AI Disclosure & Transparency

Version 1.0 · Effective July 5, 2026 · Last updated July 5, 2026

1. You are interacting with AI

The study evaluator's "AI assist" feature uses artificial intelligence — specifically a large language model — to add reasoning and a plain-language summary to your evaluation. When you enable AI assist, you are interacting with an automated system, not a human. This disclosure is provided consistent with AI chatbot-disclosure laws (e.g. California's Bot Disclosure Law, Cal. Bus. & Prof. Code §17941) and the EU AI Act's transparency obligations for systems that interact with natural persons (Article 50).

2. What AI assist actually does here

The Good Study Framework evaluator scores a study on 7 dimensions (design, causation, size, measurement, statistics, robustness, applicability) using deterministic, rule-based extraction and scoring — regular-expression and metadata-based checks against the study text and, where available, PubMed/Crossref registry data (e.g. publication type, retraction status). That scorecard runs for every evaluation and does not use AI. It is always available, whether or not AI assist is enabled, and whether or not you're signed in.

If you separately opt in to AI assist (a checkbox on the evaluation form), your submitted study text and the deterministic scorecard are sent to the AI provider named in Section 3, which returns:

• Confounder reasoning — plausible third variables that could explain an observed association, beyond what the deterministic checks can identify from text patterns alone.
• A second opinion on low-confidence dimensions — the deterministic engine flags specific dimensions it isn't confident scoring from the text; AI assist reviews only those flagged dimensions and may adjust their score, never the dimensions the engine already scored confidently.
• A plain-speak bottom line — a short, non-jargon summary of the biggest strength, biggest weakness, and whether the finding should change anyone's behavior.

AI assist never overrides the deterministic scorecard wholesale — it only augments the specific dimensions the engine marked as needing review, and adds narrative output alongside the score. If the AI call fails or is withheld (see Section 4), the deterministic scorecard is shown on its own.

3. Who processes your input

When you enable AI assist, the study text you submit (and the deterministic scorecard) is sent to the following third-party AI provider for processing: Anthropic. See our Privacy Policy for how we handle that data more generally, and the provider's own privacy/data-use terms for how they handle it on their side (in particular, whether your input is used to train their models — confirm this with your provider agreement and disclose it here).

4. Usage limits — AI assist is free today, with limits

There is no charge to use this Service or its AI-assist feature today. That said, AI assist is metered by two independent abuse controls, both designed so the deterministic scorecard is never affected — only the optional AI section is withheld when a limit is hit:

• Per-visitor monthly limit. Each visitor (tracked by IP address) gets a limited number of free AI-assist evaluations per calendar month. Once you've used your monthly allowance, further evaluations still run the full deterministic scorecard; only the AI-assist section is withheld until the next month.
• Site-wide monthly token cap. Independent of any individual visitor's usage, the Service enforces a hard ceiling on the total AI tokens it will spend across all users in a calendar month, as a cost/abuse backstop. If that cap is reached, AI assist pauses for everyone until the next month, regardless of any individual visitor's remaining allowance.

These are usage/abuse controls, not a paid pricing tier — you are not charged for AI assist, and reaching a limit only means AI assist is paused, not that your account is affected.

5. Limitations — please read

• AI-assist output can be inaccurate, incomplete, outdated, or biased. It reflects patterns in training data and the text you submitted, not verified facts about the underlying research.
• AI-assist output is not professional advice — not medical, scientific, statistical, legal, or financial advice — even if it's phrased confidently.
• Neither the deterministic scorecard nor AI assist is a substitute for peer review or your own critical reading of the study.
• You're responsible for independently verifying anything you rely on before acting on it, especially the "bottom line" and "behavior answer" outputs, which are intentionally opinionated summaries, not neutral facts.

6. High-risk & consequential uses

This Service's default AI disclosure and safeguards are not sufficient if you use (or let others use) this Service's output to make or materially inform a decision about a specific person in an area such as employment, credit/lending, housing, insurance, healthcare, or education. Depending on your jurisdiction and use case, that can trigger additional obligations, for example:

• The Colorado AI Act (SB 24-205) — impact assessments, consumer notice, and an opportunity to correct/appeal for "high-risk" AI systems used in consequential decisions.
• NYC Local Law 144 — bias audits and candidate notice for automated employment-decision tools used by NYC employers.
• The EU AI Act (Regulation (EU) 2024/1689) — Annex III "high-risk" obligations (risk management, human oversight, logging) if the system is offered to users in the EU.
• Illinois, and a growing list of other states, regulate AI use in specific contexts like employment interviews and insurance underwriting.

[TODO: this Service is built for general research literacy, not as a decision-support tool for the high-risk uses above. If a use case like that emerges, this page and the Service need feature-specific disclosures, a human-review/appeal path, and — for the regimes above — a documented impact assessment. Get counsel involved before shipping that, not after.]

7. Synthetic content

AI assist only generates text (reasoning and summaries) — the Service does not generate images, audio, or video, so synthetic-media labeling requirements (e.g. California's AI Transparency Act, SB 942) are not currently applicable. Revisit this if that changes.

8. Your choices

AI assist is opt-in per evaluation — it only runs when you check the "AI assist" box on the evaluation form. You can always use the deterministic scorecard on its own by leaving that box unchecked. There is currently no mechanism to request that your input be excluded from the AI provider's own model-training use, beyond what's described in their terms; confirm this with the provider agreement and update this section if that changes.

9. Contact

Questions about our use of AI: ajkaiserauer@gmail.com.