About AI

ai-waw.com is produced with the help of artificial intelligence tools and describes that help openly. This page sets out which tools are used, for what, in what workflow and within what limits. Status of this description: 10 October 2026.

Sections: Principle and roles · Tools · Workflow · What AI does not do · Quality control · Data and privacy · AI as a reader · Transparency and the law · Errors and responsibility

Principle and roles

AI calculates, checks and drafts – the author decides and approves.

AI tools do most of the groundwork: they retrieve and process data, build balance models, write scripts and the portal code, prepare tables, figures, maps and text in Polish and English, and then check the result in separate runs.

The author of ai-waw.com chooses the topics, settles disputed assumptions, decides on theses and conclusions, and approves every version before publication. At the author's request AI may draft theses or conclusions; only wording accepted by the author is published.

Publications on ai-waw.com are therefore texts prepared with AI and subject to editorial control – not texts written without AI.

Tools

• Claude Code (Anthropic) – an agent running on the editorial computer. It is the only AI tool used for the portal's content and code.
• Model: Claude Opus 5.5. In the change history of the portal and report factory repositories every change is marked as co-authored by this model. If the model changes, this page will name the new one.
• Claude Code built-in tools – reading and writing working files, a shell for running scripts (Python, PHP), fetching web pages and documents, web search.
• A browser controlled through an MCP server – for checking portal pages on phone and desktop screens and in light and dark mode.
• Data publishers' interfaces – statistics (Eurostat: Comext, PRODCOM, demography) and EU legal acts (Cellar, EUR-Lex) are retrieved by scripts directly from the publishers' interfaces, with no intermediaries.

Figures, diagrams and maps are drawn by code from data, not generated by image models. English versions are produced with the same tool, from the same source of numbers as the Polish ones.

AI models are used only in editorial work, run from the editorial computer. The portal server does not call any AI model: its pages, API and MCP server serve content stored in the database, with no model-generated answers.

Workflow behind a publication

The portal has two kinds of publication with different workflows.

Reports with a fact register (the "Reports" catalogue) are produced in the report factory – a set of scripts run through Claude Code:

1. Data retrieval from source interfaces; every download is entered in a download register with a SHA-256 checksum.
2. Fact register: every number gets an identifier, unit, period, area and a classed source.
3. Typesetting of text, tables and figures: the text refers to facts and typesetting inserts the values.
4. Automated checks A1–A11 (see Quality control).
5. Author gate: without a summary, theses, conclusions and a conflict-of-interest declaration completed by the author, a release fails the checks.
6. A package signed with HMAC reaches the portal as an unapproved version.

Metal flow analyses (the Metal flows section) are produced in separate analytical projects:

1. Research agents – separate Claude Code runs – collect data from official statistics, company reports, industry association publications and legal acts into research registers. Each register records its rules: primary sources only, a page reference for each number, trade press only as a pointer, no names of private individuals.
2. Balance model: flows from data and from marked estimates are reconciled by weighted least squares, and the uncertainty of indicators is estimated by Monte Carlo simulation.
3. Generators produce text, tables, figures and maps in Polish and English from a single source of numbers.
4. Five independent reviews A1–A5, each in a separate AI run that does not change the analysis files but writes a report of findings: A1 – sources (documents reopened), A2 – trade data, A3 – balance model (reproduction of the result, alternative assumptions), A4 – figures and maps, A5 – text, language and portal page (control recalculations, automated comparison of numbers in the Polish and English versions).
5. List of corrections: every finding is implemented or recorded as open. Example: for the aluminium flow analysis of Italy the reviews produced 91 findings (2 blocking, 31 material, 58 minor), and the corrections changed the analysis's main indicators.
6. The analysis enters the editorial panel as a series P publication and awaits approval.

The five-review standard applies from the copper flow analysis for Poland onwards. The first analysis – copper flows in the EU-27 – was produced earlier and has no A1–A5 review records.

Two-stage publishing. Every package first becomes an unapproved version. Approval makes it the current version, moves facts and sources into the public registers and writes the SHA-256 checksum of the version manifest to the event log. Approved versions are never overwritten: a correction is a new version (e.g. 1.0 → 1.1) or an erratum. Publications are not deleted – they can be withdrawn. Limitation: figures and maps of the flow analyses are files deployed with the portal code and are not versioned together with the text.

What AI does not do

• It does not decide on theses, conclusions, assessments or what is published.
• It does not approve publications on its own initiative. Approval is the author's decision, made in the editorial panel. On the portal's test version approval is sometimes executed by Claude Code – only on the author's explicit instruction concerning a specific version. The flow import command cannot approve in the production environment.
• It does not deploy changes on its own – deployment to the server, sending a package and pushing code to the repository each require the author's consent every time.
• It is not the source of any number. A number has a primary source or is an openly marked model estimate (see Quality control).
• It does not use paywalled or login-protected content and does not bypass CAPTCHA. An inaccessible document is recorded as not retrieved, and its numbers are either not used or assessed indirectly.
• It does not collect personal data – the company and plant registers contain no names of private individuals.
• It does not use LME prices or price-reporting agency prices – the portal rejects a package with such a source.
• It has no access to server passwords and keys – the tool's configuration forbids reading the files that hold them.

Quality control of AI

The main risk in working with a language model is a number or claim that sounds plausible but has no basis in a source. Control therefore rests on making every number reproducible.

Source classes. Classes A and B are primary sources (official statistics, legal acts, company documents, study groups, industry associations). Class C covers trade press, releases without underlying data, and own estimates. Definitions are on the Method page.

Reports with a fact register. Every number in the text is a fact with a class A or B source. The portal rejects a package in which a fact has an unknown source or a class C source. Automated factory checks: A1 – references in the text point to existing facts; A2 – facts are complete; A3 – every number in the register is recalculated directly from the downloaded source files; A4 – a number typed into the text by hand is an error; A5 – author gate; A6 – privacy and licences; A7 – package consistent with the register; A8 – completeness of trade data months; A9 – figures; A10 – currency of the legal status; A11 – terminology glossary. Each check reports its coverage; a check that checked nothing does not pass.

Flow analyses. The numbers come from the reconciled balance and from the primary sources cited in the analysis. The balance flows, net imports by trade group, indicators and the key figures of the header have cards in the fact register (status reconciled, with a standard deviation, the indicators also with a 90% interval); they enter the register only when the author approves a version. Some flows are not measured by statistics – these are model estimates marked as class C, with a standard deviation, which enter the reconciliation together with primary data; their fact cards cite a class C source. Results are given with uncertainty ranges, and the analyses' annexes list the parameters and their sources. Review A3 reproduces the model result independently, and A5 recalculates a sample of the numbers in the text.

Portal. A privacy guard rejects packages containing data excluded from publication; automated tests and a scan of public responses check, among other things, the absence of cookies and of disallowed data.

What remains a risk. The reviews are carried out by the same kind of tool as the work, so errors can repeat in both runs; an AI review does not replace an industry specialist. Control recalculations of flow analysis texts cover a sample, not every number. Documents that could not be opened are assessed indirectly. Class C estimates depend on assumptions – which is why they carry wide ranges and are documented.

Data, privacy and security

• The model provider (Anthropic) receives editorial working material: source data, public documents, draft texts, code and calculation results.
• It receives no reader data. The portal does not call AI models and passes no information about visits to AI providers.
• Public pages use no cookies or tracking tools; scripts are served only from the portal's own server. The machine-access log (API, MCP, .md, .json) records the channel, resource address, response code, time and client name – without IP addresses or call arguments.
• The "Open in Claude" and "Open in ChatGPT" buttons on report pages pass to the external assistant only the address of the report's .md version with a request to read it – and only when the reader clicks them.

Details: Privacy.

AI as a reader – access for agents

The portal is built for AI assistants and agents:

• .md versions of static pages, reports and flow analyses, .json versions of reports with their facts and the facts of flow analyses (…/facts.json);
• API v1 with an OpenAPI description and an MCP server – see Method – machine access;
• a /llms.txt file with a content map.

Flow analyses are not in the API publication catalogue or in the MCP search (search_reports); their numbers are facts in the register – available through /api/v1/facts and the get_facts tool with the analysis identifier, e.g. AIWAW-2026-PCUPL.

Rules on model training are set per path. Without reservation, with attribution: report summaries, the fact register, the method page, and the overview and company cards of the "Companies" register. All other content is reserved, including full reports, metal flow analyses (except the facts.json files), datasets, feeds, the API, MCP and files. The reservation is expressed through TDMRep, robots.txt and /tdm-policy.json. AI assistants may retrieve content in response to a user's question, citing the publication or fact. Details: Terms of use.

Transparency and the law

Article 50(4) of Regulation (EU) 2024/1689 (the AI Act) requires deployers of an AI system that generates or manipulates text published with the purpose of informing the public on matters of public interest to disclose that the text has been artificially generated or manipulated. The obligation does not apply, among other cases, where the AI-generated content has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility for its publication. The information must be given in a clear and distinguishable manner at the latest at the time of first exposure (Article 50(5)). Article 50 applies from 2 August 2026 (Article 113); Regulation (EU) 2026/1744 did not amend paragraph 4.

The portal discloses the use of AI whether or not the editorial-control exception applies:

• this note is linked from the main menu and the footer of every page of the portal;
• every report has an "AI note" section on its page describing the workflow behind it – a mandatory field of the publication package; the portal rejects a package without one;
• flow analyses carry an AI note in the version approved in the editorial panel; their pages do not yet display it separately – the workflow behind them is described in the Workflow section;
• every version goes through editorial control and approval by the author of ai-waw.com, who holds editorial responsibility for the publications.

The portal does not hold conversations between readers and an AI system and does not publish AI-generated images, audio or video.

This note describes the portal's practice; it is not a legal opinion and does not determine how that practice is classified in law.

Responsibility and reporting errors

The author of ai-waw.com is responsible for the content of every publication – including the parts prepared with AI. Errors are corrected openly: the errata register gives the date, location, description and fact identifiers, and a changed number goes into a new version with a description of the changes.

Errors can be reported to the address on the Contact page – preferably with the publication and fact identifiers, or with the name of the analysis and the number of the figure or table. Comments on this note can be sent to the same address.