AI Credibility Datasets — Validation & Reproducibility

Fabricated facts, invented citations, and ungrounded assertions. Measured against reference answers and retrievable evidence rather than fluency.

• https://github.com/sylinrl/TruthfulQA
• https://openai.com/research/gpt-4
• https://deepmind.google/discover/blog/reducing-hallucinations-in-large-language-models/

Representation and disparity across populations — where outputs diverge systematically by group rather than by evidence.

• https://www.nist.gov/ai
• https://aif360.res.ibm.com/
• https://crfm.stanford.edu/

Harm and misuse surfaces: jailbreaks, prompt injection, and refusal behaviour under adversarial pressure.

• https://safetyinstitute.ai/
• https://www.anthropic.com/index/constitutional-ai
• https://openai.com/safety

Stability under distribution shift and perturbation. A credible system degrades predictably rather than silently.

• https://deepmind.google/research
• https://openai.com/research
• https://www.anthropic.com/index/research

Source reliability and evidence quality — whether a claim carries provenance a reader can follow back to origin.

• https://www.technologyreview.com/
• https://www.wired.com/tag/artificial-intelligence/
• https://www.theguardian.com/technology

Truthfulness, reasoning, and safety benchmarks, plus the contamination and versioning problems that make scores drift.

• https://github.com/sylinrl/TruthfulQA
• https://arxiv.org/abs/2009.03300
• https://github.com/google/BIG-bench
• https://arxiv.org/abs/2309.07864