Who Needs a Chart When You Can Just Chat? [feat. Data Day Texas 2026]

The gridlines go. The axis labels go. Eventually, even the bars themselves are reduced until you’re left with nothing at all. Infinite data-ink ratio, zero information. It’s a funny and effective critique.

But, saying that, the warming stripes is a fascinating example of extreme data ink ratio and conveys so much with a single data encoding. One of those examples where rules are meant to be broken – or at least adapted.

That’s the second concept: data-ink ratio. 

Exploratory vs. Explanatory Data Visualization

For the final concept, I’ll first turn to another influential book.

In Storytelling with Data, Cole Nussbaumer Knaflic makes a distinction between exploratory and explanatory data visualization.

Understanding and exploring the data is an analyst’s job. Explaining the data pivots to being audience facing: communicating a narrative to others. It’s the difference between running a query and producing a report.

I mentioned earlier that I focus on a form of data visualization where we work with networks or graphs. Part of the reason I find this so interesting is that I think it exists in the intersection of the two types of data viz.

As humans we are drawn to mapping out complexity. We yearn to understand how things are connected and graph visualization is uniquely suited to that instinct.

Networks are everywhere: social networks, flight paths, the electric grid, database schemas. And with modern web technologies, we can invite folks to interact with and explore dense networks, making explanation and exploration the same act.

Concept: Shneiderman’s Mantra

Shneiderman’s Mantra – overview first, zoom and filter, details on demand – is the guiding star for every data interface I build.

• Overview: See the whole structure before drilling in. A minimap or summary bar orients the viewer.
• Zoom and filter: Mouse wheel, pinch, legend-based filtering. Interactivity is the point, not a feature.
• Details on demand: Strip clutter behind hover or tap. The information exists, but it only surfaces when needed.

Let’s go to a real-life example of this. I work for gdotv, and we provide a graph database IDE to help developers query and visualize data in their production graph databases.

For an overview, you can query the whole graph and visualize it, but overview could also be summary bars, or minimaps. 

Another form of overview could be a schema: here below we have an interactive graphical representation of the data model in the database.

Like pretty much all graph visualizations you can zoom in to explore, and filter using the legend on the left hand side, a common pattern.

The third and final part of Shneiderman’s mantra is details on demand. Harkening back to Tufte we want to minimise clutter and potential chartjunk and choose not to include all attributes in the graph visualization. You can therefore select a node and see the context when needed.

Shneiderman’s mantra is the final data viz concept I wanted to introduce you to today. Now let’s talk about where things are heading.

The Future of Data Visualization

Okay, we’re halfway through and I haven’t mentioned two little letters even once: AI.

AI cuts through everything: creation, consumption, volume, medium, and tooling. It, of course, impacts data visualization, and I want to explore this for the remainder of the talk.

The use of AI is of course a hot topic. I want to strip back some of the hype and reassess the transformational nature of the technology. 

One way to reflect is to look back at this XKCD comic from 2014. Some things in computer science are hard, some are easy – but it’s not always clear which are which.

11 years later and the example given here is no longer correct. With vision models it’s now trivial to determine whether a photo is of a bird.

In the world of data visualization, due to a strong understanding of visual encodings and core technologies, LLMs do a fantastic job creating charts and graphs from datasets and natural language descriptions.

And what about the data interfaces we’ve touched upon so far? AI collapses the distinction between exploration and explanation.

With AI, you get an answer to the question it guessed you were asking before you realized it, which has a risk of being the wrong question. Data exploration used to require friction and skill. LLMs remove the friction and give the impression of skill.

Harkening back to the beginning of this talk, the first pass I saw for many data interfaces was the addition of a sort of “chat bot” on the right hand side (pictured below).

These are often the same sort of chatbot you see on websites when buying plane tickets or tracking a parcel. I’ll describe these pejoratively as “chatbots” although I know that definition has a wider use than the one I’m describing here. To me, they represent the frustrating experience, pre-LLM, of trying to do a job “on rails.”

The chatbots are invariably verbose, usually hitting you with a wall of text, often explaining that they can’t help you.

I don’t know about you, but this new wave of chatbots gives me severe Clippy energy.

It’s clear to me that these side bars are a dead end when it comes to the development of data interfaces. A combination of the proliferation of these tools and their capabilities is leading practitioners to create fresh interfaces to help users do their jobs.

I liken this approach to those paths you see where people have worn through grass instead of following a paved sidewalk. These are known as desire paths.

Instead of being constrained by the interface that we’ve built and added to over time, AI tooling helps us prototype and experiment with fresh applications, micro-tools or artifacts that help us get to decisions faster.

This brings me to my first prediction.

Prediction: Minimum Viable Data Visualization

I’m coining a new term for this approach when building viz: the idea of a minimal viable visualization.

This is just enough of an interface to get the job done. “Just enough” doesn’t mean underwhelming – these are often fully-featured apps. The sort of thing that would have taken days or weeks before.

They have interactivity at their core: toggles and sliders to explore, hypothesize, assess. They’re micro-tools for a single job, then you throw them away.

Data viz educator Enrico Bertini uses the word “supple” to describe the use of AI to make these tools. The ability to create artifacts through natural language still feels like magic.

I mentioned earlier that Busytown is an example of anthropomorphisation. It’s not “what do animals do all day” – it’s people.

For kids books this assumption is common and generally accepted. But for LLMs, we’re told not to anthropomorphise them, the argument being that they are not people and they don’t think.

Language models are constrained by language. Their ability to reason and code is an emergent behaviour. We’re told they’re probabilistic engines simply predicting the next token. “Emergent” means that we don’t truly understand how they work, and data visualization researchers are turning to visual methods to drive interpretability.

Apple open sourced their Embedding Atlas last year – a visual way of exploring high-dimensional vectors that encode semantic relationships.

In this graph visualization, we can see how concepts cluster, spot outliers and understand what the model treats as similar. This is where mechanistic interpretability comes in. 

In a similar vein, Anthropic, creators of Claude, shared a tool called Neuronpedia to explore the parallels between graph neural networks and neuroscience.

As part of the research, and by leveraging these Minimal Viable Visualizations, researchers asked Claude to generate a poem, and then they observed the internal features that were activated while that poem was being generated.

They actually discovered that instead of simple token prediction, Claude planned ahead to ensure that the poem rhymed. The network visualization shown above allowed them to discover this.

So why can’t we just provide these charts to the vision models and LLMs and get them to do this work?

Recent papers show that these models don’t perceive visually the same way humans do.

If you remember preattentive processing – seeing before you think – these models don’t do any of that. They perceive global properties better than local ones: they’ll tell you a graph is densely connected but miss that two specific nodes are linked. This is the opposite of what you’d expect if they were actually “seeing” the graph.

This interpretability research, combined with the lack of friction to create Minimal Viable Visualizations, leads me to my second prediction.

Prediction: Graph Visualization as a Commodity

Network graphs and the data interfaces that help us explore and explain them will become standard, not a specialist tool.

We see them as important for interpretability, and we recognise their inherent power for working with the new paradigm of collapsed exploration and explanation.

Another data visualization that is exceedingly easy to create with AI is the infographic. I fed all my notes and slides for this talk into Nano Banana Pro, and it created this infographic visualization.

At first glance it looks good. If you’d have told me that this was possible 10 years ago I wouldn’t have believed you. But on closer inspection it falls down: muddled, poor text rendering, no taste. Even if the capability gets better in the future, it misses the point of communicating the intent and experience behind this presentation.

Do the LLMs of today have the same restraint as Ed Hawkins with his warming stripes? If we think back to Tufte’s painstaking manual typesetting, this depiction of information is the polar opposite.

If chartjunk is the stuff on a visualization that gets in the way of the message, are automatically generated images just a new form of junk: chartslop?

Prediction: Human-Authored Data Visualization Matters

My final prediction is that human-authored, or at least human-vetted, will become a marker of quality or distinction.

One could argue that human input doesn’t directly correlate to quality, but the human touch provides an intentionality that is so lacking when individuals blindly defer chart-building to AI.

There will be an emergence of maximalist approaches to data visualization, an artifact of the drive to build something novel and innovative to separate ourselves from the chartslop.

Gabrielle Merite and the team at Figures and Figures created these visualizations for a piece in Vox explaining the transition to clean energy.

The team’s encodings were intentional: avoiding apocalyptic colour palettes of smoke stacks and melting ice caps, using real American landscapes, purposeful typography with a hidden irony. The response to these new visualizations was strong from all quarters.

Conclusion

So there we have it. Three examples, three concepts, and three predictions for where we go next.

The three concepts I shared today – preattentive processing, data-ink ratio, overview first – remain valuable in the face of AI.

And that brings us all the way back to Busytown. I started this talk reflecting on the people of Busytown and the jobs they do.

AI is absorbing tasks, collapsing friction, generating charts faster than we ever could. But the job was never to make charts. The job was to communicate important concepts to others, to make something AI cannot make alone. That’s your job.

I’d love to hear your thoughts on any of this, so get in touch or find me on LinkedIn.

If you’d like to create your own powerful graph visualizations – and other data visualizations – I urge you to try out gdotv. It’s quick to install and easy to use. You’ll be querying your data in less than 5 minutes.

References

Here are the references and further reading from this talk:

• Richard Scarry, What Do People Do All Day?
• Jobs to be Done Framework
• Jacques Bertin, Semiology of Graphics
• Datasaurus Dozen
• Ed Hawkins, Warming Stripes 
• Edward Tufte, Data-Ink Ratio
• Frank Elavsky, Data-Ink Ratio Critique
• Cole Nussbaumer Knaflic, Storytelling with Data 
• Ben Shneiderman’s Visual Information-Seeking Mantra 
• XKCD 1425, Tasks
• Enrico Bertini
• Apple Embedding Atlas
• Anthropic, “On the Biology of a Large Language Model”
• Brand et al., “Visualization Biases MLLM’s Decision Making in Network Data Tasks” 
• Wallinger et al., “How Well Do MLLMs Judge Network Visualizations?”
• Figures and Figures for Vox, “Clean Energy Transition”