-
-
Backlash Predictor
Inspiration
Campaigns rarely fail because the idea is bad. They fail because different audiences read the same message differently.
I kept watching bold marketing lines trigger unnecessary backlash, not out of malice, but because teams never saw the reaction coming. Critics, first-time viewers, and brand safety reviewers all interpret the same flyer through different lenses.
Backlash Predictor started with one question: What if you could see those reactions before you hit publish?
What It Does
Backlash Predictor analyzes campaign materials, including images, flyers, and text, and evaluates how distinct audience segments are likely to interpret them.
It extracts structured campaign elements: headline, claims, tone, visual theme, calls to action, and disclaimers. Then it simulates reactions across defined audience lenses, calculates a multi-dimensional backlash risk score, flags the specific phrases causing friction, and generates targeted rewrite suggestions so you remove blind spots without losing boldness.
How We Built It
We built a multimodal pipeline powered by Gemini that processes both text and visual campaign assets.
Unstructured marketing material gets converted into structured campaign anatomy, then evaluated across four risk dimensions: overclaim, inclusivity, authenticity, and tone. Audience personas are anchored to real stakeholder types, including loyal followers, industry critics, first-time viewers, and brand safety reviewers, rather than generic demographic buckets.
The frontend surfaces trust scores, flagged phrases, and rewrite suggestions in a clean interface built for rapid pre-launch iteration.
Challenges We Ran Into
The hardest problem was making the "simulation" credible. Audience reactions had to feel grounded, mapped to consistent evaluation logic, not just plausible-sounding LLM output.
Making the backlash score transparent was equally important. A single number felt like a black box, so we broke it down by dimension and tied each score directly to flagged phrases in the original content.
We also had to walk a fine line: critique the language without flattening the creative voice. The goal was sharper messaging, not safer mediocrity.
Accomplishments We're Proud Of
We built a working multimodal risk analysis system that goes beyond sentiment into structured reputational modeling.
The interface clearly connects each audience reaction to a risk dimension and a specific language trigger, so the output is explainable, not just scored.
Most importantly, the rewrite suggestions are actionable. The tool doesn't just flag risk; it shows exactly how to reduce it.
What We Learned
Backlash is rarely about intent. It's about perception, and perception varies by audience.
We also learned that transparency is what makes AI analysis feel trustworthy rather than arbitrary. Showing why something scores high matters as much as the score itself.
And we saw that AI can serve as a genuine strategic reviewer, not just a content generator.
What's Next for Backlash Predictor
We plan to incorporate live social signals and real audience data to further ground persona reactions beyond model simulation.
Next steps include expanded risk dimensions, industry-specific evaluation frameworks for entertainment, healthcare, and finance, and side-by-side comparison of original versus revised campaigns.
Long term, Backlash Predictor becomes a pre-publish safety layer embedded into the workflow of marketing teams, agencies, and creator platforms.
Log in or sign up for Devpost to join the conversation.