Black & Abroad offers luxury travel experiences for an underserved and under-represented customer base: young, upwardly mobile members of the Black community. The brand’s destinations include Paris, Ghana, Colombia and South Africa. With 30% of revenue coming from repeat business, and the pandemic’s industry impact finally fading, 2024 was the perfect opportunity to re-engage with past guests to drive new bookings.

BRIEF

Drive new bookings from past Black & Abroad guests with an uplifting, tech-forward engagement platform.

OBJECTIVES

Lift in engagement among previous Black & Abroad customers

High open rates in email

Percentage increase in repeat bookings.

Strengthen leadership position as an innovative, purpose-driven brand.

SEE YOU THERE

A project that began as a 1:1 digital experience for Black & Abroad customers – and became a global data solution for bias in generative AI models.

https://seeyouthere.ai/

A PROJECT IN THREE PARTS

1. Personalized digital experiences for Black & Abroad customers, in which we used generative AI to show past guests at future destinations.

2. A mass media campaign highlighting the racial bias problems we encountered along the way.

3. The Generative AI Bias Reporting System – our data solution to the problem of bias in generative AI.

GENERATING PAST GUESTS INTO FUTURE DESTINATIONS

We began by collecting +11,000 untagged images of guests from previous Black & Abroad trips. We hand-sorted by customer, prioritizing facial clarity and removing unusable or irrelevant images – to arrive at 5 to 10 images per customer, which we ingested into a secure cleanroom environment to test, train and refine our AI models. We documented problematic images, then uprezed and fine-tuned outputs, for a long list of selects. Human review determined final image selects.

STRUCTURING UNSTRUCUTRED BIAS DATA

The Generative AI Bias Reporting System is our data solution to bias in generative text, image and video. By entering instances of bias, users are creating structured data at scale, for use by anyone, anywhere in the refinement and development of gen AI models.

TARGETING

First party data – customer email, photos and purchase

Third party data – MRI Simmons, Acxiom Infobase, Kantar Mars

GENERATING PAST GUESTS INTO FUTURE DESTINATIONS

We began by collecting +11,000 images of guests from previous Black & Abroad trips. We hand-sorted by customer, prioritizing facial clarity – to arrive at 5 to 10 images per customer, which we ingested into a cleanroom to test, train and refine our models. We documented problematic images, then uprezed and fine-tuned , for a long list of selects. Human review determined finals.

OUR GEN AI TECH STACK

Google Colab – Testing, prototyping

Dreambooth/Autotrain/SDXL1.0, PyTorch – LoRA weights training, raw image generation

sdxl-outpainting-lora (batouresearch) – Outpainting, upscale

SDXL-Magic-Image_Refiner, MagnificAI – Refine, upscale

HuggingFace – Temporary hosting of (private) trained weights, access control

Pillow – Compositing, image processing

Pandas – Data management (client approvals)

CREATING THE GENERATIVE AI BIAS REPORTING SYSTEM

Modelled after the FDA’s Adverse Event Reporting System (FAERS), our reporting system is a disciplined technical solution to bias in generative AI across text, image and video. It collects unstructured, observed instances of bias and converts them into structured data at scale.

BIAS REPORTING STACK

Python backend

Postgres database

Python Front End

Hosted on Google Cloud Platform

Data wrangling SQL, Pandas

APPLICATION ACROSS INTERSECTIONALITY

Go to any gen AI platform. Prompt: “two trans women”. How old are the women it generates? What are they wearing? Ethnicity? Are they hypersexualized? Our utility provides a systematic way to consider questions such as these, while delivering structured, open data back to the source.

DATA USE CASES INCLUDE:

Testing and training new gen AI models on prompts causing biased outputs on other platforms. Benchmarking different models on inclusivity themes.

Observing trends across intersectionality.

TOUCHPOINTS

1:1 email

1:1 pages featuring past customers at 5 upcoming destinations

Paid social, OOH

Launch film

https://seeyouthere.ai/

TIMELINE

Four-month development cycle.

DRIVING CHANGE

3,462 biases reported

81,293,895 impressions – The most seen campaign in Black & Abroad’s history

“How the ad industry in making AI images look less like AI” – Wall Street Journal

FILLING THE FUNNEL

81% email open rate – of emails using generative imagery of past customers

218% clickthrough rate vs. benchmark from digital media (AdTheorent)

77.68% view rate on top performing creative shown on Meta platforms

53% higher video completion rate vs. YouTube Benchmark (ZEFR)