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.
Malaria thin-smear cells from 201 patients (U.S. National Library of Medicine). 50 models, each trained on 800 cells, all tested on the same 200 real cells from 50 patients none of them saw.
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 a convincing cell in seconds: for this project one person made . So the programme 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 maker is our framing. Generative AI makes the training data here; the screening model itself is an ordinary image classifier.
The NIH dataset has 27,558 cells from the blood smears of 201 patients. Cells from one patient look alike, so letting a patient’s cells into both training and testing would flatter the model.
We locked 50 patients away first (dark dots) as the final exam. The model trains on 800 cells from the other 151 patients and is tested once on 200 cells from the 50.



Every image, real or ChatGPT, is resized the same way to 96 × 96 pixels, so the model cannot tell them apart by size or file format.
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.
Prompt design mattered. Both classes used this template, so the model cannot learn “ChatGPT style” as a shortcut for one class. Only the class sentence changed; the variation phrase rotated. We checked every tile and threw out “healthy” cells that ChatGPT drew with a parasite-like dot, plus 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 built ten training sets, from all ChatGPT to all real, kept everything else identical, and trained each one five times to average out luck. Every model sat the same exam.
Real cells are added, never swapped
Each bigger real set contains the smaller one, from 8 cells up to all 800. So a better score comes from adding real cells, and not from drawing luckier ones.
Small dots: the five models at each mix. Line: their average. Dashed line: the pass mark, all-real minus 0.01.
Misses cost more than false alarms
Where to stop
How to read it
What it changes
Rule: flag when P(infected) × cost of a miss > P(healthy) × cost of a false alarm, with the model’s probability re-weighted from 50:50 training to the expected share of infected cells.

The lesson for managers
A generator draws its idea of a typical case. The rare, borderline cases a screening aid most needs to catch are where it falls short. Check generated data for coverage of hard cases, and not only for how real it looks.
Lessons for managers

To the head of laboratory services, national malaria programme
Real labels needed = 800 × 30% + the 40 reference cells shown to ChatGPT = 280. Labels avoided = 800 − 280 = 520.
Every cell came from one hospital and one camera. Our stain test says 30% holds, but it is a simulation. Manage it: test on a few hundred local real cells before go-live.
ChatGPT draws textbook parasites and missed the faint ones on the rim. Manage it: never train on fakes alone, and review every infection the model misses.
We uploaded 40 public, de-identified cells, and ChatGPT copied none of them. Manage it: with hospital data, check consent and the vendor’s data terms before uploading anything.
If future generators learn from images like ours, the rare real cases fade out (model collapse, Shumailov et al., Nature 2024). Manage it: add fresh real labels at every retraining.
Use it as a screening aid: a microscopist confirms every result. Regulators such as the US FDA treat diagnostic software as a medical device and expect retraining plans to be agreed in advance.
Thank you. Backup slides follow for questions.
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.
Seeds
Tolerance
All ChatGPT
results/results.csv holds all 50 runs; results/mix_table.csv holds this table.
Heatmaps from Grad-CAM, a standard way to see which part of an image drove a model’s answer. Warmer means the area pushed the model towards “infected”. Averaged over five models each.


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; the team divided the work by stage and checked samples of the images by hand. Every number on these pages is generated from the result files by scripts/10_build_site_data.py.
Option A ·
expert labels
Option B ·
expert labels
“Chance of passing” is the share of 4,000 re-runs (test cells and training runs resampled) in which the mix stays within 0.01 AUC of all-real.