Responsible AI practices
Responsible AI Challenges in Traditional AI and Generative AI
Accuracy: bias and variance
The story: Three students prepare for a math exam.
- The first only learns "the answer is usually around 10". They get most practice questions wrong and most exam questions wrong too. Too simple.
- The second memorizes every practice question and its answer, including the typos. They ace the practice set, then fail the exam because the numbers changed.
- The third learns the actual methods. They do well on both.
In AI/AWS terms: A model only knows its training data, so poor training gives inaccurate results. Accuracy is the top problem developers face.
| Student | Name | Bias | Variance | Result |
|---|---|---|---|---|
| "Around 10" | Underfitting | High | Low | Too simple. Misses important features and performs poorly even on training data |
| Memorizer | Overfitting | Low | High | Memorizes noise. Great on training data, poor on new data |
| Learned the methods | Balanced (the goal) | Low | Low | Captures the real patterns without the noise |
- Bias is the gap between the model's predictions and the true values: how far off the first student is.
- Variance is how much predictions change with different training data, a measure of sensitivity to noise: how thrown the second student is by new numbers.
- The bias-variance trade-off is finding the balance between the two.
For the exam: Good on training data but bad on new data = overfitting (high variance). Bad on both = underfitting (high bias).
How to fix bias and variance errors
The story: How a tutor would help the memorizer and the oversimplifier:
- Give surprise quizzes with questions they haven't seen, to catch memorizing.
- Give them more and more varied practice questions.
- Mark them down for overly specific tricks that only work on one question.
- Teach a simpler approach to the memorizer, and a richer one to the student who oversimplifies.
- Cut out distracting details that don't matter.
- Stop drilling before they start memorizing the practice sheet.
In AI/AWS terms:
- Cross-validation: train on subsets of the data and test on the rest to detect overfitting (the surprise quiz).
- Increase data: add more samples.
- Regularization: penalize extreme weights to prevent overfitting.
- Simpler models: reduce overfitting. If a model underfits, it may be too simple.
- Dimension reduction (for example PCA): cut the number of features while keeping the information.
- Stop training early: so the model doesn't memorize the data.
For the exam: Cross-validation detects overfitting. Regularization, more data, simpler models, and early stopping reduce it.
Challenges specific to generative AI
The story: A brilliant new writer joins a magazine. Problems appear:
- Sometimes they write something offensive, and the editors argue about what counts as offensive versus just blunt, which differs by country.
- They quote a study that doesn't exist, because it "sounded right" as they wrote.
- They copy a paragraph word for word from a book, or paint exactly in a famous artist's style.
- Students start handing in essays this writer wrote, and teachers can't tell who wrote what.
- Other writers worry they'll lose their jobs.
In AI/AWS terms:
- Toxicity: offensive or inappropriate content. Hard to define, and the line between filtering and censorship depends on context and culture.
- Hallucinations: plausible but false claims, such as invented scientific citations. They come from next-word sampling: the model picks likely-sounding words, not checked facts.
- Intellectual property: models reproducing training data verbatim, or imitating an artist's style.
- Plagiarism and cheating: AI-written essays and job applications. It's hard to verify who authored content.
- Disruption of the nature of work: worries that some professions will be replaced or transformed.
For the exam: Know the five generative AI challenges: toxicity, hallucinations, intellectual property, plagiarism and cheating, and disruption of work.