AI Transparency Policy

Last updated: May 20, 2026

Framework: NIST AI RMF 1.0 (GOVERN function) System: AI-Assisted Blog Post Generator — torreyladams.com Organization: Torrey Adams Version: 1.1 — May 2026

Advisory Notice: This document is provided for informational and governance transparency purposes only. It does not constitute legal advice or a formal compliance determination. For jurisdiction-specific regulatory guidance, consult qualified legal counsel.

System Overview

The AI-Assisted Blog Post Generator is a content production workflow used to research, draft, and publish AI news and governance commentary on torreyladams.com. The system takes a source input — either a curated briefing document or a web research query — and produces a structured, SEO-optimized blog post in MDX format. Every post generated by this workflow is created as a draft and reviewed by the site author before publication.

This system is not used to make decisions about any reader, visitor, or third party. Its sole output is editorial content for a personal portfolio blog.

Intended Use

Primary intended use: Drafting AI news analysis and commentary posts for torreyladams.com, drawing on source documents and web research to produce structured, human-reviewed editorial content.

Intended users: The site author (Torrey Adams), acting as sole operator, editor, and publisher.

Intended deployment context: Personal portfolio blog covering AI technology, governance, and policy. Posts are published for a general technical audience.

Out-of-scope uses: This system is not designed for — and should not be used for — generating content that will be published without human editorial review, producing content that makes claims about specific individuals without factual basis, or generating legal, medical, or financial advice.

Foreseeable misuse: The most likely misuse is publishing AI-generated drafts without editorial review. The workflow mitigates this by setting all posts to draft: true at creation, requiring explicit author action to publish.

System Description

The workflow is powered by Claude (Anthropic), accessed via the Cowork desktop environment using a custom ai-blog-post skill. The pipeline operates as follows:

  1. Source intake — a .docx briefing document or web research query is provided as input. If a .docx is used, embedded author metadata is stripped before content extraction.
  2. Research — web searches are conducted to supplement source material with supporting facts, data points, and context.
  3. Drafting — the model generates a structured 450–550 word MDX post with Keystatic-compatible frontmatter.
  4. Internal linking — the post is checked against a local post index and relevant internal links are woven into the body copy.
  5. Human review — the draft is saved with draft: true and reviewed and edited by the author before any publication action.

The system produces text only. No images, personal data, or automated publishing actions are taken without human initiation.

Human-AI interaction design: The author reviews every draft in full before publishing. The author may edit, reject, or substantially rewrite any generated content.

Training Data

This system uses Claude, a large language model developed and trained by Anthropic. Training data characteristics, coverage, and preprocessing are documented in Anthropic’s published model cards and usage policies at anthropic.com. The ai-blog-post skill itself does not involve additional fine-tuning or custom training.

Source documents used as runtime inputs are curated weekly briefings produced by the author. Web search results are retrieved at time of use and are not stored as training data.

Performance and Evaluation

Evaluation approach: Outputs are evaluated qualitatively by the author at each draft review. No automated benchmarks are applied to published posts.

Known performance characteristics: The system produces factually grounded content when source documents are high quality. Output quality degrades if source documents contain errors, outdated information, or ambiguous claims — these are caught at the human review stage.

Known limitations: The model may occasionally misattribute a quote, overstate the certainty of a regulatory outcome, or produce a framing that does not fully reflect the nuance in a source document. The author’s editorial review is the primary control for catching these issues before publication.

Failure modes: Posts generated without a high-quality source document may produce plausible-sounding but unverified claims. The workflow is designed to always start from a curated source; ad-hoc generation without a reference document is discouraged.

Risks and Limitations

Identified risks:

1. Factual inaccuracy. AI-generated drafts may contain subtle errors in dates, attribution, or regulatory interpretation. Severity: medium. Likelihood: low when a high-quality source document is used.

2. Editorial voice drift. Repeated use of AI drafting may gradually shift the blog’s tone and perspective away from the author’s authentic voice. Severity: low. Likelihood: low with active editorial review.

3. Source reliability. Web research retrieved at generation time may include outdated or inaccurate third-party information. Severity: medium. Likelihood: low for well-established regulatory topics; higher for breaking news.

Mitigations in place: All posts are created as drafts requiring explicit author review. The workflow strips source document metadata. Source documents are author-curated before use. Internal links are verified against a known-good post index.

Residual risks: The author’s review is the final control. No automated fact-checking is applied to web-sourced claims beyond cross-referencing with the provided source document.

Human Oversight and Accountability

All content published on torreyladams.com under this workflow passes through a mandatory author review stage. The author reads the full draft, verifies factual claims against the source document, edits for voice and accuracy, and takes explicit action to change the post status from draft: true to draft: false before publishing.

No post is published automatically. The author retains full editorial control and is accountable for all published content regardless of how it was drafted.

Readers who identify factual errors or concerns about a published post can contact the author directly via the contact information on torreyladams.com. If a published post is found to contain material factual errors, the author corrects or removes the post directly in the CMS.

Privacy and Data Handling

The blog post generation workflow does not collect, store, or process personal data about readers or visitors. Source documents used as inputs are internally produced briefings. No reader data is used as input to the AI system.

Source .docx files have embedded author metadata stripped prior to processing (author name, last-modified-by, revision history). Stripped files are stored locally and are not shared externally.

Applicable frameworks: GDPR (EU), for any EU-based visitors to the blog — standard site analytics and cookie policies apply separately and are outside the scope of this disclosure.

Regulatory Classification

EU AI Act (Regulation 2024/1689): This system is not classified as high-risk under Annex III. Content generation for a personal portfolio blog does not fall within the high-risk categories defined by the EU AI Act. Article 13 transparency obligations do not apply. This assessment will be revisited if the system’s scope or use context materially changes.

Maintenance and Updates

The ai-blog-post skill is maintained by the author and updated as the blog’s formatting requirements, SEO standards, or content strategy evolve. The underlying model (Claude) is managed by Anthropic and may be updated by Anthropic independently. Material changes to how AI is used in content production on this blog will be reflected in an updated version of this model card.

No automated monitoring or drift detection is in place; the author’s ongoing editorial review serves as the primary quality signal. If this workflow is discontinued, the AI Disclosure footer on all previously published posts will remain in place.

Contact and Further Information

Technical and editorial contact: Torrey Adams — contact

Model provider documentation: Anthropic — Claude Model Card

Version history:

  • 1.0, May 2026 — initial publication
  • 1.1, May 2026 — added advisory notice, EU AI Act classification statement, end-of-life note, and incident correction language; aligned with NIST AI RMF model card (governance/model-card-blog-post-generator.md)