Unblocking $30M with AI for ads integrity
Risk and training AI in data-sensitive use cases
Problem
30 millions ads are rejected daily for using promotional images, text or video content that doesn't comply with our Advertising Standards. 70% of all ads rejected for these issues are remediable, but only 7% of advertisers actually attempt to fix them. Some issues are more clear-cut, such as explicit nudity, whereas other issues are more ambiguous, such as using comparison images which might 'make people feel bad about their body'.
Creating and uploading ad campaigns is time-consuming and expensive; image galleries can sometimes contain up to 500 images and when an ad is rejected it can be challenging to understand what within the ad (or 500 images) actually caused the issue. How can we empower advertisers to identify and resolve issues with ad creative.
What I did
- Created system prompts
- Created model architecture, novel eval and labelling criteria
- Created AI agents to review over 7000 LLM-generated responses to benchmark response quality
- Created Meta's first AI integrity playbook to define guardrails for data ingestion, and transparency UX per policy
- Created UX principles for AI
- Created and tested AI response principles for integrity
- Trained and iterated AI models on our internal policies
- Designed E2E UX
Who I worked with
1 Designer 1 Researcher 1 Product Manager 4 Engineers 1 Data Scientist 1 Policy Manager 1 Lawyer 1 Product Marketing Manager
Constraints
- Limited real estate AI is known for it’s (unfortunate) verbosity. We had the space of an error message to explain why a particular image/text/video was violating and needed to be succinct. How to overcome?
- Real-time abuse vector AI ads editing tool could become a new abuse vector. If we allowed unlimited queries, bad actors could abuse this and map our rules to circumvent our policies.
- Policy and legal concerns To avoid legal liability should the model hallucinate, we were unable to give advertisers instructions on how to fix, e.g. Say “Remove this image”. How can we frame guidance to be useful?
- Surface-owner alignment We don’t own the components, or interactions we were proposing and needed to align with the surface-owner team, which was slow- especially as some of the things we were proposing were net-new.
How I did it
🔎 Step 1. Defining AI UX principles
"AI" comes in many forms and we wanted it to be intentional and practically useful to the user, vs. just creating another chatbot.
We organised and ran a 3-day sprint with UXR and designers from across the company to create a set of best practices for using AI in product.
We came up with a golden principle that the level of user input needed for diagnosis and resolution should dictate the AI modality we offered.
- For tasks that are easy for Meta to diagnose and solve, automation is the best approach.
- For tasks that are difficult for Meta to diagnose and resolve, direct chat assistance is required.
- Tasks that are easy to diagnose but challenging to resolve should provide contextual AI guidance, with chat support available. This was where ad rejects fitted, so agreed that we would need both an in-product experience and a chat experience.
👀 Step 2. Mapping AI risks
Before we started prompting, we needed to understand the risks and guardrails.
I organized a workshop with ICs from policy, legal, engineering as well as CDs working in prompt engineering from across Meta to map principles, risks and workflows of AI ingestion.
For instance what happens if someone tries to socially engineer or abuse the chatbot into doing what it wants; how should AI respond, or we detect a bad actor engaging with our technology. Do we need systems to remote turn off AI conversation if we have a regional emergency?
I developed Meta's first transparency framework that divided policies into tiers based on risk and reformability.
These tiers could be plugged straight into the model architecture the engineers were building and would dictate the type of responses we could give for each issue.
The idea was that transparency operates on a sliding scale.
- Low risk, high reformable policies are allowed high transparency; or where we allow the model some degree of freedom to return personalised guidance.
- High risk, low reformable policies however receive canned responses.
👟Step 3. System prompting and evals
Solution
Impact
$35M unblocked revenue
81% increase in edit attempt
50% increase in successful edits