InTheWeights Rates People on LLM Familiarity
Visibility in training corpora matters for practitioners because it can bias model outputs, affect attribution, and change evaluation of model fairness. According to a July 6, 2026 post on Overcoming Bias, the website InTheWeights.com assigns people a 0-1000 score that estimates how well they are known by large language models. The blog post notes the author's unusually high score, and recounts sample scores for colleagues and acquaintances. The post attributes the high rating to factors including writing on broadly-discussed topics and taking distinctive positions that create repeated textual fingerprints; it also suggests that current LLM training may underweight source prestige compared to sheer volume of discussion, per Overcoming Bias. The author furnishes examples and a short list of comparative scores to illustrate the point.
Editorial analysis
For practitioners building, auditing, or deploying LLMs, measures of "who is visible to models" are a practical probe of dataset composition, information retrieval behavior, and downstream bias risk. Simple scalar ratings like those reported by InTheWeights.com surface how frequently a person or notion appears across training-like corpora, which in turn can affect model completions, hallucination targets, and mistaken attribution.
What happened
According to a July 6, 2026 post on Overcoming Bias, the website InTheWeights.com scores individuals on a 0-1000 scale estimating how well they are known by large language models. The post reports the author's own score as unusually high and includes comparative scores for numerous colleagues and acquaintances to illustrate distributional differences, per Overcoming Bias. The post also offers a candid hypothesis that the author's high score reflects long-term writing on widely debated topics and clear, repeatable positions; the post further suggests that current LLM training regimes may not weight source prestige as humans do, so lower-status but extensive online discussion can boost a name's footprint in model data, per Overcoming Bias.
Editorial analysis - technical context
Scalar "visibility" metrics are noisy but useful signals for auditing. They conflate multiple data-generation factors: raw token frequency, co-occurrence structure, cross-posting and quote chains, and the presence of derivative summarizations. As a result, a high score can reflect either concentrated primary-source material or diffuse secondary discussion. Practitioners should treat single-number scores as prompts for deeper corpus sampling rather than definitive provenance proofs.
Editorial analysis - implications for modeling and evaluation
Visibility influences model behavior in two concrete ways. First, high-visibility persons and terms are more likely to be surfaced confidently, increasing both correct recall and confident hallucinations. Second, evaluation sets that include publicly visible figures risk overestimating model competence on less-documented populations. Observability tools like InTheWeights are therefore valuable as triage instruments for dataset curation and benchmark design.
What to watch
observers should track whether such tools document methodology (corpus sources, query strategies, time window) and whether model providers publish alignment between measured visibility and model output frequency. If the methodology is opaque, scores remain suggestive rather than diagnostic.
Key Points
- 1Visibility metrics reveal training-corpus concentration, which can raise both recall and confident hallucination rates for high-scored subjects.
- 2Single-number scores are coarse; practitioners need corpus sampling and provenance checks to diagnose why a person scores highly.
- 3Tools measuring LLM familiarity are useful triage for dataset curation, benchmark design, and bias auditing in deployments.
Scoring Rationale
A public tool that quantifies how well people appear in model training-like corpora is practically useful for dataset auditing and bias checks, but it is not a major technical breakthrough.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
