IIM Ahmedabad, Applied Deep Learning and Generative AI for Managers, November 2026

How many real images does a malaria AI need?

real labelled cells
drawn by ChatGPT
fewer expert labels to pay for

The answer in one slide

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.

Every training label costs a trained microscopist’s time

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.

0 cells read0 infected

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.

We tested on 50 patients the model never saw, as in a real roll-out

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.

An original cell image
original,
The same cell padded to a square with black
padded square
The cell resized to 96 by 96 pixels
96 × 96 PNG

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.

151
training patients, 800 cells
50
test patients, 200 cells, locked
0
patients on both sides

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.

We showed ChatGPT five real cells at a time and asked for sixteen new ones

20 + 20reference cells, training pool only
5rotated into each new chat
4 × 4 grids from Create image
tiles, each one inspected
400 + 400kept, 96 × 96
These are . Create one square image that is a 4 by 4 grid of 16 new, separate images in the same style, camera angle and colour as these references, but each different in content: slightly pinker tint, rounder cells, and a different cell size and texture in each image. Match the references closely: each cell fills almost its whole square, with pale natural colours and the soft, slightly blurry look of a real low-resolution microscope photo. Do not copy any reference image. Thin white lines between images. No text, labels or borders.

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.

A raw 4 by 4 grid of infected red blood cells generated by ChatGPT

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.

Spot the fake. Tap every cell you think ChatGPT drew

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

A fair test: 50 models, each trained on 800 cells, and only the mix changes

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.

Same modela standard image classifier, pre-trained on everyday photos (ResNet-18)
Same training10 passes over the 800 cells, never tuned on the exam
Same exam200 real cells from the 50 locked-away patients
Pass markwithin 0.01 AUC of the all-real model, fixed before any run
10 real cells10 ChatGPT cellsone training run (seed)

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.

Cheapest mix

Infected test cells the ChatGPT-only model calls uninfected
Infected cells the ChatGPT-only model calls healthy. p is how likely it thinks infection is; all-real is the same cell scored by the all-real model.

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

  1. Benchmark a free, open model before paying for labels or training.
  2. A model built for your field (medical) beats a general one.
  3. When the quality bar is tight, fine-tuning still wins, and ChatGPT cells still save labels.

Test cells as they are and under three simulated other-lab shifts
The same seven test cells as collected, then made to look like a lab with a purpler stain, a paler smear or another camera. Training cells were left as they were.

To the head of laboratory services, national malaria programme

Decision

At scale

Cut-off

Before go-live

After launch

Real labels needed = 800 × 30% + the 40 reference cells shown to ChatGPT = 280. Labels avoided = 800 − 280 = 520.

Risks, and how we would manage them

A new site looks different

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.

Generated data has blind spots

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.

Much of the skill is borrowed

Patient data and copying

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.

Fakes made from fakes

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.

Accountability

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.

What we ask you to approve

  1. Commission expert-labelled cells instead of 800.
  2. Have ChatGPT draw more, with our logged prompt and a check of every image.
  3. Flag cells from a cut-off set by agreed error costs (about 8% today), not the default 50%.
  4. Top up to 440 labels before a national roll-out, and check a few hundred local cells at every new site.
  5. Keep a microscopist in charge of every result, and track what the model misses.
real labelled cells to collect
cells for ChatGPT to draw
fewer expert labels

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 accuracy
FID real vs ChatGPT; two real halves give
tiles that look like a copy of a reference
Sixteen real parasitized cells beside sixteen ChatGPT cells
Parasitized. Left, 16 real cells; right, 16 ChatGPT cells.
The six synthetic tiles most similar to a reference, each next to that reference
The six closest ChatGPT-to-reference pairs in the whole pool are still different cells.

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.

Grad-CAM heatmaps of the ChatGPT-only and all-real models on missed rim-parasite cells
The five test cells the 5%-real model gets wrong and the all-real model gets right
The brief’s check: cells the 5%-real model gets wrong and the all-real model gets right. PAR infected, UNI healthy.

Everything here can be re-run from the project folder

reportreport/report.pdfthe written report
notebooknotebooks/project.ipynbend to end, outputs saved
resultsresults/results.csvone row per training run
logdocs/generation_log.mdevery prompt, reference and discarded tile
plandocs/milestone1.mdmetric and tolerance, fixed before the runs

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

Labels saved
Chance of passing
Missed per 100 infected

Option B ·

expert labels

Labels saved
Chance of passing
Missed per 100 infected

“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.