AI Ethics
AI Ethics
What is AI ethics?
- Principles of fairness, transparency, safety, and accountability
- 'Can do' is not the same as 'should do'
- What to consider before adopting AI
AI ethics is not about writing an impressive declaration. It is about building “stop points” into every step of how models are built and used. Putting in writing where data comes from, which variables are left out entirely, who signs off before deployment, and who pulls the switch when something goes wrong elevates technical decisions into management decisions. The change people usually notice in practice is one thing: instead of “development is done, please review it,” everyone sits down together from the planning meeting.
Ethics overlaps with legal compliance but is not the same thing. The EU AI Act divides AI into four tiers (prohibited, high-risk, limited-risk, and minimal-risk). For uses classified as high-risk, such as hiring, credit scoring, and healthcare, it makes data quality management, technical documentation, log retention, and human oversight mandatory. The law is the floor, and ethics is the judgment you exercise above it. In areas regulation has not yet reached, such as turning internal performance data into summary reports or running sentiment analysis on customer calls, something can be legal and still not be accepted by employees and customers. Keep in mind, too, that bias mitigation techniques and explainability (XAI) are tools for putting ethics into practice, not ethics itself.
The most common failure in practice is posting an ethics charter on the website and stopping there. The next is over-reliance on committees. If every internal use of AI has to go through review, approvals back up by two months and business teams end up quietly using personal accounts. A design with different approval paths by risk level lasts much longer. The incidents that actually blow up are not dramatic, either. A hiring model trained directly on past hires reproduces the same discrimination through proxy variables such as club names or career gap periods, without ever looking at gender or school directly. When an audit arrives, it cannot even explain itself because there is no decision log. The cost then is less the fine than the reputation hit of “nobody trusts that company’s AI.”
John KimInstructor John Kim, author of "B2B Marketing Design" and CEO of Nanum Consulting
Seungtae ParkAX instructor who teaches by building AI business development and marketing automation himselfA field that has emerged as AI's social impact has grown. Countries, companies, and international organizations are now establishing AI ethics principles and guidelines.
Consider a hypothetical case. Company A, a savings bank with KRW 3 trillion (≈ US$2,244,501,000) in assets and 320 employees, had been running a machine learning scoring model for remote credit loan screening for eight months. Review time dropped from two days to 20 minutes, but complaints asking why applications were rejected rose to 40 a month. The call center reported, "Customers ask why they were turned down, and we have nothing to tell them."
- The head of risk management pulled six people from the data, compliance, and credit review teams to form an 'AI Review Subcommittee' and capped the time to clear a single model at two weeks.
- When the head of data science re-examined 18 months of training data, postal codes and mobile carrier plans turned out to be input variables, and rejection rates in certain regions were 11 percentage points higher than average.
- Someone said, "They claimed it never looked at region, but it was effectively using it through a proxy variable." After removing those four variables and retraining, the team confirmed the approval rate shifted by only 0.8 percentage points.
- The compliance officer wrote a notice that automatically shows the top three reasons for rejection, and changed the process so that 200 cases a month are sampled and checked, with an underwriter signing off.
- They opened a channel where a person re-reviews any customer appeal within 48 hours, and required every case where the model's prediction and human judgment differed to be logged and kept for 24 months.
- Eight months later, complaints demanding explanations fell from 40 a month to 9, and the gap in rejection rates between regions narrowed from 11 percentage points to 3. The company submitted the supporting documents the Financial Supervisory Service requested in its inspection within three days.