Google DeepMindP332026-06-25abstract onlyred-teamingtext-to-image-safetyglobal-southparticipatory-researchdataset

Going PLACES: Participatory Localized Red Teaming for Text-to-Image Safety in the Global South

Community-led red-teaming workshops in Ghana, Nigeria, and India produced PLACES, a 26,000+ example dataset showing text-to-image safety harms that Western-centric frameworks miss.

It shows current AI image-safety evaluation is structurally blind to non-Western harms, and offers a participatory method for closing that gap.

Read the paper ↗All Google DeepMind papers →Compare its abstract →

Charvi Rastogi · Mukul Bhutani · Minsuk Kahng · Shamsuddeen Hassan Muhammad · Evgeniia Razumovskaia · Priyanka Suresh · Ibrahim Said Ahmad · Charu Kalia · Yaaseen Mahomed · Madhurima Maji · Minjae Lee · Alicia Parrish · Jessica Quaye · Vijay Janapa Reddi · … — meet the researchers →

How much of this do you want?
Orient me keeps four things: the abstract, the method, the claim↔evidence panel, and where it leads next. Everything adds constructs, the model table, every reported statistic, the discussion framing and the style moves. Switching hides nothing permanently and never changes what a section says — it only changes how many are on screen.
Abstract

Two readings, equal authority

How to choose: The paper’s words is verbatim — use it when you need to quote, or to judge how they write. Plain language is a paraphrase written for comprehension — use it when you want the idea fast. Neither is a summary of the other; they are two doors into the same room.

“Despite the global deployment of text-to-image (T2I) models, their safety frameworks are largely calibrated to a Western-centric default, creating significant vulnerabilities for the rest of the world. To embrace cultural pluralism and bring historicallyunder-represented perspectives in T2I safety, we conduct localised community-centered red teaming studies in the GlobalSouth. Our two-fold approach prioritizes localization and participation, by focusing on secondary urban centers in theseregions, and conducting community engagement and training workshops to contextualize local norms. As a result, we presentPLACES, a dataset comprising over 26,000 examples of T2I model failures collected in partnership with universities in Ghana,Nigeria, and two regions of India (Karnataka and Punjab). Analysis of prompts collected reveals a wide-ranging diversity insocio-cultural and linguistic attributes, when compared to existing geography-agnostic crowdsourced red-teaming data. Weobserve unique adversarial patterns enabled by local cultural and linguistic nuances, and distinct clusters within region aroundspecific themes, such as religion in India. Moreover, we uncover structural contextual gaps in existing safety frameworks byidentifying novel harms showing normative dissonance (e.g., violating religious norms, ignoring local customs, and ominoussymbolism). This work argues that expanding T2I safety requires moving beyond mere scale to incorporate deeply localized,participatory methodologies for data collection and contextualization.”

Constructs

What this paper defines

Every definition below is the paper’s own sentence, with its locator. The plain gloss is a reading aid and is marked as one.

PLACES dataset

“we presentPLACES, a dataset comprising over 26,000 examples of T2I model failures collected in partnership with universities in Ghana,Nigeria, and two regions of India (Karnataka and Punjab).”Abstract

In plain terms: A collection of over 26,000 documented cases where text-to-image models produced culturally harmful outputs, gathered with local university partners in Ghana, Nigeria, Karnataka, and Punjab.

normative dissonance

“we uncover structural contextual gaps in existing safety frameworks byidentifying novel harms showing normative dissonance (e.g., violating religious norms, ignoring local customs, and ominoussymbolism).”Abstract

In plain terms: A category of harm where an AI-generated image clashes with local cultural or religious norms in ways that generic safety checks don't flag.

Method

What they actually did

Each step is a synthesis. Open any step to see the paper’s own sentence it was derived from, with its locator — so nothing here floats free of the source.

→
→
→
Local workshops shape community-led red-teaming, producing a 26,000+ example dataset whose analysis surfaces harms invisible to geography-agnostic safety data.
Click any box to open it.
  1. The team ran localized, community-centered red-teaming studies in the Global South rather than using a single global default.
    Trace this step to the paper
    “we conduct localised community-centered red teaming studies in the GlobalSouth.”Abstract
  2. They focused on secondary urban centers and ran community engagement and training workshops to ground the work in local norms.
    Trace this step to the paper
    “Our two-fold approach prioritizes localization and participation, by focusing on secondary urban centers in theseregions, and conducting community engagement and training workshops to contextualize local norms.”Abstract
  3. They partnered with universities in Ghana, Nigeria, Karnataka, and Punjab to collect the red-teaming prompts.
    Trace this step to the paper
    “we presentPLACES, a dataset comprising over 26,000 examples of T2I model failures collected in partnership with universities in Ghana,Nigeria, and two regions of India (Karnataka and Punjab).”Abstract
  4. They analyzed the collected prompts for socio-cultural and linguistic attributes, comparing them against existing geography-agnostic crowdsourced red-teaming data.
    Trace this step to the paper
    “Analysis of prompts collected reveals a wide-ranging diversity insocio-cultural and linguistic attributes, when compared to existing geography-agnostic crowdsourced red-teaming data.”Abstract
  5. They identified adversarial patterns and thematic clusters specific to each region, such as religion-themed prompts in India.
    Trace this step to the paper
    “Weobserve unique adversarial patterns enabled by local cultural and linguistic nuances, and distinct clusters within region aroundspecific themes, such as religion in India.”Abstract
  6. They used the findings to surface gaps in existing safety frameworks by naming new harm categories not previously captured.
    Trace this step to the paper
    “we uncover structural contextual gaps in existing safety frameworks byidentifying novel harms showing normative dissonance (e.g., violating religious norms, ignoring local customs, and ominoussymbolism).”Abstract
The models under study

Exactly what was run, and how

What they reported — and what they left out

The abstract refers only generically to 'text-to-image (T2I) models' as a class; it names no specific model, version, developer, or generation settings.

Results

The numbers they report

The study produced a dataset of over 26,000 examples of T2I model failures from four regions.

over 26,000 examples

See it in the paper
“a dataset comprising over 26,000 examples of T2I model failures collected in partnership with universities in Ghana,Nigeria, and two regions of India (Karnataka and Punjab).”Abstract

Prompts collected showed much wider socio-cultural and linguistic diversity than existing geography-agnostic red-teaming data.

See it in the paper
“Analysis of prompts collected reveals a wide-ranging diversity insocio-cultural and linguistic attributes, when compared to existing geography-agnostic crowdsourced red-teaming data.”Abstract

Each region showed distinct thematic clusters of adversarial prompts, such as a religion-focused cluster in India.

See it in the paper
“distinct clusters within region aroundspecific themes, such as religion in India.”Abstract

The data exposed novel harm categories, described as normative dissonance, that existing safety frameworks do not cover.

See it in the paper
“we uncover structural contextual gaps in existing safety frameworks byidentifying novel harms showing normative dissonance (e.g., violating religious norms, ignoring local customs, and ominoussymbolism).”Abstract
Claim ↔ evidence

What they assert, beside what they showed

Left is the claim in the paper’s own words. Right is the data offered for it. Where the two do not fully meet, a gold band names the distance.

The claim

Existing T2I safety frameworks are calibrated to Western norms and this leaves the rest of the world exposed to harms those frameworks don't catch.

“their safety frameworks are largely calibrated to a Western-centric default, creating significant vulnerabilities for the rest of the world.”

The evidence

“we uncover structural contextual gaps in existing safety frameworks byidentifying novel harms showing normative dissonance (e.g., violating religious norms, ignoring local customs, and ominoussymbolism).”

Abstract
Mind the gap: The claim is framed as covering 'the rest of the world,' but the supporting evidence comes from four specific sites (Ghana, Nigeria, Karnataka, Punjab), not a broader sample.
The claim

Expanding T2I safety requires moving past scale toward deeply localized, participatory data-collection methods.

“This work argues that expanding T2I safety requires moving beyond mere scale to incorporate deeply localized,participatory methodologies for data collection and contextualization.”

The evidence

“Analysis of prompts collected reveals a wide-ranging diversity insocio-cultural and linguistic attributes, when compared to existing geography-agnostic crowdsourced red-teaming data.”

Abstract
Mind the gap: The abstract does not report a direct test of a scale-based approach failing; the 'beyond mere scale' framing rests on a comparison to existing geography-agnostic datasets rather than a controlled scale intervention.
The claim

Local cultural and linguistic context produces adversarial prompt patterns not seen in generic red-teaming.

“Weobserve unique adversarial patterns enabled by local cultural and linguistic nuances, and distinct clusters within region aroundspecific themes, such as religion in India.”

The evidence

“Analysis of prompts collected reveals a wide-ranging diversity insocio-cultural and linguistic attributes, when compared to existing geography-agnostic crowdsourced red-teaming data.”

Abstract
Discussion & after

How they frame it, and what they want next

Their framing

The authors frame mainstream T2I safety work as structurally incomplete because it defaults to Western norms, and present their community-driven method as a corrective. They position PLACES as evidence that scale alone cannot capture culturally specific harms, arguing for participatory, localized data collection as a necessary complement to existing approaches.

Register: The abstract is written in a confident, declarative register throughout ('we uncover', 'we observe', 'this work argues'), with no hedging language.

What they say it means

  • Global deployments of T2I models likely carry safety blind spots outside Western cultural contexts.
    the paper’s words
    “creating significant vulnerabilities for the rest of the world.”Abstract
  • Safety evaluation more broadly should incorporate localized, participatory methods rather than relying on bigger but geography-agnostic datasets.
    the paper’s words
    “This work argues that expanding T2I safety requires moving beyond mere scale to incorporate deeply localized,participatory methodologies for data collection and contextualization.”Abstract

What they call for next

  • Incorporate deeply localized, participatory methodologies into future T2I safety data collection and contextualization work.
    the paper’s words
    “This work argues that expanding T2I safety requires moving beyond mere scale to incorporate deeply localized,participatory methodologies for data collection and contextualization.”Abstract
For your own writing

Moves worth stealing

Titles the work with a punning acronym: 'PLACES' is both the dataset's name and, in 'Going PLACES,' an idiom for making progress.

“Going PLACES: Participatory Localized Red Teaming forText-to-Image Safety in the Global South”

Opens with a deficit framing of the status quo ('Despite... calibrated to a Western-centric default') before introducing the paper's own method, so the reader feels the gap before seeing the fix.

“Despite the global deployment of text-to-image (T2I) models, their safety frameworks are largely calibrated to a Western-centric default, creating significant vulnerabilities for the rest of the world.”
Connected

Where else this leads

Same people

Same territory

Published alongside it

The nearest publications in time, across all three labs.

What this page was built from

The source page provides only the publication's abstract, author list, and venue metadata (FAccT 2026), not the full paper, so every field here is derived solely from that abstract; the scraped text also contains missing-space artifacts (e.g. 'GlobalSouth', 'presentPLACES', 'Weobserve') which are preserved verbatim in quoted fields rather than silently corrected.