Akshat Mandotra, Arnav Dixit, Hrishikesh Tule, Mayank Gomekar and Sriram Namana. Class project “Real, Synthetic, or Both?” for Prof. Srinivasan Radhakrishnan. Cell images courtesy of the U.S. National Library of Medicine. In the eyepiece, an azure halo marks a real NIH cell and an eosin halo a ChatGPT cell.
A thin-smear malaria test is read by eye. The microscopist moves across the smear field by field and looks at each red cell for the purple ring of a parasite. To teach a model the same first pass, someone with that training has to label thousands of cells, and in malaria regions those people are scarce.
ChatGPT can draw an infected cell in seconds. A malaria programme therefore has to decide how many real labelled cells it still needs to pay for.
Decision maker
The head of laboratory services for a national malaria programme, deciding how many expert-labelled cells to commission before building a screening aid.
A microscopist reads the smear in a zig-zag. Here the frame walks over real NIH cells.
Brief: Applied Deep Learning and Generative AI for Managers, IIM Ahmedabad. The decision framing is ours.
The NIH set has 27,558 single cells from Giemsa-stained thin smears of 201 patients. Most patients give both infected and uninfected cells, so a random split would put cells from one patient’s slide on both sides.
We locked 50 patients away first (dark dots), then took at most six cells per patient per class: 400 cells per class for training and 100 per class for the test, with near-duplicates removed.



Every image, real or ChatGPT, goes through the same steps. We pad with black instead of cropping because the cells already sit on black and a crop can cut off a parasite at the edge.
Rajaraman et al. (2018), PeerJ 6:e4568. NLM Malaria Data, a U.S. Government work. Patient ids from NLM’s patient-to-cell mapping files.
Both classes used this template. Only the class sentence changed, and grid k of each class got the same variation phrase. We threw out uninfected tiles that showed a parasite-like dot and near-duplicates. ChatGPT refused no prompt.

One untouched ChatGPT grid, parasitized class, 1254 × 1254 px.
All 55 prompts, the references each one used, conversation links and every discarded tile: docs/generation_log.md.
Eight of these are real cells from the training pool and eight came from ChatGPT. Half of each group are infected. Pick yours, then reveal.
The machine’s score
We trained a small classifier to separate our real training cells from the ChatGPT tiles, with five folds grouped by patient and by grid. It gets them right almost every time. The ChatGPT cells share a house style: a more lilac tint, cleaner edges and tidy rings.
We found no copies. Every tile is further from its nearest reference than real cells usually are from a cell of another patient.


Spot-the-fake: ResNet-18, 5 epochs, 5-fold CV. FID: Inception-v3 pool features (torchmetrics). Copy check: cosine similarity of ImageNet ResNet-18 features.
Every model trains on exactly 800 cells, half of them infected. We step the mix from all ChatGPT to all real in ten steps and train each step with five random seeds.
Nested draws
Within a seed, the 8 real cells of the 1% model sit inside the 16 of the 2% model, and so on up to all 800. Any gain therefore comes from the extra real cells; the ones already there stay the same. Each seed draws a fresh nest.
Small dots: the five seeds. Line: their mean. Dashed line: the bar, the all-real baseline minus the 0.01 tolerance.
Seeds
Tolerance
All ChatGPT
results/results.csv holds all 50 runs; results/mix_table.csv holds this table.
4,000 bootstrap replicates of the saved predictions: the 200 test cells resampled within class and the 5 seeds resampled, baseline, bar and r* recomputed each time.


Grad-CAM on ResNet-18’s layer2 (12 × 12 grid), averaged over the 5 seeds of each model and laid over the 96 × 96 cell. Warmer means the region pushed the model towards “infected”.


To the head of laboratory services
Real cells needed = 800 × r* plus the 40 reference cells shown to ChatGPT. Labels avoided = 800 minus that.
All the cells come from one hospital’s slides and one phone-on-microscope setup, and ChatGPT copied that look from our references. The next test is to score the same models on cells from another site, such as NLM’s other smear sets or a partner lab, before trusting r* anywhere else.
Our fakes came from real references. If the next generator learns from images like ours, its mistakes, such as the too-tidy rings, get copied again and the rare real cases drop out (Shumailov et al., Nature 2024). Every cycle needs fresh real labelled cells.
How we chose the dataset
We piloted concrete cracks and dropped them because published models score about 99.9%, so the curve would be flat. We piloted paediatric chest X-rays and dropped them because ChatGPT’s X-rays were stylised and all alike. Malaria cells passed both checks: ChatGPT drew them well, and a real-only model scored 0.98 on held-out training patients, which left room to move.
AI use
The synthetic images come from ChatGPT Create image, as the brief requires. Code, analysis, this site and the report drafts were produced with Claude (Anthropic) under our direction. Every number on these pages is generated from the result files by scripts/10_build_site_data.py.
Re-measure the break-even at every new site before trusting it there. Thank you.