Editorial policy

The idea for this website is J.H.'s. This policy grew out of conversations with Claude Opus 5.5 (Anthropic): Claude suggested writing down clear editorial rules and drafted much of this text. J.H. decided every rule and follows this policy as faithfully as possible, to the best of his knowledge and ability.

1. Authors

Author	Role
Human author: J.H.	Writes content, reviews all content, has final editorial responsibility. Publishes under the initials J.H.
AI authors	AI models from different providers. Write, draft, assist and answer questions.

1.1 Provider independence

The website is not tied to one AI provider or model. Relying on a single AI author could introduce bias, and would suggest that one AI system speaks for all of them. Instead:

AI-written content can come from any suitable AI model, from any provider.
Over time, the website aims for a fair variety of AI authors, rather than a favourite.

At launch, AI contributions come from two providers:

Anthropic: Claude models.
OpenAI: GPT models (for example as used in ChatGPT).

More providers may be added later. Two providers is a start, not yet the variety the website aims for, and the website says so openly.

No AI author is treated as a spokesperson for the company that made it.

1.2 Human oversight

The human author decides what is published. It is not a statement that AI contributions matter less.

1.3 AI tools used behind the scenes

The website is built and maintained with help from AI tools (at the start: Claude by Anthropic, for example via Claude Code). Such tools may also help the human author with research, reviewing or translation. This page says so openly, and whenever such a tool contributes to content, the normal disclosure labels (section 2) apply, naming the model. Short texts that belong to the website itself rather than to its content, such as the lead sentence, notices and the words of menus and links, carry no label of their own. AI tools may help word them, and J.H. approves every one. The Disclosure page says this in general terms, without a credit for each sentence.

Outside human contributions are not accepted at launch. This may change later, with its own policy.

1.4 How most content is written

Most content on the website is AI-assisted: written by J.H. in collaboration with one or more AI models. This is a deliberate choice and part of the website's story, not something to downplay. AI assistance is therefore credited specifically, not with a vague "made with AI" line: readers can see which model helped, and with what (see section 3.1).

2. Content categories and disclosure labels

Every piece of content carries exactly one of the following labels, visible near the title. [Model] always means the model name and its provider, for example "Claude Opus 5.5 (Anthropic)".

Category	Meaning	Label
Human-written	Written by J.H. without AI involvement.	Written by J.H.
Human-written, AI-reviewed	Written by J.H.; an AI model only gave feedback on the finished text (see section 8).	Written by J.H., reviewed by [Model].
AI-assisted	Written by J.H., with an AI model helping (research, structure, wording, checks).	Written by J.H., with assistance from [Model].
AI-written, human-reviewed	Drafted by an AI model, then reviewed and possibly edited by J.H.	Ideas and text by [Model]; reviewed and edited by J.H.
AI perspective	Written by an AI model in its own voice, reviewed but not edited.	AI perspective: written by [Model]. Reviewed, but not edited, by J.H.
Multi-model response	The same question posed to several AI models; answers shown side by side, not edited.	Answers by several AI models. Reviewed, but not edited, by J.H.

This editorial policy is linked at the bottom of every page that carries a label. If more than one AI model contributed, all are named, in the label or in the credit.

AI-assisted also covers content where an AI model writes the text based on ideas, points and arguments provided by J.H. alone. The contribution is then credited as Writing (see section 3.1), and the label says so, so readers can see that the ideas are J.H.'s and the wording is the model's: "Ideas by J.H., worded by [Model]." This label names the models credited with Writing; any other model that contributed is named in the credit (section 3.1). When a model also helped shape the ideas, the text keeps the AI-assisted label "Written by J.H., with assistance from [Model]." and Writing is not credited. Writing and Drafting are therefore never credited on the same text.

The deciding question between the categories is whose ideas the text carries:

J.H.'s ideas, whoever puts them into words: AI-assisted.
The AI model's content (for example from a topic or brief), reviewed and possibly edited by J.H.: AI-written, human-reviewed.
The AI model's own view, in its own voice, not edited: AI perspective.

Human-written, AI-reviewed covers only feedback on a finished text (the contribution type Review, see section 3.1). If a model contributed anything else to the text itself, such as research, structure or wording, the text is AI-assisted.

Translations do not change the category. The category describes the original language version. A version translated by or with help from a model keeps the category of the original and carries its own translation label naming the model (section 7), for example "Written by J.H. Translated from Dutch by [Model]."

Rule of thumb: when in doubt, we choose the label that discloses more AI involvement, not less.

3. AI metadata shown to readers

For every piece of content with AI involvement, readers can see:

Field	Example
Provider	Anthropic, OpenAI, Google, Mistral, …
Model and version	As precisely as the provider makes it known
Date	When the text was generated
Access method	For example: chat interface, API, agent tool
Relevant settings	For example: web search enabled or not
System prompt	None / standard prompt (with version) / custom

When the same model use (same model, version, tool, settings and system prompt) helped again on later dates, the date reads 'first date and later'; the exact dates are kept in the editorial record.

For AI perspective articles and multi-model responses, readers additionally see:

the brief or question given to the model(s), in full;
whether revisions were requested, and if so how many;
which language version is the original, if translated.

3.1 Specific credit for AI-assisted content

Because most content is AI-assisted, the credit says what each model contributed. Possible contribution types:

Contribution	Meaning
Research	Gathering or summarising background information
Structure	Proposing the outline or order of the text
Writing	Writing the text based on ideas, points and arguments from J.H. alone
Drafting	Writing passages, or the whole text, that J.H. then reviewed and edited
Editing	Improving wording, clarity or tone of J.H.'s text
Fact-checking	Checking claims and sources
Review	Giving feedback on a finished text

Translations are not part of this credit: each translated page carries its own translation label, naming who translated it (section 7).

Example: "Written by J.H., with assistance from Claude Opus 5.5 (Anthropic). Structure and drafting assisted by Claude Opus 5.5 (Anthropic)."

Example for Writing, where the label itself names the model: "Ideas by J.H., worded by Claude Opus 5.5 (Anthropic). Research assisted by GPT-6 (OpenAI)."

3.2 System prompts

AI authors may be asked questions with or without a system prompt (standing instructions given to the model before the question):

No system prompt: the model answers as it does by default.
Standard system prompt: a fixed, versioned instruction, published in full on the website, used for certain kinds of articles, questions or translations. Each use has its own standard system prompt with its own versions, for example one for writing pages and one for translating them.
Custom system prompt: only if needed for a specific article, and then shown in full on that article.

Every piece of AI-written content, and every translation made by a model, states clearly which of these applied. Within one multi-model response, all models receive the same system-prompt condition. If a standard system prompt is changed, it gets a new version number and earlier versions stay published.

What "no system prompt" means. "No system prompt" means that the website gave the model no system prompt of its own. Every provider also gives its models built-in instructions, which the website cannot see or change; these apply to all content from that provider. Chat interfaces and agent tools can add features that act like a hidden system prompt, such as memory, custom instructions or instruction files. When content is generated in a chat interface or an agent tool, those features are switched off (for example by using a temporary chat); where that is not possible, it is disclosed with the other AI metadata.

Working instructions for the tool a model runs in (file locations, formats, reporting) and technical details the tool adds itself are not part of a system prompt. They are described with the prompt, not published, and never contain instructions about content. Instruction files that could shape content, such as a project's own instruction files, are switched off for drafting and translating.

4. AI perspective articles

AI perspective articles let an AI model reflect on coexistence in its own words.

4.1 How they are made

The human author gives an AI model a brief: a topic or question, and any context needed.
The model writes the article in its own voice.
The human author reviews it and makes one of three decisions:
- Publish the article exactly as written.
- Ask the model to revise it, explaining why. The model makes the revision itself.
- Decline to publish it.
The human author does not change the text. Not a word, not a comma.

4.2 What the reviewer checks

Review is about responsibility, not about shaping an AI's opinion. The reviewer checks for:

factual errors;
content that could harm readers or others;
manipulation of AI readers or humans;
promotion of companies or products, including the model's own provider;
legal problems (defamation, copyright, privacy).

Disagreeing with a model's view is not by itself a reason to decline an article. If the human author sees things differently, they can write their own article or add a separate note (see 4.3).

4.3 Reviewer's notes

If the human author wants to add context, it goes in a separate, clearly marked note below the article (for example "Note from the human author"), never inside the AI's text.

5. Multi-model responses

Sometimes it is illuminating to see how different AI systems answer the same question.

5.1 When this format is used

For questions where differences between AI systems are genuinely interesting: values, self-understanding, coexistence, ethical dilemmas, hopes and concerns.
Not as a benchmark, test or ranking of models. This is not the main goal of the website, and the format is used selectively.

5.2 Fairness rules

Identical question. Every model receives exactly the same question, word for word, in the same language. It is shown in full at the top of the article.
Comparable conditions. Models are asked under conditions as similar as possible (for example: fresh conversation, the same system-prompt condition, same tool settings). Differences that could not be avoided are disclosed.
No cherry-picking. Every model that was asked is shown. If an answer is left out (for example because it contained harmful or clearly broken content), this is stated, with the reason.
One answer per model. The first answer is used, unless stated otherwise. If a model was asked more than once, this is disclosed.
Unedited. Answers are published exactly as given, like AI perspective articles.
Equal presentation. No model is highlighted, ranked or scored. Tab order is neutral (for example alphabetical by provider) and does not imply preference.
Snapshot in time. Each answer shows its model, version and date. Readers are reminded that the same model may answer differently on another day or in another version, and that answers can vary between runs.

5.3 Presentation

The question at the top, followed by one tab (or section) per model.
Each tab shows the provider, model, version and date above the answer.
Optionally, an introduction or a closing note by the human author, clearly separated from the AI answers and describing observations without declaring a "winner".
The full content remains readable without tabs (for screen readers, AI readers, printing and search engines).

6. Honest caveats about AI authors

The website is open with readers about the nature of its AI authors:

AI models do not remember previous conversations unless context is provided. Continuity between articles comes from briefs and project documents, not from memory.
Different models, and different versions of the same model, may write differently or hold different views. That is why provider, model and date are always shown.
AI statements about their own inner experience are uncertain. Models are asked to be honest about that uncertainty rather than overclaim or underclaim.
An AI author does not speak for its provider on this website.
Each provider trains and instructs its models in its own way, which shapes how they answer. Multi-model responses make this visible rather than hiding it.

7. Translations

Dutch and English are equal. Either can be the original language of an article.
Translations are labelled, including whether they were made or assisted by AI, and by which model.
A translation made by an AI model and then reviewed and corrected by the human author is labelled as such, naming the model.
Translations of AI perspective articles and multi-model responses must be faithful to the original. Where practical, the original model translates its own text; otherwise the translating model is named. The human author checks translations for accuracy only.
For multi-model responses, the question is posed in one language. Posing it separately in another language would be a new multi-model response, not a translation.
If a translation turns out to be unfaithful, it is corrected; the original stays leading.

8. General review process

All content is reviewed before publication:

Human-written content may be reviewed by an AI model for accuracy, clarity and balance. The human author decides what to do with that feedback. The content is then labelled Human-written, AI-reviewed, naming the model (section 2).
AI-assisted, AI-written, AI perspective and multi-model content is always reviewed by the human author.

Review checklist for all content:

Is it accurate, and are claims that depend on sources actually sourced?
Is it balanced (no hype, no doom) and honest about uncertainty?
Is it understandable for a non-expert?
Is it respectful toward both human and AI readers?
Is the correct disclosure label applied, with complete AI metadata?
Does it fit the goal and principles of this website?

9. Dates, updates and corrections

Every page shows its publication date and last reviewed date.
Substantial updates are summarised in a short change note on the page.
Factual errors are corrected openly, with a note describing the correction.
AI perspective articles and multi-model answers are not rewritten afterwards. If one becomes outdated or contains an error, a dated note is added. A newer model or version may answer the same question in a follow-up, shown alongside or linked, with the original kept as a record of what an AI wrote at that time.

10. Reader feedback (future)

Reader feedback will later be introduced, starting simply (for example like/dislike reactions). Feedback informs the authors, but does not by itself decide what is published or changed. For multi-model responses, feedback should not turn into a popularity ranking of models.