This five-chart visual story follows a three-act narrative: first, how generative AI is entering student cognitive work; second, why short-term assisted performance may not fully translate into independent learning; and third, how the infrastructure behind frontier AI is becoming concentrated among powerful institutions and countries.
Generative AI isn’t about robots, it’s about a text box. It has subtly become a way “of doing everyday thinking” and this initial graph illustrates how far its reach has now extended, far beyond writing to summarising, research, coding and study support (Aristovnik et al., 2025).
Bar chart with horizontal bars representing the percentages of higher education students surveyed who use ChatGPT often or always on twelve higher education tasks, ranked by the highest to lowest percentages. Some of the most common tasks include coding assistance, research assistance, and summarising, with a significant percentage of students stating that they always use AI for these tasks. This chart shows that AI adoption has gone beyond writing, with AI now being integrated into other key cognitive tasks such as brainstorming, translation, study, exam preparation, and more.
To help someone perform is not the same thing as helping to learn. In another randomised experiment, high-school mathematics students’ performance was higher when they used AI to practice, but dropped substantially when they were asked to perform on their own without the aid of AI (Bastani et al., 2025). The gain proved to be borrowed, it turned out.
The three lines connecting the Practice performance scores and Independent performance scores on the slopes represent the scores obtained for the different treatment groups: Control, GPT Base, and GPT Tutor. AI-supported groups showed an improvement over the control group during the practice session, while all groups were matched during independent work. The annotated region of the two columns is shaded in to represent the area where the performance improvement due to the AI is not significant, and the point where the two curves meet is marked as “Practice gain does not fully transfer. As can be seen in the chart, though there was an increase in performance outcomes in the short term, there didn’t seem to be a corresponding increase in independent learning ability when using AI.
Attention is another more difficult-to-measure learning foundation. This chart shows that as screens take over more of the day, this is how screen time correlates with attention span and productivity (Zaidi, n.d.) - a snapshot of the attention span and productivity students are now attempting to focus on.
A stacked proportional bar chart of the percentage of individuals who reported their productivity as low, medium, or high across 6 bands of screen time spent per day, from less than 2 hours to more than 10 hours. 100 percentage of the respondents from each band of screen-time are represented in each bar. The higher the productivity, the bigger the share of the bar at lower screen times. The percentage of low productivity increases and high productivity decreases as daily screen time increases beyond 4-6 hours. A drop down filter enables readers to compare this pattern across all activity types, academic screen use, work screen use, and entertainment and social screen use, showing that the loss of productivity as a result of increasing screen time is greater in entertainment contexts than in contexts of academic use.
When you zoom out of the person, another question emerges: Who’s creating these systems? To train a frontier model today requires computing power that most institutions don’t have access to - and this chart shows how such compute has scaled, and who has increasingly provided it (Epoch AI, 2026).
Log-scale scatter plot of significant AI models from 2010 to 2026, where the horizontal axis represents the publication date of the models and the vertical axis represents the amount of compute used in training the models. The points are each individual model, with their colour and shape determined by whether they were in industry (red circle), in academia (dark triangle), in collaboration (green square), or otherwise (grey diamond). Parameter count (if available) is represented by the size of the bubble. While industry-developed models have continued to grow in compute used for training, the gap has been steadily increasing after 2020, as evidenced by the two smoothed trend bands. The following models are directly labelled with their names: GPT-4, Gemini 1.0 Ultra, Llama 3.1-405B, Grok 3, GPT-4.5. Hovering over any point will show the model name, organisation, sector, country, date, number of training compute and number of parameters.
That power is not equally distributed. The tools changing the way millions of students learn and work are being designed by a very few (Stanford Institute for Human-Centered Artificial Intelligence, 2026) - Private investment, data centres and AI compute focus in a few countries. Mind on screen can be borrowed, but the question raised by this story remains of whom.
Grouped dot chart comparing countries with respect to three dimensions of AI infrastructure: AI private investment (billion US dollars); number of data centres; number of AI supercomputers, each measured as an indexed score (0-100) on each dimension, with the three plotted on the same axis. The countries are ranked in reverse order according to private AI investment. The United States is far ahead of all other countries, with scores at or close to 100 on all three dimensions. China is second in investment and second in supercomputers, but not so high in data centres. The United Kingdom, Germany and Canada are at the top level in investment, but have relatively low infrastructure ratings. The dropdown filter feature enables the readers to toggle through all the countries and the top 10 countries for investment. Hovering over any dot will show the country, dimension, indexed score, raw value, and global rank. The cases of AI supercomputer are only regional figures and presented as infrastructure context, not the actual number of supercomputers per country.
Anthropic. (2026). Claude (version) [Large language model]. https://claude.ai
Aristovnik, A., Ravšelj, D., Keržič, D., Tomaževič, N., Umek, L., Brezovar, N., Kessler, S. H., & Schäfer, M. (2025). Higher education students’ evolving perceptions of ChatGPT: Global survey data from the academic year 2024–2025 [Data set]. Mendeley Data. https://data.mendeley.com/datasets/nv2343nwsb/1
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, O., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
Bastani, O. (2025). GenAICanHarmLearning [Data set]. GitHub. https://github.com/obastani/GenAICanHarmLearning
Epoch AI. (2026). Notable AI models [Data set]. https://epoch.ai/data/notable-ai-models
OpenAI. (2026). ChatGPT (version) [Large language model]. https://chat.openai.com
Sajadieh, S., Fattorini, L., Perrault, R., Gil, Y., Parli, V., Santarlasci, L., Pava, J., Maslej, N., Altman, R., Brynjolfsson, E., Brodley, C., Clark, J., Dignum, V., Kumar, V., Landay, J., Lyons, T., Manyika, J., Niebles, J. C., Shoham, Y., … Weld, D. (2026). The AI Index 2026 annual report. Stanford Institute for Human-Centered Artificial Intelligence. https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf
Stanford Institute for Human-Centered Artificial Intelligence. (2026). AI Index 2026 public data and charts [Data set]. https://drive.google.com/drive/folders/1zJTOg0iR0j5SijCwFutwWvDt143lW277
Zaidi, M. A. R. (n.d.). Screen Time Data: Productivity and Attention Span [Data set]. Kaggle. https://www.kaggle.com/datasets/muhammadalirazazaidi/screen-time-data-productivity-and-attention-span
I have used Claude (Anthropic, 2026) and ChatGPT (OpenAI, 2026) in the development process of this assignment for the purpose of structuring my narrative, debugging the R code involved in the visualisations, resolving issues related to Plotly and ggplot2 formats, ensuring that colours meet The Conversation’s brand colour scheme, and improving the accuracy of the annotations and captions in the assignment. The output produced with the help of the AI programs was critically evaluated and edited. All data and visuals used in this assignment are my own work.