If a personalized assistant remembers a private conversation, can you see what it retained or change it? Transparency means making an AI system’s data, methods, evaluations, and behavior visible enough for people to assess it. For an assistant, that includes what it remembers and how people can inspect, edit, or reset that memory. section 10.4 section 10.19
Transparency has several layers:
- Dataset documentation describes the data used to develop a system. section 10.19
- Model disclosures describe training methods, benchmarks, and evaluation results. section 10.19
- System-level tools help people interpret, debug, and correct outputs. section 10.19
Published benchmark results can help people compare systems, but a score is not a complete performance guarantee. Results depend on what a test measures and can change with details such as evaluation formatting. Transparent reporting should help users judge suitability for their own task—and help developers see a model’s limits. section 10.19
Table 10.1 A leaderboard for code-generation benchmark BigCodeBench
| Rank |
Model |
Pass@1 |
Year |
| 1 |
Gemini-Exp-1206 |
62.4 |
2024 |
| 2 |
DeepSeek-V3 |
62.2 |
2024 |
| 3 |
Llama-4-Maverick |
61.4 |
2025 |
| 4 |
GPT-4o |
61.1 |
2024 |
This mid-2025 BigCodeBench leaderboard compares models on real-world programming tasks using a first-attempt success measure. Its dated, task-specific results illustrate useful disclosure, not a universal ranking. section 10.19
Disclosure also matters beyond test scores. Humanlike chatbots can blur boundaries even when their nature is disclosed; changes to a companion chatbot have left users distressed. In workplaces and schools, transparency about AI use belongs alongside consent, privacy policies, and accountability for sensitive data. Openness alone does not resolve these risks, but opaque systems are harder to question or hold accountable. section 10.13 section 3.12 section 3.16 section 11.6
