Skip to content
GoldCompass

AI disclosure

Where we use AI, and where we deliberately don’t

The market analysis on this site is drafted by an AI system and reviewed by a person before it is published. We put that at the top rather than in a footnote, because on a site about money the reader should get to decide how much that matters. This page sets out exactly which parts are AI-written, which parts a model never touches, and what can go wrong anyway.

The dividing line

The single most important design decision we made is that the model explains the market but never reports it. Retrieval and reasoning are kept apart: the numbers are fetched deterministically and handed to the model as ground truth, and the model writes analysis on top of them. A language model asked to recall a gold price will sometimes produce a plausible one, and a plausible price is worse than no price.

AI drafts this

  • The written analysis on the outlook page — the market narrative, the reasoning behind each call, and the risks section.
  • The BUY / HOLD / SELL signals themselves, their confidence levels, and the invalidation conditions attached to them.
  • The insight articles: the topic is chosen by a person, the draft is written by the model.
  • The key-level labels that summarise support, resistance, and notable prices.

AI never touches this

  • The spot price, the daily change, every point on the 30-day chart, and the macro figures — the dollar index and Treasury yields. These come from market data feeds and are passed to the model as fixed inputs it is told not to contradict.
  • Every figure the calculators produce. That is deterministic arithmetic running in your browser — no model is involved at any point.
  • The source list under each piece. Those are the URLs the model reported using, published unedited so you can check them yourself.
  • This page, the methodology page, the disclaimer, and the rest of the site's editorial copy.

What the drafting system is

The drafting is done by a general-purpose large language model from a commercial AI provider, accessed over its API. It is given a web-search tool so it can take account of recent market news, capped at a small number of searches per run, and it is required to return the sources it actually used.

We describe it that way rather than naming a version deliberately. Which model we use is a configuration value we revisit as better ones become available, and a disclosure page naming last quarter’s model would be less accurate than one describing what the system actually does. No other AI system is involved in producing what you read here, and we do not use AI-generated images.

What does not change is the constraint around it. Whichever model is in use, it is handed the numbers rather than asked for them, it must cite the sources it relied on, its output must satisfy a strict schema, and a person approves the result before it publishes. Those guarantees belong to our process, not to any vendor’s.

What happens between the model and the page

Model output is not published as it arrives. Five things happen to it first, and any of them can stop it.

  1. Market data is fetched first and injected into the prompt as authoritative ground truth, with an instruction not to contradict or re-estimate it.
  2. The model researches and drafts, then must return a single structured object — not free-form prose — containing the analysis, the calls, the key levels, and its sources.
  3. That object is validated against a strict schema. A missing signal, a missing rationale, or an article with no sources is rejected. One corrective attempt is allowed; after that the draft is discarded and the previously published version stays live.
  4. HTML is stripped from the text before storage. We keep Markdown and render it through our own components, so nothing retrieved from the web can inject markup into a page.
  5. The result is opened as a proposed change for review. A person reads it and decides whether it publishes. Nothing reaches the site on its own.

What “human reviewed” means here

The reviewer checks that the quoted figures match the fetched market data, that the cited sources exist and support the claims made, that the reasoning is consistent with the signal, and that the tone stays educational rather than promotional. Anything that fails is corrected or discarded.

It is worth being equally clear about the limits of that. Review is a competent editorial check on market commentary — not a professional audit, not a verification of every underlying fact, and not something that can make an uncertain forecast certain. Our methodology page covers the surrounding process in more detail: data sources, signal definitions, and update cadence.

Nothing you type goes to a model

Content generation happens on a schedule, away from the site, and it takes no input from visitors. There is no chatbot here and no feature that forwards what you enter to an AI provider.

The calculators run entirely in your browser — your budget, your holdings, and the figures you enter are never sent to us at all, which is also why a shared calculator link carries its inputs in the URL rather than on a server. The only thing you can submit is a newsletter email address, which goes to our email provider and nowhere near a language model.

What can still go wrong

Every mitigation below leaves something behind. Listing only the safeguards would be the dishonest version of this page, so each row states what remains after ours have done their work.

The model states something confidently that is wrong

What we do: Hard numbers are supplied to it rather than requested from it, and a reviewer checks the claims against the sources before anything publishes.

What remains: Review is an editorial check, not an audit. A plausible-sounding misreading of a genuine source can survive it.

A source is cited that does not support the claim

What we do: At least one source is required or the draft is rejected, and the reviewer follows the citations.

What remains: A source that is real and roughly on-topic but does not actually support the specific sentence is the hardest case to catch.

Web search returns low-quality or manipulated content

What we do: The prompt requires reputable sources, and any HTML is stripped from what the model returns before it is stored, so retrieved content cannot inject markup into our pages.

What remains: We cannot rule out a well-dressed but poor source shaping the framing of a piece.

The analysis reads as more certain than the evidence justifies

What we do: Every call carries an explicit confidence level and a stated condition that would invalidate it.

What remains: Fluent prose is persuasive regardless of whether it is right. Treat confidence as a description of the evidence, not a probability.

The market moves after publication

What we do: The outlook regenerates daily and every page shows when it was last updated.

What remains: Anything you read is a view as of its timestamp, not as of now.

The model produces a broken or incomplete draft

What we do: Drafts are validated against a strict schema. A draft missing a required field is rejected, retried once, and otherwise discarded — the previously published version stays up.

What remains: This one fails safe: the visible outcome is an older outlook, not a corrupted one.

Things we will not do

  • Publish AI-drafted analysis without a person approving it.
  • Let a model supply a price, a percentage, or any other hard number we display.
  • Present AI-written text under a fabricated human byline or persona.
  • Use AI to generate personalized advice — we cannot see your circumstances and do not try to.
  • Send anything you type into this site to a language model.

Important: Everything on GoldCompass is educational information only — not financial, investment, tax, or legal advice. That an analysis was reviewed by a person does not make it advice, and does not make it right. Read our full disclaimer.

Last updated . We will revise this page if the model, the pipeline, or the review process changes.