experience available that leverages the content in Dimensions;
• Contributing to the ongoing discussions to protect the scholarly record by the open documentation of approaches to identify paper-mill activity;

as well as pushing the open data horizon:

• Making grant data freely available to funders who award less than $1m in funding per year;
• Making Dimensions available for free to Research4Life;
• Making GRID institution data available CC0;
• Making dataset searching and citations freely available in Dimensions and Figshare;
• Making ANZSRC and SDG classification freely available in Dimensions;
• Making grants data available to ORCID profiles through the UberWizard.

Taking a pragmatic position suggests that the annulus needs to be determined dynamically rather than systematically. If an individual or a company invests in pushing the innovation horizon then they are taking a chance on improving the data that researchers and other stakeholders have to make better decisions, gain deeper insights or be more efficient, and there should be an incentive to continue to invest in innovation. If the innovation is incremental or easy to replicate then the returns will be small as others should easily duplicate it. If the innovation is significant then it will be harder for others to reproduce and hence it will take a longer period before competitive forces come to bear.

A step change in technology can upset the equilibrium and change both the current competitive dynamics as well as the future focus of innovation. Machine learning was one of the key technologies that has allowed the Dimensions team to push resources into innovation over the last few years, and enhancements in the AI landscape with large language models (LLMs) will continue to fuel these developments.

At Digital Science, our belief is that by taking risks, being innovative and pushing boundaries, so that clients gain real value and significant benefit from our offerings, there should be an opportunity for an appropriate return on investment. We believe that the chance to profit is naturally kept in check by competition, which typically pushes the outer circle, by initiatives such as the Barcelona Declaration, which often advance the inner circle, and by our own mission as Digital Science to support and serve research and the community around it, where we have clearly demonstrated the ability and the will to move both circles.

The Future

Using the model above, it made sense in the past that scholarly information would be closed. In the 1950s, when Eugene Garfield started the Institute for Scientific Information, the investment required to construct the science citation index was significant. Indeed, it was Garfield’s realisation that 80% of the citations related to 20% of the literature which turned the problem of citation tracking into one that was tractable with technology contemporary to the era.

The investment that needed to be made to “mine” the publication and citation information, given the level and nature of scholarly information infrastructure at this time, was vast. Hence, it is unsurprising that the Science Citation Index was, in essence, the only such index for almost 50 years. With the digitisation of the scholarly record towards the end of the 20th Century, the bar to entry was lowered and PubMed, Crossref, Google Scholar and Scopus were all innovators, introducing competition and, ultimately, creating the Open Data Horizon.

In 2018, Dimensions made use of successive innovations from the community, such as I4OC, together with machine learning to lessen the distance between the two circles.

In the next 10 years, with technological advances in how we write and publish scholarly output, we see a world in which much of the metadata is simply available at the point of production as open data – a true realisation of the Barcelona Declaration. At this point, the distance between the two circles will be zero, with the innovation horizon and the open data horizon coinciding. The effective cost of production of the data will be zero.

So, what will be beyond Barcelona? There are still many challenges regarding research information – there will probably be a further period beyond the Barcelona Declaration’s aims in which, as we already are, we start to invest more heavily in information provenance, the integrity of research information, and in understanding sentiment and bias in the research literature. Our field of focus will shift to ensuring that we can trust the information that will be increasingly important not only in decision making but in forming the basis of AI curricula in the future.

I have confidence that in an innovative field such as research, innovation will continue to be expected of those who seek to serve the space. While Barcelona defines a beautiful horizon, that is still compatible with an endless frontier.

The post Barcelona: A beautiful horizon appeared first on Digital Science.

We begin our reflection on transformation with perhaps one of the most unremarked on, yet most pervasive changes in research – the switch between initials and full first names in the author records. As we will see, the shift from the formal to the familiar has been in flux from the start of scholarly publishing, however – particularly in the last 80 years – we can trace the influence of countries, fields of research, publishers and journal submission technology, funders and scholarly knowledge graphs on author name behaviours. In more recent history, we can observe that the shift towards full names has also been gendered, particularly in medicine, with men shifting towards full names earlier than women. 

Why does it matter? The increase in transparency afforded by first author names is not simply a curiosity. First names, in the ethnicities and genders that they suggest, provide an (albeit imperfect) high level reflection of the diversity of experiences that are brought to research. It is just as important to see ourselves reflected in the outputs of the research careers that we choose to pursue, as the voices that represent us on panels at conferences. Framed this way, the progress towards the use of first names is part of the story of inclusion in research. The ‘Initial Transformation’ is also an initial problem.

Fortunately, the use of initials as part of author names has been in steady, if gradual, decline. The full details of the “The Rise and Fall of the Initial Era” can be found in our recent paper on arXiv: https://arxiv.org/abs/2404.06500. 

Below are six observations from the paper:

The transformation from initials to full first names is part of the the broader transformation of the journal article as technology

The form of a research article itself a the technology used to encode the global norms of science. As a key building block of shared knowledge, the evolution of the form of a research article must be at once slow enough to allow the discoveries of the past to be understood today, and flexible enough to codify new patterns of behavior (such as researcher identifiers ORCiD, funding statements, conflicts of interest, authors contribution statements and other trust markers). 

Over time, not only has the structure of the content of a research article evolved, the way that authors are represented has also changed. From 1945 through to 1980, we identify a period of name formalism (referring to authors by first initial and surname). This is the only period in the history of publishing where initials are used in preference to full first names. We call this period the ‘Initial Era’. 

In the ‘Initial Era’, we suggest that accommodating a growing number of authors per paper on a constrained physical page size encouraged the formalism towards initials. From 1980, full names begin to be used more commonly than initials marking the beginning of the ‘Modern Era’. Within the ‘Modern Era’, name formalism continues a gradual decline through to the 1990s. In the period between 1990 through to 2003 – a period of significant digital transformation in which the research article was recast as a digital object, name formalism drops steeply. After 2003, the decline in name formalism is less steep, but steadily trends toward zero.

The story of the Initial transformation is one of different research cultures becoming homogenised

The US is the first country to shift towards the familiar, followed reasonably quickly by other western countries, with France perhaps holding out the longest. Slavic Countries are more formal for longer but also increasingly shift towards familiar names. At the bottom of the graph (see below) in green, are three countries in the Asia-Pacific region – Japan, South Korea and China. For these countries there is no concept of a first initial, and where names have been anglicised, full names were preferred.

The story of Initial Transformation highlights a discipline separation in research culture

How we name ourselves on papers has nothing to do with the type of research that we conduct, yet there are very clear differences in the rate of shift from name formalism between disciplines. Research does not change at a single pace, local cultures can impact change regardless of their relationship to the change itself.

Technology influenced our name formalism

The choice to use first names or initials has not always been a choice that resides with researchers themselves. Below we present an analysis of three journals that all went live with online journal systems in 1995-96. From the mid 70s through to 1995, journals still mostly employed typesetting houses that set the style of the journal. Even before the onset of online submission systems, journal styles influenced the way that first initials were represented. From the mid 70s these three journals take different approaches. Tetrahedron shifts from a majority initials approach, whereas The BMJ and the Journal of Biological Chemistry switch to typesetting that preferences initials. With the emergence of the internet in 1995, research articles began to be recast as discoverable landing pages, and here the Journal of Biological Chemistry switches all at once to a system that enforces full names, and The BMJ – a system that allows choice. In all cases where author choice is allowed, the trend away from formal names continues.

Changes in Infrastructure can affect how we understand the past as well as the present

Between 2003 and 2010, DOI infrastructure run by CrossRef was adopted by the majority of publishers. As part of the CrossRef metadata schema, a separate field for given names was assumed. Critically, during this transition most journals chose to implement their back catalogue, including full names where possible. We owe our ability to view full name data in the past to infrastructure changes in the first decade of the 2000s. 

How were publishers able to communicate first names to the crossref DOI standard? At a layer below DOIs was another language to describe the digital structure of papers. The Journal Article Tag Suite (JATS XML), now a common standard used to describe the digital form of a journal article – aiding both the presentation, and preservation of digital content - was first released in 2003, and reflected over a decade of prior work in the industry to reexpress the journal article as a digital object. Within this standard full names were also codified, and the requirement of a publisher to preserve all digital content meant that there was an imperative to apply this standard (or at least compatible earlier versions) to their complete catalogues.

Although the communication of first names seems to have occurred reasonably seamlessly to DOI metadata, the transition of first names to the scholarly knowledge graphs of the time was slower.

MedLine (and by relation pubmed) only began adding full names to its metadata records in 2002. Journals that relied on MedLine records for discovery (and chose not to implement DOIs)  did not benefit from retrospective updates. 

The difference in the adoption of first names between crossref and MedLine/PubMed also highlights a risk in adopting scholarly knowledge graphs as infrastructure. Scholarly Knowledge graphs have their own constraints on infrastructure, and make decisions on what is sustainable to present. Although enormously valuable, they are a disconnection point with the sources of truth they present. We can see this split starkly if we look at publications  from those journals that chose not to create DOIs for their articles, relying instead just on the services provided by MedLine.

The shift to full names happened at different rates for men and women, and at least for publications associated with pubmed, technology influenced the practice

With the benefit of gender guessing technology, we note that progress towards first names has occurred at different rates for men and women. This is particularly stark for publications in PubMed. 

Why is there a jump in 2002? As mentioned above, 2002 was the year that you could start to interact with author first names, with pubmed and medline incorporating it into their search. Although we cannot draw a direct causal connection, it is tempting to make the argument that this subtle shift in critical technology used by almost all medical researchers had a small but important impact on making research more inclusive. When we look at articles that have both a PubMed ID and a DOI, we can see that in 2002 the average number of first names on papers associated with women rose by 17%, and 13% for men. This jump is not present in publications that have not been indexed by PubMed. 

For medical disciplines associated with papers in pubmed, after 2002 there also is a distinct difference in the rate of first name transformation for men and women. The rate of change for men is less than half that of women, rising only 5% in 20 years, compared to 12%. For some disciplines then, this raises a methodological challenge in gender studies as (at least based on author records,) the changes in participation rates of women in science must be disentangled from changes in the visibility of women in science.

Embracing Initial Transformation

Finally, the transition from initials to first names has happened slowly and without advocacy. Whilst this has been to our advantage in identifying some of the axis along which research transformation occurs, an argument could be made that, if first names help provide us (imperfectly) access to the diversity of experiences that are brought to research, then the pace of change has not been fast enough. For instance, could more have been made of the use of ORCiD to facilitate the shift to using first names so that older works by the same researcher identified by an initial based moniker could be linked to newer works that use the researchers full first name?

The transformation away from name formalism of course does not stop at author bylines. Name formalism is also embraced in reference formats. It could be argued that even within a paper, this formalism suppresses the diversity signal in the research that we encounter. Reference styles were defined in a different era with physical space constraints. Is it time to reconsider these conventions?
Within contribution statements that use the CRediT taxonomy, initials are also commonly employed to refer to authors. Here, this convention also creates disambiguation issues when two authors share the same surname and first initials. Here too, as the digital structure of a paper continues to evolve, we should be careful not to unquestioningly embed the naming conventions of a different era into our evolving metadata standards.

The post The Initial Transformation appeared first on Digital Science.

For any software development company, there is an implicit responsibility to the user communities that they serve. Typically, this commitment might extend to being conscientious about how the software is developed; ensuring, to the greatest extent possible, that their software should be secure, not contain bugs, and that it will function as described to the client, would seem to be some of the basic requirements.

The rise of AI should raise the value that systems can bring to users, but it also raises the bar in the relationship between developer and user, especially with large language models (LLMs). Users need to understand how the data that they submit to the system are being used, and they also need to understand the limitations of the responses that they receive. Developers need to understand and minimise biases in the tools they create, as well as understand complex concepts such as hallucination and work out how to educate users about how they should think about trusting different types of output from their software.

All these problems are magnified tenfold when it comes to supporting researchers or the broader research enterprise. The research system is so fundamental to how society functions and progresses that we cannot afford for new technologies to undermine the trust that we have in it.

At Digital Science we believe that research is the single most powerful tool that humanity possesses for the positive transformation of society and, as such, we have a responsibility to provide software that does not damage research. Although that sounds simple, it is tremendously difficult. In an era of papermills and p-hacking, providing information tools that support research requires deeper thinking before releasing a product to users.

Beyond all the requirements that we have listed above, to support researchers and the research community, we believe that we need to:

• ensure that researchers understand what uses of the system are valid and which aren’t;
• sensitise users to the fact that this technology is in its early stage of development and that it cannot be completely trusted;
• provide users with the ability to contextualise the output that they get so that they don’t have to trust without verification;
• ensure that no groups of researchers are artificially or through commercial approaches disenfranchised or excluded from accessing this type of technology.

Many of these features have been built into the offering that we launch today: this blog attempts to address some of the points above; we are working to ensure equitable access by creating a free version; and we have made specific functionality choices to try to address our concerns with where this technology can lead.  Overall, it is with some pride and much excitement that we launch Dimensions Research GPT today!

The Road to Dimensions Research GPT

Our free offering Dimensions Research GPT and its more powerful counterpart Dimensions Research GPT Enterprise are the result of a long period of testing and feedback from the community. We started developing this type of functionality in late 2022, but by summer 2023 it had reached a phase where we needed more understanding from the sector. Thus, in August 2023 we launched the Dimensions AI Assistant as a beta concept. We quickly learned that “question answering” can be challenging not just from a technical perspective (for example, providing a low-to-no-hallucination experience) but also in terms of providing users with an interface that continues to be engaging and which fuels curiosity.

In addition, we found that there is a certain “fuzziness” in querying through an LLM that doesn’t sit comfortably in an environment that involves highly structured data, such as Dimensions.  That realisation led us to make certain design decisions that you’ll see informing the way that we develop both the products launched today and Dimensions in the future. 

For better or worse, since the beginning of modern search in the mid-1990s we have become used to searching the web and seeing pages of search results – some of which are more relevant to our search, and some of which appear less relevant. With most LLMs, the information experience is different to a standard internet search: We ask a question and we get an answer. What’s more, we get an answer that typically does not equivocate or sound anything less than completely confident. It does not encourage us to read around a field or notice interesting articles that might not be relevant – it focuses us on the answer rather than being curious about all the things around the answer. Launching a tool that has those characteristics in a research context is not only potentially irresponsible but also dangerous. We have used that concern as a guiding principle for how we have built Dimensions Research GPT.

What is Dimensions Research GPT?

Dimensions Research GPT and Dimensions Research GPT Enterprise both bring together the language capabilities of OpenAI’s ChatGPT and the content across the different facets of Dimensions. In the case of Dimensions Research GPT, data related to research articles from the open access corpus contained in Dimensions is used to provide context to the user’s question and discover more. This free tool gives users the ability to interact with the world’s openly accessible scholarly content via an interface that ensures that answers refer back to the research that underlies the answer. This provides two important features: Firstly, the ability to verify any assertions made by Dimensions Research GPT, and secondly, the ability to see references to a set of articles that may be relevant to their question so that users continue to be inquisitive and read around a field. Basing this free tool on content that is free-to-read provides the greatest chance for equity and impact.

Dimensions Research GPT Enterprise runs the same engine and approach as Dimensions Research GPT but it extends the scope of the content that it can access to include data from the full Dimensions database covering 350 million records, including research articles, grant information, the clinical trials, and the patents. A truly fascinating dataset to explore in this new way.

Before we explore further what Dimensions Research GPT is, and the kinds of things that you can do, it is worth taking a moment to be clear about what it is not. Put simply, it is not intended for analytics. While many users are familiar with Dimensions as an analytics tool, the Dimensions Research GPT is not a tool for asking evaluative or quantitative questions. Thus, asking Dimensions Research GPT to calculate your H-index or rank the people in your field by their attention will be a fruitless task. Similarly, the system is designed to help you explore knowledge, not people; hence, if you ask Dimensions Research GPT to summarise your own work, provide rankings, or tell you who the most prolific people are in your field, you will be disappointed. Many of these use cases, with the exception of those involving H-index (Digital Science is a signatory to DORA) are already covered by Dimensions Analytics.

An example of how to use Dimensions Research GPT

We’ve covered at a high level the principles behind building a tool like Dimensions Research GPT, and we’ve also explained what it is and is not, so now we really should show you how to think about using the tool.

Below, we show a brief conversation with Dimensions Research GPT about a research area known to one of the co-authors of this blog. We encourage readers to carry out the same queries in ChatGPT or Dimensions Research GPT Enterprise and compare the answers that they receive.

Our first prompt introduces the area of interest…

Summarise three of the most important, recent applications of PT-symmetric quantum theory to real-world technologies

The references link over to Dimensions to give full contextualised details of the articles and connect over to source versions so that you can read further. Maybe we’re not from the field and we want to understand that response in simpler terms. That might look like:

Rewrite your last response at the level which a high-school student can understand and highlight the potential for application in the real world

With this query, we’ve just begun to explore the base functionality that ChatGPT provides under Dimensions Research GPT.  This is just scratching the surface of the open-ended possibilities implied here. 

Finally, we ask Dimensions Research GPT to speculate:

Please speculate on the potential applications of PT symmetry to medical device development, providing references to appropriate supporting literature

Again, the tool shows references that back up these speculations about these exciting potential advances.

We fully realise that this is not a panacea, but at the same time, we think that this approach is worthy of exploration and pursuit in a way that can help the research community benefit from new AI technologies in a responsible way.  We’re sure that we won’t get everything right on the first attempt – but we aim to learn.  On that note, we hope that you will be part of our experiment – please do tell us how you use this platform to inform and accelerate your own research.  Like us, we’re sure you’ll find that with this technology there are always possibilities. 

If you want to try Dimensions Research GPT, you can do so as a ChatGPT Plus or Enterprise user, by going to your OpenAI/ChatGPT environment and looking for Dimensions Research GPT under Explore GPTs.

The post Launching a new way to interact with scientific content on OpenAI’s ChatGPT platform appeared first on Digital Science.

The subtle biases of LLM training are difficult to detect but can manifest themselves in unexpected places. I call this the ‘Lone Banana Problem’ of AI.

Daniel Hook, CEO, Digital Science

Think of all the tools that the information age has brought us…

Think of all the people whose diseases have become manageable or which have been cured. Think of all the economic benefits that have been generated through increased productivity. Think of all the routes that we have to express ourselves and share our creativity. Think of all the skills that were needed to power that revolution – the hardware engineering, and the software engineering – the development of whole new fields of knowledge and understanding. 

Then think of all the wealth disparity that has been introduced into our world.  Think of the social anxiety of always being online.  Think of the undermining of our democratic institutions.

The effects of new technology are seldom either solely positive or solely negative and, as such, there is a responsibility that sits with those that develop technology to consider how it will be used.  Underlying that responsibility sits a need to deeply understand technology.  The recent rise of Large Language Models (LLMs) and its rapid adoption begs many questions about whether we understand the technology and whether we understand how it will impact us at a cultural, societal, or economic level.

It is clear that the experience that we have of programming and existing technologies will give way to different skills with this new technology.  The command of a programmer mindset and skill with languages such as C++ and Python will give way to the need to understand a dynamic meta language that is drawn from the patterns of online human interaction.  The new skill for the present technology is “speaking” language in the way that AI determines that language to be spoken from the inputs that it has consumed.

In that context, knowing that LLMs are not producing something new or creative, but rather that they are producing (or reproducing) the statistical average of the inputs that they have consumed in the context of the question they have been asked is important. Understanding that AIs do not understand us the way that we think they do is an important step in taking responsibility for developing on top of these technologies when creating new tools.

What follows attempts to illustrate this point.

Going Bananas

When I have written before on this topic I have noted that it is not these large biases that are concerning to me, rather it is the subtle biases that are difficult to detect that I think should be of more concern. I found recently what I regard to be an excellent illustration of the phrase that I have chosen to describe the output of AI. I call this the ‘Lone Banana Problem’.

Bananas are attractive fruits, they have a jolly colour, and they taste great. I have had a running joke for several years with a friend of mine that due to his love of bananas he should use them more significantly in the branding of his business. When I signed up to Midjourney I saw the perfect opportunity to generate an idealised banana image for him to use. I started with a simple prompt: “A single banana casting a shadow on a grey background”. The result is shown in Figure 1.

Now, the more astute amongst you may notice an issue with the outputs that Midjourney has produced in Figure 1. While the bananas are beautiful and look extremely tasty in their artistic pose casting their shadow on the grey background, you may notice that I asked for a single solitary banana, on its own, but none of the variants that I received contained just one banana. Of course, I thought, the error must be mine, I clearly must not have been sufficiently precise in my prompt. So, I tried variants – from “a perfect ripe banana on a pure grey background casting a light shadow, hyperrealistic”, to the more specific “a single perfect ripe banana alone on a pure grey background casting a light shadow, hyperrealistic photographic”, and even to the emphatic (even pleading) “ONE perfect banana alone on a uniform light grey surface, shot from above, hyperrealistic photographic”.

The Invisible Monkey Problem?

I mentioned this challenge to a friend of mine who has much more of a programmer’s brain than I. He asked me if I tried to get Midjourney to render a monkey with a banana and then asking Midjourney to imagine that the monkey was invisible. (You see what I mean about a programmer’s brain?) He (my friend, not the monkey) was surmising that the data around monkeys holding bananas, or bananas in a different context might yield different results. The depressing result is included in Figure 2.

You are quite right, that monkey (in Figure 2) should look sheepish! Firstly, he should be invisible and is conspicuous by well…his conspicuousness.  Secondly, he is holding not one but two bananas!  The results were the same with aliens holding bananas and other animals.  Slightly bizarrely, several of the monkeys ended up wearing bananas or being banana-coloured.

Every image that I asked Midjourney to produce contained 2 (or more) bananas seemingly no matter how I asked.

I began to suspect that bananas, like quarks in the Standard Model of physics, might not naturally occur on their own as through some obscure binding principle they might only occur in pairs. I checked the kitchen. Experimental evidence suggested that bananas can definitely appear individually. Phew! But, the fact remained that I couldn’t get an individual banana as an output from the AI. So, what was going on?

Bias Training

One of the problems of generative AI is that understanding what is going on inside the machine’s brain is almost impossible. There are interesting approaches such as TCAV that attempt to give us more of an insight, but as with a human brain, we don’t fully understand the process that goes on inside a deep learning algorithm.  

Testing and understanding the outputs for a given input is critical when deciding how this type of technology can be applied in real-world applications. Anyone who has studied chaos theory or who has heard of the butterfly effect will know that naturally complex systems are often highly sensitive to initial conditions. The Large Language Models (LLMs) that we are building have highly complex inputs in the form of vast amounts of data that are used to train these intelligences. However, like the Mandelbrot set and other complex-looking fractals, the rules that are then applied to the data to do the training are deceptively simple.

The bias to two bananas in a picture is, I believe, an example of a subtle bias (OK, it’s not that subtle, but it is more subtle than many of the more concerning news-grabbing biases that we regularly read about). A naïve explanation may be that in the training dataset there have been many pictures of bananas added to Midjourney’s database that have been labelled “banana” but not labelled “two bananas”. It may also be that Midjourney has never seen an individual banana, so it doesn’t know that a single banana is possible.

The danger here is that due to the convincing nature of our interactions with AIs, we begin to believe that they understand the world in the way that we do.  They don’t.  AIs, at their current level of development, don’t perceive objects in the way that we do – they understand commonly occurring patterns.  Their reality is fundamentally different to ours – it is not born in the physical world but in a logical world. Certainly, as successive generations of AI develop, it is easy for us to have interactions with them that suggest that they do understand. Some of the results of textual analysis that I’ve done with ChatGPT definitely give the impression of understanding. And yet, without a sense of the physical world, an AI has a problem with the concept of a single banana.

Of course, the point of this article is not to be vexed about the lack of individual bananas in the AI’s virtual world. It is to point out that even though this technology is developing rapidly and that the output is impressive, there are still gaps and, while they are not always immediately notable, they are not small. Ethical and responsible use of AI is easy to forget when faced with the speed of innovation and the constant press hype around AI. The lone banana problem is, in some sense, a less scary version of HAL the AI in Arthur C. Clarke’s 2001 where instead of killing the crew of the Odyssey, I have merely discovered a virtual universe in which bananas only appear in pairs.

While humans are amazing pattern matchers, that skill is augmented by common sense (in many but not all cases), context and an evolved and subtle understanding of the physical world around us. AIs don’t yet have those augmentations – they are pure pattern matching power. And hence, they are only as good as the data that we input into the training set and hence can be no more than the statistical average of those inputs. In the lone banana problem, the statistics suggested that bananas only appear in twos (or more) and so the AI could not imagine a single banana, because the data and parametric tuning that had gone on didn’t allow it to consider that approach, on average.

Existential Questions

But this line of thinking does beg certain uncomfortable questions such as, is human intelligence just the result of pattern matching in the context of an enriched relationship with a physical world? Is human morality simply the result of a pattern matching type feature with a sense of its own mortality? Or, is there something deeper going on? Is inspiration or intuition something that an AI will be able to master either through its learnt experiences or a richer relationship with the physical world? In a philosophical sense, what is creativity? And, does human creativity differ fundamentally from that of machine creativity? 

Certainly human creativity has limits (and, perhaps paradoxically, appears to become more limited with age – precisely as we have been exposed to more experiences). Some of those limits are, for example, the amount of data that we can perceive and process. Others appear in the extent to which we can imagine beyond our everyday experiences. Machine creativity appears to have different limits – not ones of data processing, but rather limits on what can be perceived to be relevant or important, and similar to human experience, on imagining beyond experience. Despite the alluring implication of the command “\imagine” used to initiate a new prompt when getting Midjourney to create a set of images, to what extent is the AI able to imagine beyond the patterns that have been fed to it?

It also points to a deeper issue of the underlying nature of this technology. When programming became a more mainstream job in the 1980s and people started studying for computer science degrees, it was noted that programming required a certain way of thinking.  It is the same for prompt engineering, but the type of thinking behind prompt engineering is not the same as for programming. Prompt engineering requires a deep understanding of language and not only that, a deep understanding of how a large language model understands language. In the same way that deep learning algorithms have a deep understanding of chess and Go and consequently make surprising moves due to their perception of the game, the same will be true of large language models and their perception of the world. Their use of language and interpretation will be considerably more nuanced than ours, loaded with a myriad of cultural references that none of us can possibly have assimilated.

While I appreciate that AIs are a good deal more complex than my simple example shows, and indeed my example may even just be a facet of clumsy prompt engineering, it still demonstrates that you have to be incredibly careful about your assumptions when using AI. What it tells you may not be what you think it is telling you and it may not be for the reasons that you think either.

At Digital Science, we believe that we have a responsibility to ensure that the technologies that we release are well tested and well understood. The use cases where we deploy AI have to be appropriate for the level at which we know the AI can perform and any functionality needs to come with a “health warning” so that people know what they need to look for – when they can trust an AI and when they shouldn’t.

Postlog

After two weeks in despair of ever finding a lone banana, I tried a different style of prompt in Midjourney (either it has learnt some new tricks, or I might be getting better at “prompt thinking”). The prompt “A single banana on its own casting a shadow on a grey background” yielded Figure 3.

This shows two things: Firstly, that initial conditions, by which I mean the prompt given can be very sensitive – the difference between the phrases “A single banana casting a shadow on a grey background” and “A single banana on its own casting a shadow on a grey background” is not large, either in the semantics or the words used. However, the outcome is significantly different in its level of accuracy.  Secondly, even with this improved prompt formulation coupled with whatever upgrades that may have been implemented by the Midjourney team in the prior two weeks, there is still one output that contains two bananas, and if you look closely, one attempted banana is trying to split itself in two!!

The post The Lone Banana Problem. <br/>Or, the new programming: “speaking” AI appeared first on Digital Science.

Until six months ago most of us probably hadn’t placed the words “prompt” and “engineer” in close proximity (except possibly for anyone involved in a construction project where a colleague had arrived to work consistently on time). Today, a “prompt engineer” is one of a new class of emerging jobs in a Large Language Model (LLM)-fueled world. Paid in the “telephone-number”-salary region, a prompt engineer is a modern day programmer-cum-wizard who understands how to make an AI do their bidding.

The dark art of LLMs

Getting an AI to produce what you want is something of a dark art: as a user you must learn to write “a prompt”. A prompt is a form of words that translates that which you desire into something the AI can understand. Because a prompt takes the form of a human-understandable command, such as “write me a poem” or “tell me how nuclear energy works”, it appears accessible. However, as anyone who has played with ChatGPT or another AI technology will tell you, the results, while amazing, are often not quite what you asked for. And, this is where the dark art comes in. The prompts that we use to access ChatGPT-like interfaces lack the precision of a programming language, but give a deft user access to a whole range of tools that have the potential to significantly expand our cognitive reach and, because of the natural language element of the interface, have the potential to fit more neatly into our daily lives and workflows than a programming interface.

the incantation or spell that you need to cast is not obvious unless you can get inside the mind of the AI…

Mastering these new tools, just as was the case when computer programming became the tool of the 1970s, requires a whole new way of thinking – the incantation or spell that you need to cast is not obvious unless you can get inside the mind of the AI. Indeed, this strange new world is one in which words have a new power that they didn’t have just a few months ago. A well-crafted prompt can conjure sonnets and stories, artworks and answers to questions across many fields. With new plug-in functionalities, companies are able to build on the LLM frameworks and extend this toolset even further.

Problematic patterns

What can seem like magic is actually an application of statistics – at their hearts Large Language Models (LLMs) have two central qualities: i) the ability to take a question and work out what patterns need to be matched to answer the question from a vast sea of data; ii) the ability to take a vast sea of data and “reverse” the pattern-matching process to become a pattern creation process. Both of these qualities are statistical in nature, which means that there is a certain chance the engine will not understand your question in the right way, and there is another separate probability that the response it returns is fictitious (an effect commonly referred to as “hallucination”).

Hallucination is one of a number of unintended consequences of AI – whereas a good search on Google or a Wikipedia article brings you into contact with material that may be relevant to what you want to know, it is most likely that you will have to read a whole piece of writing to understand the context of your query and, as a result, you will read, at least somewhat, around the subject. However, in the case of a specific ChatGPT prompt, it is possible that you only bring back confirmatory information or at the very least information that has been synthesized in specific response to your question, often without context.

Another issue is that of racism or other biases. AIs and LLMs are statistical windows on the data they contain. Thus, if the training data are biased then the outcome will be biased. Unfortunately, while many are working on the high-level biases in AI and also trying to understand how one might equip AIs with an analog of common sense, it is not yet mainstream research. More generally, I am quite hopeful that, since the issue of bias is becoming better articulated with time, we will conquer obvious biases that take place in interactions.  However, what is more concerning to me are the cumulative effects of subtle biases that AIs adopt.  The whole point of AIs as tools is that one thing that they do better than humans is that they are able to perceive factual information more completely and with greater depth than a human.  This is why we can no longer beat AIs at Chess or Go (except by discovering critical flaws!), and why we may not be able to perceive biases that are not simple.  In “2001”, the famous Stanley Kubrick Arthur C Clarke collaboration, the AI HAL interpreted conflicting instructions in an unpredictable way. In fact, the rationale for HAL’s psychosis is obvious by modern standards, but it serves as a prescient tale – the outcome of a confluence of a large number of complex instructions could be a lot more challenging to identify.

Fear and loathing in AI

Twenty years ago, we would not have recognised “social media influencer” as a job type and in twenty years from now, we may not recognise taxi driver as a job type.  Prompt engineers may form just one new class of jobs, but how many of us will be prompt engineers in the future? Will prompt engineers be the social media superstars of tomorrow? Will they be the programmers that fuel future tools? Or, like using a word processor or search engine, will we all be required to learn some level of witchcraft? 

Previous industrial revolutions have been responsible for massive wealth creation as sections of society have significantly increased their productivity. But, at the same time, revolutions have also led to destitution in the parts of society that have either lacked access to new technologies or who have lacked the capacity to adapt. As can be seen in Figure 1, following the first industrial revolution in the UK the average GDP per person fell between the 1760s and the 1820s – it was not until the 1830s (50 years after the introduction of the steam engine and a full 80 years after the beginning of the industrialization) that the GDP per person regained and eventually eclipsed its level in 1750. The same trend took place in the late 1860s where GDP per person fell until the early 1900s when, by 1925, it was well on the way to its previous high, again to eclipse its prior peak by 1950. We are currently in a third cycle in which GDP per person is now lower than its 1975 high, initiated by the advent of the PC, followed by the World Wide Web and further fuelled by the rise of AI.

Viewed as a continuum since the dawn of the information age with the widespread introduction of the PC, we stand in the midst of an exponential revolution where the technologies we have built are immediately instrumental in designing and building their successors, superseding themselves over shorter and shorter timescales.

Moore’s Law may have held, predicting the growth in potential of computer chips, through the end of the last millennium. While this power has determined the power and capability of AI up until this point, data  are now the dominant factor in how the power of AI will develop. And, as data limitations come into place, the power of quantum computing may take us further into the future.

If we view AI as the tool that it is, and think about how it can complement our work, we begin to position ourselves for the new world that is emerging.

Despite Elon Musk’s prophecies of doom and even the training of evil AIs, or those which create deadly toxins, most people just want to know if their livelihoods are under threat in the near future. While world domination by an AI is not likely in our imminent future, it is almost a certainty that jobs will change. However, if we view AI as the tool that it is, and think about how it can complement our work, we begin to position ourselves for the new world that is emerging. Progressive researchers and businesses alike will be investing not just in learning how the current technology will revolutionize our world but in continual professional development for their team members as, in an exponential revolution, we can expect to need to retrain not just one or twice in a career, but potentially all the time.

While Web 2.0 in the early 2000s gave us massive open online courses (MOOCs) with the potential to retrain and develop ourselves, it has developed as a tool for companies to deliver information such as safety and compliance training, as well as a tool for us to learn new hobbies.  Importantly, it has broadened access to education by, for example, making college and university-level courses available for free or at low cost.  However, for the majority of people, MOOCs and learning platforms have not been integrated into our lives on a regular basis.  In an AI-fueled revolution, our capacity to learn may become critical in being employable – continuous professional development may be the new norm, with your education even being tailored to your needs by AI.

Who prompts the prompters?

ChatGPT and similar technologies show great promise in a research context. They have the capacity to extend our perception, detect subtle patterns, and enhance our ability to consume material. It is clear that their existing ability to summarize and contextualize existing work is impressive. Several authors have already attempted to add ChatGPT as a co-author, causing some publishers to bring out guidelines that ChatGPT may not be a co-author, since it cannot take responsibility for its contribution. At the same time there are concerns over ChatGPT’s tendency to “hallucinate” answers to questions. Since LLMs are essentially statistical in nature – their answers correspond to the statistically most likely responses to questions – which means that their answers are only statistically based on fact and there is always a chance of untruth.  This means that a literature review written by an AI may be written quickly but it could, at some level, be both uninnovative and incorrect since it is the most statistically likely statements that others might make about the subject.

Such approaches can already be observed as publishers seek to address an onslaught of papers from papermills

This leads to two obvious challenges (and many more besides): One is the bias of research to particular points of view that are reinforced by repetition in the literature; the second, is that AIs are clearly already capable of producing papers that look sufficiently “real” that they may pass peer review. Depending on who is prompting the AI, such an AI-written paper may be designed to contain fake news, factually questionable information, or otherwise damage the scholarly record.  Such approaches can already be observed as publishers seek to address an onslaught of papers from papermills.

On the one hand, LLM technology may be able to perceive patterns beyond the capability of humans and hence point to things that humans have not thought of, yet on the other hand, LLMs also don’t generate true understanding-based intellectual insight, although it can seem to be otherwise.  This leaves AIs at an uncomfortable intersection of capabilities until they are truly understood by their users.

But, as Harari argues, the effect of AI tools on how we interact with each other and how we think has the potential to be profound.  In Harari’s example, it is narrative storytelling that is the critical capability that can undo society.  In research, we need to preserve young researchers’ ability to formulate narrative arcs to describe and relate their research to other researchers – while machine-readable research has been a long-held aim of those of us in the infrastructure community, human-readability must not go away.  Asking an AI to write a summary or narrative for you using the current technology is a way to potentially miss key patterns, key results and realisations.  For a mathematician this is analogous to using a symbolic mathematics program to perform algebraic manipulations rather than having a sense of the calculation oneself.  If one lacks calculational facility then it is impossible to know if the result really is true or whether the machine has made an assumption that is inappropriate.In spite of these challenges, AI tools, used as a set of intellectual augmentation capabilities, hold intriguing possibilities for researchers. AIs already play significant roles in the solution of certain types of differential equations, in calculating optimal geometries for plasma fusion and detecting cancer – essentially anything where human pattern matching is not subtle enough or where one has to carry too many variables in a single mind.

Is the real gain to come from AIs being prompt engineers for us?

We started this blog by considering humans writing as prompt engineers for AIs – but is the real gain to come from AIs being prompt engineers for us? When we stare at a blank page trying to write an article, an AI does an excellent job of writing something for us to react to. Most of us are better editors than we are authors so is one possible future one in which the AI helps with our first draft? Or, does it challenge our thinking if we appear to be the ones who are biased? Is AI the tool that ensures that our logical argument hangs together? Or, is it the tool that takes our badly written work and turns it into something that can be read by someone with a college-level education?

When writing a recent talk, I used ChatGPT as a sounding board – “Would you consider the current industrial revolution to be the third or fourth?” I asked.  And then, “It is generally agreed that the phase the exponential revolution is synonymous with the fourth industrial revolution?” Later I prompted, “Please make suggestions for an image that might be used to represent the concept of community.” The responses were helpful, as though I had an office colleague to bounce ideas off.  Not all of them were accurate but they gave me a starting point to jump off into other searches on other platforms or other books that I had available to me.

One, perhaps scary, notion is that an AI may herald a reversal of the recent trend towards team research and back toward the days of the lone researcher. Over the last half century the world has been moving forward with an increasingly collaborative research agenda (see Fig. 2) – collaboration has become critical in many areas. This need for collaboration often has its genesis in needing resources such as equipment that is too costly for a single institution or single country to hold. But, collaborations often also come from wanting to work with colleagues with particular skills or perspectives. In this latter case, LLMs can be a helpful tool. I have written before that it is a conceit to believe that you can be the best writer, idea generator, experimentalist, data analyst and interpreter of data for your research. But, with LLMs, you may only need to be capable of just a few of these in order to work once again alone.

Regardless of the actual future of AI and its role in research, it is clear that we are living in a world where AIs will take a role in our lives.  Arthur C Clarke once commented that “Any sufficiently advanced technology is indistinguishable from magic” and that is certainly how the outputs of LLMs appear to many of us, even those of us in the technology industry are impressed.  But, there is another side to witchcraft. In John Le Carré’s Tinker, Tailor, Soldier, Spy – Operation Witchcraft takes over the Circus (the centre of the British intelligence service), leaving them completely vulnerable to false information from their enemies.  In what is coming, it seems that we have no choice but to master witchcraft.

Daniel Hook is CEO of Digital Science, co-founder of Symplectic, a research information management provider, and of the Research on Research Institute (RoRI). A theoretical physicist by training, he continues to do research in his spare time, with visiting positions at Imperial College London and Washington University in St Louis.

Read more articles by Daniel on TL;DR.

The post Tinker, Researcher, Prompter, Wizard appeared first on Digital Science.

We hope that it will be a friendly, safe environment where we can encourage diverse but respectful opinions and where we can challenge ourselves to ensure that we become the best supporter of the community that we seek to serve.

I’m looking forward to contributing my own articles to TL;DR in the near future, and in the meantime you might enjoy one of my more light-hearted looks at how a certain figure in popular literature and film makes a regular appearance in research.

Or for a more in-depth analysis of a more serious topic, here are three examples where the Dimensions database has provided useful context to one of the global challenges facing the world today:

The post The importance of adding context appeared first on Digital Science.