Virtuous Cycle
The more users you serve well, the more profitable the business, and the more you can bid for ads and get more traffic, creating a positive feedback loop that helps good companies do better.
Ethical Use Recommender
Ethical Use of Recommender Systems
Section titled “Ethical Use of Recommender Systems”Potential for Harm Despite Profitability
Section titled “Potential for Harm Despite Profitability”Even though recommender systems have been very profitable for some businesses, there are some use cases that have left people and society at large worse off. The goal should be to use recommender systems and other learning algorithms only for things that make society at large and people better off.
Configuration Choices and Their Implications
Section titled “Configuration Choices and Their Implications”System Goal Setting
Section titled “System Goal Setting”When designing a recommender system, there are many choices in setting the goal and deciding what to recommend to users. The label y can represent different user actions: engagement, clicks, explicit likes, or other interactions.
Seemingly Reasonable Approaches
Section titled “Seemingly Reasonable Approaches”User Satisfaction Focus
Section titled “User Satisfaction Focus”- Recommend movies most likely to be rated 5 stars - This seems fine and appears to show users movies they would like
- Recommend products users are most likely to purchase - This seems like a reasonable use of a recommender system
Revenue-Driven Approaches
Section titled “Revenue-Driven Approaches”- Show ads most likely to be clicked on - Many companies show ads likely to be clicked where the advertiser had put in a high bid, since revenue depends on whether the ad was clicked and what the advertiser bid per-click
While this is a profit-maximizing strategy, there are possible negative implications of this type of advertising.
Profit Maximization
Section titled “Profit Maximization”- Recommend products that generate the largest profit - Many websites don’t show the most relevant product or the product you’re most likely to purchase, but instead show products that generate the largest profit for the company
Engagement Maximization
Section titled “Engagement Maximization”- Show content that leads to maximum watch time - Video websites or social media websites that run on ad revenue have an incentive to keep you on the website for a long time to show you more ads
Problematic Use Cases
Section titled “Problematic Use Cases”Advertising as an Amplifier
Section titled “Advertising as an Amplifier”Advertising can amplify both beneficial and harmful businesses.
Positive Example: Travel Industry
Section titled “Positive Example: Travel Industry”- Good travel companies provide excellent experiences to users
- Success comes from truly serving users well
- More profitable businesses can bid higher for ads
- Higher bids lead to more traffic from advertising sites
- Creates a virtuous cycle where better service leads to more success
Problematic Example: Payday Loan Industry
Section titled “Problematic Example: Payday Loan Industry”- Payday loan industry charges extremely high interest rates, often to low-income individuals
- Companies efficient at exploiting customers become more profitable
- Higher profits enable higher ad bids
- Higher bids lead to more traffic sent to exploitative companies
- Creates a positive feedback loop that causes the most harmful companies to get more traffic
Content Engagement Problems
Section titled “Content Engagement Problems”Conspiracy Theories and Toxic Content
Section titled “Conspiracy Theories and Toxic Content”Maximizing user engagement (time spent watching videos or on social media) has led large platforms to amplify conspiracy theories or hate and toxicity because:
- Conspiracy theories and certain types of hate/toxic content are highly engaging
- They cause people to spend significant time consuming them
- The effect of amplifying such content is harmful to individuals and society
Partial Solutions
Section titled “Partial Solutions”- Content filtering: Try to filter out problematic content such as hate speech, fraud, scams, certain types of violent content
- Challenge: Definitions of what exactly should be filtered out are surprisingly tricky to develop
User Awareness Issues
Section titled “User Awareness Issues”Many users don’t realize that apps and websites are trying to maximize profit rather than necessarily the user’s enjoyment of the recommended media items. Users often think the system is trying to recommend things they will like.
Ethical Guidelines and Solutions
Section titled “Ethical Guidelines and Solutions”Transparency Recommendations
Section titled “Transparency Recommendations”Encourage companies to be transparent with users about the criteria by which they decide what to recommend. Although this isn’t always easy, being more transparent about why recommendations are shown can increase trust and cause systems to do more good for society.
Harm Prevention Strategies
Section titled “Harm Prevention Strategies”- Refuse ads from exploitative businesses - Though defining what constitutes an exploitative business is challenging
- Invite diverse perspectives - Get multiple opinions from multiple people when making design choices
- Open discussion and debate - Engage in dialogue about the implications of system design decisions
Development Principles
Section titled “Development Principles”Ethical Development Approach
When building recommender systems or any machine learning technology:
- Think through not just the benefits you can create, but also the possible harm
- Invite diverse perspectives and discuss and debate implications
- Only build things that you really believe can make society better off
Call for Responsible Development
Section titled “Call for Responsible Development”Recommender systems are very powerful, profitable, and lucrative technology, but they also have problematic use cases. The goal should be for everyone in AI to only do work that makes people better off.
Key Considerations:
- Consider both intended benefits and potential unintended consequences
- Engage multiple stakeholders in decision-making processes
- Prioritize societal benefit alongside business objectives
- Maintain transparency about system goals and methods
- Continuously evaluate and adjust systems based on their real-world impact
The responsibility lies with developers, companies, individuals, and governments to continue wrestling with these complex problems and work toward solutions that benefit society as a whole.