Greetings from La Banlieue

Read the campaign story

Heetch and BETC launched Greetings from La Banlieue in November 2023. Their campaign account showed examples in which adding a reference to the Paris suburbs changed the tone of generated images. The proposed response was a collection of real photographs with descriptions, communicated through postcards addressed to Midjourney’s team. The One Club credits BETC, Heetch and Sovage; the agency’s published account identifies the November timing and the corrective image collection.

The source material presents this as a challenge to stereotypes about the suburbs and a brand expression rooted in Heetch’s relationship with those places. This study does not treat selected image examples as a complete technical audit. Nor does it claim that postcard delivery demonstrates model retraining or lasting bias reduction. Establishing those outcomes would require evidence about data use, model versions, evaluation methods and representative results beyond the campaign narrative.

Editorial interpretation: the idea makes a remote technical system accountable through a familiar human communication object. A postcard normally says that a place is worth seeing; here it offers a counterexample to an unflattering representation. The contrast between physical mail and generative technology makes the intervention easy to explain. Its strength is not that paper can automatically change a model, but that a visible correction can make a representation dispute accessible to a wider audience.

Editorial interpretation: the people represented should have influence over the correction. Replacing negative stereotypes with uniformly idealised images would still flatten a varied community. A responsible project needs breadth, consent and context rather than a polished tourism filter. The brand also needs to demonstrate why it belongs in the conversation through its service and relationships, not merely through the publicity opportunity created by a fashionable technology debate.

Distinguish examples from an audit

Document the prompt, date, model version and selection method for any comparison. Repeat tests and include ordinary as well as striking results. A few dramatic images can identify a question, but they cannot establish how often a system behaves that way. Use precise language about what was observed. If a technical claim is important to the campaign, commission an appropriate independent evaluation instead of allowing a compelling film to substitute for one.

Build corrections with the represented community

Invite people to describe what is missing or distorted and agree how their contributions can be used. Obtain permission for photography, captions, distribution and any proposed machine-learning use separately where needed. Include everyday complexity rather than only attractive landmarks. Credit contributors and offer a way to raise concerns. A campaign about representation should not reproduce the same power imbalance by selecting and exporting community images without meaningful participation.

Define what a successful response would be

Separate acknowledgement, dialogue, data acceptance and demonstrated system change. Ask the recipient for a specific, realistic next step and avoid harassment or unsolicited contact with private individuals. If a company says it will investigate, report that commitment accurately without announcing that the problem is solved. Keep the campaign’s own outcomes visible too: people may value a better public portrayal even when the technical change remains uncertain.

A small test for another brief

With a willing local organisation, review a small set of public descriptions of a neighbourhood and create a consent-based correction package for the appropriate editorial owner.

Measure acknowledged corrections, improved description accuracy and contributor satisfaction. Do not claim broad reputational or technical change from a single response.

Stop if participation becomes harassment, image rights are unclear or the evidence cannot support the accusation being made.

• Record examples and distinguish factual errors from subjective framing.
• Invite residents to contribute context and obtain explicit permission for each use.
• Submit the package through an appropriate organisational channel and track the response.