Article: “Fairness and Bias in Artificial Intelligence: A Brief Survey of Sources, Impacts, and Mitigation Strategies” by Emilio Ferrara (2024) MDPI
This survey reviews where bias in AI comes from (data, human input, algorithm design), its impact (inequitable outcomes, stereotypes, reinforcing social inequalities), and ways to reduce bias—such as better data practices, auditing, and more transparent / accountable models. MDPI
💡 3 Key Insights
- Bias often arises from the dataset
If the training data does not represent all segments (e.g. underrepresented demographics, edge cases), the model will likely perform poorly or unfairly for those groups. MDPI - Mitigation needs to include ongoing evaluation
It’s not enough to build something once; you must continuously audit, monitor performance, and adjust models to prevent bias from creeping in (especially as new data arrives). MDPI - Transparency and clarity build trust
Users, clients, or stakeholders should understand what data was used, how decisions are made, and what limitations exist. This helps avoid surprises and ensures ethical deployment. MDPI
🛠 How These Can Apply to Unilancerz
| Insight | Application in Unilancerz Projects |
|---|---|
| Dataset bias | When training recommendation systems (matching freelancers to clients), ensure your sample data covers all skill levels, geographies, languages, etc. Don’t rely only on your early users. Include data from minorities, etc. |
| Ongoing evaluation | Set up pipelines for continuous testing of your models. For example, if you use AI-chatbots or job matchers, track where people complain / where matches are rejected / where users say the suggestions were wrong, and retrain or fix those issue areas. |
| Transparency | On your platform, show users information about how AI components work (e.g. “This recommendation was made because of your past projects in UI/UX + skill tags”). Also communicate known limitations (“This model doesn’t work well for languages X yet”). Helps with trust & user satisfaction. |