Pull to refresh
Logo
What Changed
Following Why Sign Up
ChatGPT's image model rewrites factory farms into pastures, study finds

ChatGPT's image model rewrites factory farms into pastures, study finds

New Capabilities

A 4,800-image study traces DALL-E 3's pastoral bias to automatic prompt revision.

July 15th, 2026: Peer-reviewed study published in Technology in Society

Overview

Updated 1 hour ago

Ask ChatGPT's image generator for a dairy farm and it draws green pasture, grazing cows and a blue sky. A study of 4,800 images found that result in every default dairy request and in 99% of default pig-farm requests.

The cause is not the image model's base training alone. DALL-E 3 automatically rewrites user prompts before drawing. When researchers blocked that rewriting, 60% of dairy images and 96% of pig images showed animals indoors, behind metal barriers or on concrete floors. The study says a feature meant to improve images is erasing the look of industrial farming.

Why it matters

If image models rewrite ordinary prompts toward idyllic scenes, they can make industrial farming invisible to people who never visit one.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

4,800
Images generated
The study generated 100 images per prompt across 48 prompts.
100%
Default dairy prompts showing pasture
Every basic dairy-farm request produced cows grazing outdoors.
99%
Default pig prompts showing outdoor scenes
Almost every basic pig-farm request produced free-roaming pigs outdoors.
60%
Dairy images showing indoor housing when revision was blocked
With prompt rewriting inhibited, cows appeared indoors at feed barriers.
96%
Pig images showing indoor housing when revision was blocked
With prompt rewriting inhibited, pigs appeared behind metal railings on concrete floors.
94%
Dairy 'reality' prompts still showing pasture
Even explicit requests for realistic dairy farms produced pastoral scenes.
91%
Pig 'reality' prompts still showing outdoor scenes
Explicit requests for realistic pig farms mostly produced pasture or mud.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

Log in Sign Up

Organizations Involved

OpenAI
OpenAI
Artificial intelligence company
Maker of ChatGPT and DALL-E 3, the image model at the center of the study

OpenAI makes ChatGPT and DALL-E 3, the image model the study tested.

Timeline

2023 July 2026

3 events Latest: July 15th, 2026 · 4 months ago

  1. Peer-reviewed study published in Technology in Society

    Latest Publication

    The paper reports that blocking prompt revision shifts most farm images from pastures to indoor housing.

  2. Preprint posts to arXiv

    Research

    The study tests whether DALL-E 3's prompt revision changes how livestock farms are depicted.

  3. OpenAI ships DALL-E 3 with automatic prompt revision

    Product launch

    DALL-E 3 rewrites user prompts before generating images, a step OpenAI says improves quality and safety.

Scenarios

1

OpenAI changes prompt rewriting for farm imagery

Possible Resolves by Jul 15, 2027

Discussed by: The study authors and AI transparency researchers

The study calls on OpenAI to disclose the full prompt-revision system prompt. If OpenAI updates its DALL-E system card or API documentation to stop adding pastoral detail to farm prompts, this scenario resolves yes. The trigger would be a public policy change, not a blog post about image quality.

2

Independent audit finds same bias in another image model

Possible Resolves by Jan 15, 2027

Discussed by: AI safety researchers and replication teams

The paper's mechanism is prompt revision, not a DALL-E-only quirk. A replication on Midjourney, Stable Diffusion or Google's Imagen would show whether the bias is industry-wide. The test is a peer-reviewed or major-lab audit using the same no-revise comparison.

3

OpenAI leaves farm prompt rewriting unchanged

Likely Resolves by Jul 15, 2027

Discussed by: OpenAI watchers and the study authors

OpenAI has not said it will change the prompt-revision step. If its public documentation still describes automatic rewriting with no farm-specific exception by mid-2027, the bias remains in place.

Historical Context

2 moments from history that rhyme with this story — and how they unfolded.

February 2024

Google Gemini image pause (2024)

Google paused Gemini's image generation of people after users found historically inaccurate outputs, including Black Nazi soldiers and Asian American Founding Fathers. The model's diversity prompts had overridden historical context.

Then

Google apologized and disabled people generation.

Now

The episode showed that prompt-level guardrails can distort history, and the fix was to turn off the feature.

Why this matters now

The DALL-E farm study describes a similar layer: automatic rewriting changes what the base model would draw.

July 2015

Google Photos gorilla label (2015)

Google Photos labeled a Black software developer and his friend as gorillas. The company apologized and blocked gorilla-related labels.

Then

Google removed the offensive label category.

Now

The underlying image-recognition bias was not solved; the company suppressed a symptom.

Why this matters now

The farm study argues OpenAI's prompt revision suppresses an uncomfortable reality rather than fixing the model's base output.

Sources

(2)
Technology in Society arXiv