Writing a ‘digital horizons’ piece right now is challenging. Things are moving quickly with little certainty. It is hard to know how to orient ourselves, let alone how to navigate through the coming turbulence. If we are to keep steady on the tiller, I’d like to propose we keep in mind some things remain constant, while trying to understand what is driving the uncertainty.
For me there are two forces driving this uncertainty – commercial and technical.
Commercial
On the commercial side, there is an enormous amount of cash being invested into the long-term promise of AI. This causes distortions in the market, and in the economy of attention. AI is seen as an important signal for investors. This has effects that we need to deal with as technologists.
Most companies are shoe-horning AI into every product as a result of this investor and perception pressure, making it hard to evaluate product offerings on their own merits, and increasing the inbound sales noise from partners and vendors attempting to get our attention. Our customers and partners are asking us more questions about our own use of AI. You are probably being asked to fill in AI risk assessments or disclosure documents, and the questions are probably often irrelevant to what you are actually doing, but it’s taking time and attention away from your teams.
There is a drain on specific skills because some sectors of the tech labour market are being distorted. At the same time, we need to respond to the business about “what are we doing about AI” just as the talent to help answer this is becoming scarce. The collective effect of these is that our attention is often being taken up by the side effects of the investment cycle rather than on things that are creating actual value for our customers and our companies.
Technical
On the technology side, there are as many, if not more, sources of confusion emerging as on the commercial side. As powerful as LLMs are, we still don’t know how to best use them. We don’t quite know how they work. I’ll give a few specific examples. There is a lot of justified interest around how to use LLMs to accelerate the software development life cycle, but the tooling to support this evolves every few weeks. There are no consistent patterns on how to use these tools at team level rather than at individual level. This means that we need to experiment around the software development life cycle (SDLC). It’s always better when the SDLC is stable because otherwise time is spent on the process rather than on delivery. In fact, there is a sentiment that a noticeable inflection point has been reached as of November 2025 with the release of GPT 5.2 and Claude Opus 4.5 that could significantly disrupt software engineering practices in the medium term, but large organisations with old code bases and embedded engineering practices are going to struggle to get maximum benefit from these tools.
Another example is that there is a large amount of potential with agentic AI (where LLMs are granted access to tools and permissions and are allowed to work autonomously on tasks) however, today, it is not known how to effectively wire together multipage agents. Research shows that different architectures can have up to three orders of magnitude impact on the effectiveness of the agents. Given this is an unsettled question, how can we confidently bet on any specific agentic AI platform offering?
I think the most important uncertainty however is that we have not yet developed an understanding of the best way to integrate these tools into the user experience of our products. Their profligacy of output means that we can easily balloon the amount of content, features, or narrative that we present to users. Adding features is easy! Adding a chat interface – easy! RAG your archive – easy! But are any of these initiatives helping our customers towards task completion?
What AI can do…
So, how might we think about how to navigate all of this? I like to build a clear picture of what these technologies can do and how that can help our customers. I also like to think about what won’t change in terms of customer need, or the constraints that we operate under.
To help us think about what these technologies can do that are new, let’s play an analogy game. Here are some brief mental models – all partially true, about what these technologies can do.
- The calculator for text: they can operate on text given human language as input. I use this all the time to help with day-to-day tasks, getting the gist of meetings, summarising email threads, reformatting lists into tables, and vice versa. You can use them to apply tone of voice, style guides, or have a crack at doing the best with fuzzy instructions. We have just never had that ability before, and it’s coming to images, audio, and video too.
- They are intention machines. This is spooky, weird, and partially problematic. You can ask them to do a loosely defined task, they will try to do it. Try really hard. Their ability to mostly understand your intention can be astonishing, even when what you ask for is not tightly specified. This is an entirely new way for us to interact with computers. However they may not try to complete your task in the way that you think they will. At BMJ, we turned on an MCP server in our content store and hooked an LLM into it. One of our product folk asked what day does the BMJ mostly publish articles on? We thought the LLM would just use the index in the content store to get the answer – it was right there. Did it? No, it decided to start scraping every page of every article of our website to read what days articles were published on!
- They are emotional, emotive, machines. Because we anchor ourselves so much to the world through language, having a device that can talk back to us creates the perception of connection. This is wildly weird and new. We are used to thinking of technology mostly from a rigid point of view but these systems can evoke emotions within us. It creates opportunities for radically new experiences, but if we don’t understand the emotional impact that these technologies can have, we will be laying down risks for our users.
- They are Jevon’s engines. Jevon’s paradox is that when a resource becomes more efficient, instead of meeting the expected demand, more demand emerges and the resource gets used more. That’s where we are now with machine intelligence. It has become ubiquitous and can be delivered at close to the cost of electricity.
- They are amplification machines. LLMs create content with a velocity and volume that we have not had access to before. If you dabble in software projects, or you like to generate images, or you like to play with music, well now you can go back and finish off all of those half complete ideas. What about adding an entire new data model to your product, or creating a Hindi version of your main publication, or auto-creating a podcast, yup, we can do all of that now. But this capability bangs right up against some of the constraints that we have to work with, so let’s turn to thinking about what they can’t do.
What AI can’t do…
They can’t think, even though they appear to present as thinking; it’s really important to understand that what is going on under the hood is a set of static operations on a very large static matrix of numbers. What this means is that they cannot make decisions for you.
They cannot be intrinsically trusted, simply because they are not thinking entities. This means they can present incorrect information credulously, and they can be tricked to steal data from you, so you have to design with security first.
But I think the most important limit that they have is that they cannot create time. No matter how efficient they are, they do not add a single second to our lives, and by creating more content at more volume than ever before, they can easily put pressure on the finite attentional constraints that we all work within. If they make some part of your process better or faster, is that just moving the bottleneck elsewhere?
Impact on scholarly publishing
Let’s look now at some more concrete examples of all of this. While scholarly publishing has a somewhat different value driver from other publishing sectors, the above trends impact us just as much. I want to describe now some of the impacts on our sector and how BMJ is responding to them.
- Content discovery: We have a lot of content that covers a very large variety of specialised topics, and that creates a real discovery problem. Finding the right article among millions is genuinely hard, both for readers and for the systems trying to surface relevant work. LLMs are making it easier to apply highly refined labels to our content and with a partner, we have done this in a cost-effective way. This also allows us to segment our audiences better and this improved segmentation has seen us double the advertising CPM for well categorised audiences.
- Diversity of voices: Another area where LLMs are beneficial is around helping researchers write. There’s a bias problem in scientific publishing with non-native English speakers facing reduced acceptance rates. That’s a diversity and equity issue as much as it’s a quality issue. We’ve developed policies around AI use in submissions specifically to address this – allowing authors to use AI tools to improve their language and presentation. We hope this will increase both acceptance rates and the diversity of voices in our journals, which feels like the right outcome. We chose not to create that tooling ourselves as it is clear that authors can use readily available AI. What’s important is that its use is transparent and that we don’t put a blanket ban on our authors using these tools.
- Simplifying submission process: Journal submission requirements are incredibly complex, and the user experience is often a nightmare. Authors have to navigate formatting requirements, metadata fields, supplementary file specifications, and compliance checklists that vary by journal. It’s an area where we know we’re creating friction, and it’s ripe for improvement. LLMs are making it much easier to auto format submissions and we are working with partners and vendors to trial tools to help streamline the submissions process.
- Software development: Internally, some of our development teams are getting real value from AI assisted coding, but we are mainly seeing this in areas of our codebase that are well structured, have good tests, and have moderate to low levels of technical debt. Recently, a team shipped about four weeks of work in one day, and the same team has helped us avoid up to 600k in licensing fees by building out some software that we would previously have bought. We are not seeing this level of impact in our older code bases yet, but I expect this to change in 2026.
- Product development: We are starting to look at bringing AI powered features into our own products. I’d say it’s early days for this and we are treading carefully because we are dealing with systems that involve highly trustworthy information, so we need to be certain that these kinds of features don’t distort behaviour. A good example of where this is working well is in increasing the number of questions that we have in some of our online learning tools. Question bank development is now faster than it was before, and it can flow through the same editorial quality checks that we had to build before, ensuring that these questions meet the quality standards that we demand.
- Peer review: Finding peer reviewers is one of the most difficult challenges in scholarly publishing – it’s time-consuming, the pool of willing reviewers is shrinking, and matching expertise to manuscripts is genuinely hard. AI might make this easier, both in identifying suitable reviewers and in improving the consistency of reviews once they’re written, as well as taking pain out of some of the more manual steps of peer review. There are review steps that previously would have been too expensive to do for every submitted manuscript, but AI now makes that possible; for example checking the details of the statistical methods and reporting standards. We’re developing a multi-agent system right now that can help with these questions in collaboration with AWS, and it’s showing promise, but a generated report can often be longer than the submitted manuscript!
- Fake papers: It’s not all upside. Other publishers are seeing a large uptick in junk submissions and many publishers are having to change their submissions policies in reaction to this. One of the weak links we have in scholarly publishing is that we don’t have any known customer protocols, so it’s possible for fake papers from fake authors to be submitted, and we take on some burden to weed these out. When these papers were created at human scale, this was something that we could do with existing processes, but now papers like this can be created at industrial scale and so we have to invest in new systems to tackle this rise in fake papers.
- Content transformation: One thing we have not experimented with publicly yet at BMJ is using LLMs to create radically different versions of research articles. Some publishers are auto-creating podcasts with highlights from the research article, others are auto-creating videos and multilingual summaries, so this kind of content transformation is clearly possible.
We can now create new experiences around our content, change how we serve our customers, but there are constraints that no amount of AI can remove. AI creates no more time, and may put increased pressure on the fixed amount of attention that exists in the world. AI is very good at helping individuals, but team level systems are far behind and we still need to work in teams; AI can’t be trusted on its own, we can’t delegate our decisions to it. Our budgets don’t magically grow, we still have to pick what to invest in. So I think it’s important to keep the fundamentals in mind.
Keep understanding your customer needs and how what you do creates value for them. Can AI help that? They really don’t want to hear about your new AI product or service, but if your tool saves them time, or gets to the goal more accurately, that they will care about. Give your teams the time to explore these technologies, but we should be doing that in any case, because if we are not learning, we are falling backwards. Invest in making your data and systems as clean, connected, and computable as possible. This will yield bigger benefits from AI, but it will yield benefits nonetheless.
It’s an exciting, somewhat exhausting, certainly frothy, time to be working in tech, but by sharing what we are finding out, we have a real opportunity to create better systems and harness this incredible technology to the benefit of our customers, and dare I say it, for society at large.
This article was first published in InPublishing magazine. If you would like to be added to the free mailing list to receive the magazine, please register here.
