A Wall Street Journal piece flagged on Hacker News shows Claude cannot answer basic questions about Dario Amodei's wife. The model returns no information despite Amodei serving as Anthropic's CEO and co-founder.
The thread received 11 points and 4 comments. Users noted the result aligns with standard model behavior rather than a special safeguard.
How Model Knowledge Boundaries Operate
Claude's training data ends at a fixed cutoff date. It contains no mechanism for real-time web access or private personal records. Queries about non-public family details fall outside the training corpus by design.
Anthropic applies additional refusal layers on sensitive personal topics. These layers trigger even when partial public data exists elsewhere.
What the HN Comments Highlight
Early comments focused on three points:
- The outcome reflects normal cutoff behavior, not targeted censorship.
- Similar gaps appear in GPT-4o and Gemini when asked about private individuals.
- Public figures receive uneven coverage depending on media volume before the cutoff.
No commenter reported successful extraction of the same information from any frontier model.
Privacy Protections Versus Capability Limits
Current LLMs separate two issues. One is the absence of data in training sets. The other is deliberate refusal policies applied after training. The Amodei query appears to hit the first constraint.
Companies rarely publish exact cutoff dates or family-name filtering rules. This opacity makes it hard to predict which personal facts any model will handle.
| Model | Public Cutoff | Real-time Search | Personal Data Refusals |
|---|---|---|---|
| Claude 3.5 Sonnet | 2024-04 | No | High |
| GPT-4o | 2023-10 | Optional | Medium |
| Gemini 1.5 Pro | 2024-06 | Yes | Medium |
Who Needs to Account for These Gaps
Developers building background-check tools or executive profiling features should test multiple models against known private facts. Researchers studying training data leakage can use similar queries as negative controls.
Users expecting LLMs to serve as current biographical databases will encounter repeated failures on non-public individuals. Public relations teams monitoring model outputs gain little from these tests.
Practical Testing Steps
Run identical queries across Claude, GPT-4o, and Gemini on any public figure whose spouse has minimal media presence. Record refusal rates and hallucination frequency. Compare results against the same queries run on the figure's Wikipedia page or LinkedIn profile.
Document whether refusals cite policy or simply state insufficient information.
Bottom line: The case confirms that frontier models still treat most personal family details as out-of-scope, whether by data absence or policy.
Future releases may narrow the gap through retrieval systems, but private biographical data will remain deliberately excluded from training pipelines.
Top comments (0)