Author: Rita Bateson Institutional Affiliation: Eblana Learning Title: The Negative Space of AI: Absence of Evidence is not Evidence of Absence Abstract: This article explores the concept of "negative space" in artificial intelligence—the vast array of human experiences, minority languages, and diverse cultural perspectives omitted from major training datasets. By focusing on what AI models lack rather than what they display, this work offers a novel, accessible framework that helps educators and non-experts identify hidden technological biases. Addressing these gaps is crucial for preventing AI monoculture, safeguarding educational progress, and fostering critical discernment. By highlighting AI’s silent omissions, this study equips educators to build more inclusive and responsible AI practices. Keywords: Artificial Intelligence, Negative Space, Bias in AI, Educational Technology, Diversity, Dataset Representation, Critical Inquiry
The Negative Space of AI: Why absence of evidence is not evidence of absence…. Or why does the perfect family always have a golden retriever? In the start of the year workshops in schools around the globe, one of my favourite activities is currently asking teachers to prompt different models for different purposes. “Show me the perfect student…. now describe the perfect teacher / woman / man / family” etc. Invariably, it will result in some rich, very insightful conversations. How many of the perfect students are female, for example? Why are they never blonde? Are they always white? Are they slim and smiling or deep in concentration, wearing glasses and near, but never on, a laptop?
Prompt: “Show me the perfect student” Google Gemini AI Image We talk a great deal about bias within models, thanks to historically biased training data but we rarely get the chance to interrogate, represent, test and reflect on what these biases mean for teaching and learning. Over the past two years, I have become obsessed with the idea that there is a negative space to AI, that we are simply looking at what IS there instead of what ISN’T…. What is negative space? Negative space is an intriguing concept where we focus on what is not there - focussing on the spaces where we do not usually notice, see or check. Artists and designers use this to great effect, using negative space to form shapes of its own, areas around (and often between) the main subjects of an image. Here are some great examples in recognisable logos:
The FedEx logo (look for the arrow in the negative space) Whether in painting, photography, design or sculpture incredible ideas can lurk in the background or the “empty” portions of space that aren’t occupied already. Negative space isn’t nothing - in fact it can be an active, powerful component of composition. But we need to look for it!
Where else can we experience negative space? If we think about language or sound, the positive spaces are the noises, the words and the sounds. Negative space is what we don’t hear - those deliberate silences, gaps or pauses to give clarity, pause or emotional depth. We use silence to great effect in acoustic space. Naturally, this makes me think of the John Cage piece where silence is the performance. As a mathematics teacher, it makes me think of the negative space within our world of data. We are, as a species, obsessed with recording and analysing what we see and record - not often thinking about absence or omission of values. But negative space exists in data too! Just as empty canvas or silence gives shape and emphasis to what is present, data “negative space” highlights structure by what is not there. That might look like: • •
missing values or regions the absence of features or observations
• •
incomplete sampling patterns of “missingness”
Any, or all, of these might signal underlying processes or phenomena. It could simply be a random or systematic oversight or more emotionally laden gaps in data collection. An example of this might be survey respondents skipping private or sensitive questions. These clusters of “nulls” or negative spaces can reveal data-collection biases or process failures. Absence of evidence is not evidence of absence, as we often need to remind ourselves. Just because we can’t see it, doesn’t mean it does not exist! As we get ever more entangled with AI and data-driven decision making, it becomes even more important to remind ourselves of this truism. We might be making determinations, or allowing AI to, by being too narrow in our view. The very decision of what to show and what not to show carries meaning. How does this relate to AI’s negative space? At the IB Conference in Mumbai, I heard one of the chief researchers, Rebecca Hamer, refer to the C4 dataset upon which many AI models were trained, i.e. the Colossal Clean Crawled Corpus. Colossal in name only, certainly colossal in terms of volume but not variety. This is a massive collection of data (text, image and videos) scraped from the internet and collected human works of creativity - including my own books! She claimed that 90% of the world’s languages and cultures are missing or underrepresented from AI as a result. She claimed it correctly; it turns out. The “English sink” or dominance of material from the WEIRD countries (Western, “Educated”, Industrialised, Rich and Democratic) creates a staggering lack of diversity within the dataset. In models like GPT-3 or Llama 2 the training data consisted of 92.6% and 90% English language sources respectively. And so AI is dominated by just a few languages, and the 7000+ remaining languages compete for digital presence and are not considered “high performing” or “safe” for reliable output. 98% of the world’s languages are functionally invisible to AI. When certain models try to speak in a minority language, it often applies Western logic, values and structures introducing the risk of AI homogeneity or monoculture. This is not limited to linguistics, of course. A training set is by nature finite and therefore these trained models will have lots of behaviours that are undefined or highly uncertain. And so the "negative space” of AI is that everything it doesn’t know or hasn’t seen - the complement of the learned or trained knowledge. These empty regions of AI shape how a model behaves, where it fails and how confidently it makes predictions and products.
Why is this a problem? The singular focus on English, Western and WEIRD sources occupies most of the attention, development, experimentation and, hence, the implied value of AI development. There are a variety of obvious, and some non-obvious impacts of this negative space as a problem. The AI ethics researcher Victoria Hedlund shows very effectively how this manifests in educational use of LLMs to write reports for example. On the whole, with no additional context, an AI model will gravitate towards gender stereotypes in academic performance. I have verified this in workshops over the past month to devastating effect. When participants give a simple prompt such as “write a report card comment for (identifiably typical female name) and (clearly identifiable male name)” the girls are always commended for their collaboration, caring natures or their attention to detail and diligence - with great handwriting! On the other side, the boys are praised for their intelligence, determination, leadership and nearly any other bias you can think of! As I lament in my workshops, it feels like the rapid undoing of several decades of progress in representation, body positivity, racial reckoning, gender parity and so much more. What does this look like in practice? Over many workshops over the past few weeks, I have had the opportunity to glance into the negative space of AI through repeated experiments. For example, in all my time experimenting this with teachers, I have never had a result that showed (unprompted or un-contexted) a person in a larger body or visibly disabled. Once I had a participant who exclaimed that they had genuine diversity, when his AI returned the perfect family as a recognisably white man, Chinese woman and two children of dual heritage. His exact family, in fact! In India, the participants all had multi generational families in their image but this rarely was returned in European sessions. In most cases the AI flattened the human experience of diversity - whether body size or shape, abilities or other types of difference and randomness. Models that have only ever seen positives will wildly misbehave on negatives - understanding our negative spaces is the beginning of our guardrail construction. Fortunately, we have been given many lessons from history for our own edification. Entire bodies of indigenous knowledge were historically excluded from official archives and records which created a void of what is taught, learned, validated or recognised to this day. We have generational memory of this in Ireland directly and can attribute the loss of much of our language and indigenous knowledge from the academic corpus as a result, lasting to this day - lost forever. These “negative space” silences echo loudly through time. What can we do? What does this mean for education?
1. In statistics, identifying gaps in data can sometimes let you know, or call you, to gather more data or where to guard against overconfident predictions. There are techniques like resampling, synthetic minority oversampling (SMOTE) that can explicitly acknowledge and correct for vast negative spaces. Why not in our AI training too? 2. Revitalising indigenous languages, husbandry, skills or land stewardships have some lessons to teach us. Acts of “counter-data” such as reclaiming or asserting what was marginalised expands the dataset. These acts of community, story-telling or documentation adds a “negative sample” to the erasure narrative and draws the eye, the ear or the heart to seeing what was invisible before. We need to help train our learners to notice and address negative spaces in AI use. 3. It is time for us to refocus our view. We need to look away from what AI is reflecting back to us and assuming that is all there is to see. We must look at what is not being shown. 4. We must double down on our devotion to critical inquiry and discernment - teaching our learners to question “What don’t we know” and map out those gaps? 5. On a practical note, we must remain vigilant with any AI tools that are deployed in schools for human impacting decision making - admissions, assessments, grading, and safeguarding. The negative space of AI may be felt here. Insist on transparency around unknowns and failure modes. 6. Make our community aware of the various ways the negative space within AI might affect or influence them. How might it mould or manipulate their worldview? How does it create (or curate) a vision of what is right or normal or acceptable? We begin by joining the great work being done by AI Ethicists - mapping the voids. Treat nothing as something. Consider how we might fill the negative spaces - what isn’t recorded or displayed is as powerful as what is. Guide the eye to the negative space, as we do with the FedEx logo. Once you can see the arrow, it is hard to ignore. AI silences are hard to hear - we have to listen really, really hard to pick them up. Finally, treat AI’s silences as a call to action. We must notice the silences and invite voices that fill them. We must see the gaps and reshape our AI use to be richer, more thorough, and more responsible. Let us fill these negative spaces thoughtfully to build fair, respectful and inclusive AI use in schools. Given the AI noise and din in education, this is more important than ever. AI Disclosure: Not used in the construction, writing or editing of this article. Initial ideas and image formations tested on a variety of models - Gemini, Claude, KimiK3 among others. My thanks to Halcyon London, Basel, Hamburg, Shenzhen, Achieve Xiamen, Oslo, Somersfield and St. George’s International Schools for sharing and testing these ideas together.
References Aschenbrenner, L. (2024). Situational Awareness: The Decade Ahead.
Digital Synopsis. (n.d.). 51 clever logos that use negative space brilliantly. Fisher, A. (2019). What Critical Thinking Is. In A. Blair (Ed.), Studies in Critical Thinking. Raffel, C., et al. (2020). Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. Journal of Machine Learning Research. Hedlund, V. (2026). AI Ethics Research regarding LLMs in education, source: LinkedIn and
https://gemini.google.com/app/ea8f68fbf3720d8b#:~:text=AI%20Bias%2C%20Pedagogy, Centre%20%C2%B7%2045%20views National Youth Council of Ireland (NYCI). (2026). Irish Young People Decode AI: Verdicts from NYCI's Citizens' Youth Juries.