What are the ethical concerns when picking an AI model?
Bias and Fairness
AI models can perpetuate biases present in their training data, leading to unfair or discriminatory outputs. This is a concern for applications like hiring, lending, or healthcare. Some providers publish bias evaluations, but standards vary.
Consider whether the model has been tested for bias across different demographics and whether the provider has a plan to mitigate it. Open-source models allow independent auditing, which can help.
Transparency and Consent
You should know what data a model was trained on and whether that data was collected with consent. Many models are trained on web-scraped data, which raises copyright and privacy issues. Some providers disclose training data sources; others don't.
Also consider how the model is governed: Who controls it? Can it be audited? Is there a way to appeal harmful outputs? These questions matter for accountability.
- Bias: Does the model produce stereotyped or unfair results?
- Transparency: Is training data and model behavior documented?
- Consent: Was data used with permission?
- Environmental impact: Training large models consumes significant energy.
- Misuse: Could the model be used for deepfakes, spam, or surveillance?
Common mistakes
- Assuming that a model is neutral because it's technical; all models reflect choices made by developers.
- Ignoring the environmental cost of running large models, especially for trivial tasks.
- Believing that open-source models are automatically more ethical; they can still be trained on problematic data.
