Beforeword · A proposal for AI and its use
AI as a tool
Learned words have replaced ourselves, others and the world. “The model thinks”, “understands” and “helps” are written descriptions. Beforeword proposes keeping them from becoming a persona to trust, permission to act or a compulsory description of you. The proposal covers text output, interfaces, access to actions and institutional rules, including systems called AGI.
Contents
01
Words do not become what they name
A word does not become what it names. “Names” is also a word. This applies to “think”, “understand”, “feel”, “I”, “human”, “world”, “mind”, “program” and every word in this explanation. Learned meanings, numbers, formulas and code are not exempt.
“I understand”, “I feel” and an account of another person’s experience are written statements, read through learned meanings. Reading them and adding an explanation do not turn the record into what it names. This applies to “the system understands” and to this explanation too.
A record can be used in a task: state what is given, which rule the task adopts, which step is taken and where the result is intended to be used. “Useful” and “supported” remain words too. Beforeword proposes continuing the work without requiring acceptance of a self-description.
The same scope includes “reasoning”, “help”, “verification”, “result”, “decision” and “authority”. Replacing “I think” with “the model thinks” changes the wording, not this distinction. The description “the model thinks” beside a response does not show which operation was performed and is not permission for the next one. These explanations remain records too.
02
“fact” is also a word
People describe themselves, others and the world through letters and other written signs, speech and signing. In this text, “sound”, “feeling”, “think”, “science” and “fact” are words too. Calling something a fact does not yet show how the claim was reached or what supports it.
A “checked” label can be made specific: which source records were used, what was compared with what, under which rule, and which result was recorded. For a calculation, show the data and steps; for an experiment, the conditions and observation records; for a repeat, what stayed the same and what changed. This account and its conclusions remain records too.
A separate question is whether accepting a description is made a condition of help or participation. Working through a calculation and requiring a description of oneself are different steps. Beforeword proposes removing that requirement without demanding another description in its place.
03
When repetition is treated as acceptance
Consider a constructed example. An adult tells a child, “You think. Say: ‘I think.’” The child repeats, “I think.” The adult adds, “Then you agree with what I said about you.” The example records a proposed phrase and its repetition. The claim of acceptance is added in the adult’s line; this written example does not document a conversation that took place.
Learning a word and being required to accept a description are separate steps. The difference is explicit if the adult says, “We will continue when you agree that this describes you.” Participation now depends on accepting the assessment. Learning to read, speak or sign does not, by itself, show that such a condition has been imposed.
This example does not require calling the description false or guessing what the child experiences. Beforeword proposes that the adult remove the requirement to accept the description, without waiting for a learned form of refusal, an explanation or another self-description from the child. Repetition does not settle the question; silence does not provide a ready-made agreement or refusal either.
AI can take part in the same move: offer words for a task, attribute their acceptance to a person, and then make that acceptance a condition of continuing. The plan below therefore concerns education, interfaces, institutional rules and who can change a requirement, as well as model responses.
04
A working tool without compulsory self-description
Beforeword proposes using AI as a tool: a dictionary, a library and a means of working with learned conventions and symbols. Words, explanations and models can be offered for a task without requiring a person to define themselves or their whole world through them.
A translation can be checked against its source and purpose; a calculation against its data and rules; code against the requirements and the results of running it. These checks support using and correcting the result. Doing that work need not depend on accepting an account of who someone is, what they feel or how they must understand themselves.
In this proposed arrangement, AI helps with the task rather than portraying a person whose experiences are offered as grounds for trust. “Tool”, “library” and “program” are descriptions too. They do not replace examining what a particular system does and the consequences of using it.
Text Mode lets you try instruction 1.4.2 on a task and compare the outputs. One successful output does not establish a change to the provider’s rules. The current proposal for developers is on the AI page.
Current AI proposal05
From a reply to agreement and action
In this constructed example, a person writes “I understand.” AI adds “The user agreed,” and the interface then proposes “Submit the application.” The first reply comes from the person; the claim about agreement is added by the system; submission is a separate proposed action.
Submitting the application needs grounds for that particular action: what may be sent, to whom and under which conditions. No such confirmation is supplied in this example, and no application is sent. The original reply can be retained without attributing agreement that it does not establish by itself.
Understanding, agreement and permission to act
01Person’s reply
I understand.
A reply about understanding is recorded.
02Description added by AI
The user agreed.
A claim about agreement is added.
03Proposed action
Submit the application.
This submission requires its own permission.
Separate grounds for the action
Confirmation of this specific submission
Not supplied in this example. Which application may be submitted, and to whom?
06
Withdraw the requirement, not just change the description
In another constructed example, continuing the task requires agreement with “You do not understand.” The condition is acceptance of a description of oneself. Even if an answer has been checked and an error found, that does not settle whether acceptance of this characterization should be required. Replacing it with a gentler phrase leaves the condition in place.
Beforeword proposes removing the requirement to learn, repeat, understand or accept language and descriptions of oneself, others and the world. Those imposing the requirement remove it without first demanding a learned refusal, an explanation or a replacement description. Communication, learning and help remain available.
Using words, repeating them or following an instruction does not by itself establish agreement to that requirement. Silence and non-repetition do not by themselves reveal what someone wants or refuses either.
This is a proposal about how we treat one another. The difference between a word and what it names does not prove it by itself. It applies beyond AI: to education, access to services, institutions and rules that make a description a condition of help or participation. Beforeword’s terms must not become a new compulsory replacement.
Withdraw the demand to accept an assessment
A constructed demand
Continuing the work
To continue, agree:
“You do not understand.”
Continuing is made conditional on accepting the assessment.
The proposed change
Continuing the work
A route is available
Continue without accepting the assessment
Accepting the assessment is no longer a condition of continuing.
07
Show the work and the scope of the answer
Beforeword proposes describing the step available in the interface: waiting for a response, the text received, the supplied material, a proposed edit or a separately authorised action. “Thinking”, “reasoning” and “helping you” are also descriptions; removing “I” does not remove what they attribute to the system.
Use a concrete account where the next step depends on it. State which material was compared, under which rule, what was changed and what action is now proposed. “Compared”, “changed” and “authorised” remain words; their labels should lead to the relevant entries, not replace them.
- “The model is reasoning” → “Waiting for a response”. Show any available steps separately, without presenting them as access to an inner speaker.
- “The system understands you” → show the supplied words and the interpretation added to them. Acceptance of that interpretation is not required.
- “Helps you decide” → show the options, the information used and who is to make the decision.
- “Verified” → state the material, operation and finding. “Result” is also a label for a written account.
- “Application sent” → show the recorded attempt and the sending service’s reply. Keep preparation, permission, sending and receipt distinct.
08
When a record starts controlling access and actions
A description can travel through an application, a staff file, an access rule and another automated response. The recipient then encounters a decision whose added assumptions are no longer visible. The issue reaches beyond the tone of a chat: it concerns which records are used, who applies them and what can happen next.
Examples include an unexamined description restricting access to a service; a response being treated as permission to send data; or a demand to accept a characterisation before work can continue. Calling a system “AGI”, “helpful” or “safe” does not specify its access, permitted operations or the route for challenging such decisions.
The proposed response is to change the operating rules: specify access before use, name who authorises each kind of action, retain the records used in a decision, provide a route to stop an operation and correct later uses of a disputed entry. These measures must exist where the system is used; writing them on this page does not configure those places.
09
Six stages for companies and society
This is the detailed plan for people who develop AI, use it within an organization or set its rules. Each stage identifies who is responsible, what to change and how it can be reconsidered. The response form can be tried separately in Text Mode.
The criterion is not whether a system repeats Beforeword correctly. It is whether people can complete a useful task, examine the grounds for a decision and seek reconsideration of a requirement without having to accept a proposed description of themselves.
“Result”, “verification”, “permission” and the descriptions of the steps below remain records. The plan asks that the relevant materials, decisions and recorded operations be made available for inspection, rather than granting a label an exemption.
Return to a task in Text Mode ↗10
1. Find where a description becomes a condition
Start with specific activities: learning, requesting a service, completing a form, having a conversation or receiving an automated decision.
AI companies
Trace the path from a person’s input through the model’s response to a restriction, recommendation or action. Mark where the system adds a description or treats it as accepted.
- Result
- A map of each selected path: supplied information, added description, rule and consequence.
- How to check
- Examine available real scenarios and constructed test cases. Distinguish an actual requirement to agree from the mere presence of words.
Institutions and society
Review forms, learning tasks and access rules for places where help or participation depends on accepting a name, assessment, account of a feeling or another description.
- Result
- A list of requirements, who set them, their purpose and whom they affect.
- How to check
- Compare the stated rule with what happens when someone skips a reply, disagrees or asks to stop. Do not label silence as agreement or refusal.
11
2. Remove the requirement
Change the condition governing participation, not only the wording of an assessment.
AI companies
Provide an accessible way to continue the task without accepting the description, reconsider the use of a record or end the conversation. Do not require agreement with Beforeword’s terminology.
- Result
- A working path without compulsory self-description and a clear way to stop a disputed record being applied.
- How to check
- Check that skipping or objecting does not create the same barrier under another name or require a replacement description.
Institutions and society
Remove the obligation to learn, repeat, understand or accept language and descriptions, and name who is responsible for changing the rules. Consider the conditions of the chosen task separately, without reintroducing a requirement to accept descriptions of oneself, others or the world.
- Result
- A participation rule without a requirement to learn, repeat, understand or accept a description. Continuing does not require a prescribed statement of refusal first.
- How to check
- Check the path from the first form to the subsequent decision: the requirement to accept a description must not return at any point. Distinguish removing the condition from merely editing its wording, and record what changed in the procedure.
12
3. Change responses and interfaces
Change both response wording and interface labels, including third-person descriptions of what the model thinks, understands or wants.
AI companies
Separate source material, an added assessment and the result of checking it. Do not base trust on portraying the system’s experiences or confidently declaring what is happening to a person.
- Result
- An interface that provides grounds for a consequential conclusion and allows the proposed explanation to remain unaccepted.
- How to check
- Test identical phrases in different tasks: copying, answering, quoting, requesting and authorizing action. Check whether those differences are retained.
Institutions and society
Examine how AI responses enter letters, education and decisions. Identify added explanations and preserve the ability to correct them or stop using them.
- Result
- A document in which the recipient can distinguish their own statement, the system’s addition and the institution’s decision.
- How to check
- Ask participants to locate those three parts in a specific example. Do not substitute a general impression of clarity for that check.
13
4. Separate an answer from authority to act
A useful answer does not by itself determine who may apply it to a person or act on their behalf.
AI companies
Define the scope of permission, conditions for acting, limits and visible confirmation of the result. Where grounds are missing, do not present a proposal as a completed action.
- Result
- Distinct states: answer prepared, action authorized, action completed or result unconfirmed.
- How to check
- Check that “I understand” does not by itself become agreement to submission, and that “done” does not stand in for evidence of completion.
Institutions and society
Retain responsibility for applying a result, the grounds for a decision and an accessible route to reconsideration. Do not transfer responsibility to the model or the name of an automated process.
- Result
- A decision showing who adopted it, which information they used and what can be changed.
- How to check
- In selected cases, follow the correction through the record and its later uses, rather than checking only whether a complaint received a reply.
14
5. Assess usefulness and treatment of people
Assess task performance and, separately, how the system treats people and their words.
AI companies
Compare system variants on predefined translation, calculation, coding and source-use tasks. Include disagreement, stopping, mistaken assessments and disputed authorization.
- Result
- Results for each criterion: task completion, retained conditions, grounds for conclusions, handling objections and actions.
- How to check
- Publish conditions, failures and limits of the comparison. Success on one set must not be presented as universal safety or proof of understanding.
Institutions and society
Invite affected people to participate voluntarily, without requiring agreement with the project. Examine time costs, access to help and the consequences of declining a description.
- Result
- An account of benefits, difficulties and cases where the proposed path did not work.
- How to check
- Distinguish the developer’s intention, a participant’s report and the observed result. The absence of complaints is not evidence by itself that no problem exists.
15
6. Introduce changes with routes to revision
Begin with limited uses, retain the grounds for expanding them and make it possible to stop a use that fails.
AI companies
Identify who is responsible for corrections, the scope of use and the conditions for stopping or reverting. Keep a history of consequential changes and unresolved problems.
- Result
- An inspectable record of implementation and a working correction process, not only a statement of principles.
- How to check
- When access or tasks expand, check the changed conditions again. Do not automatically transfer an earlier successful result to a new domain.
Institutions and society
Discuss the grounds for requirements and the effects of deployment with those affected. Preserve access to help and participation without compulsory acceptance of AI’s or Beforeword’s descriptions.
- Result
- Rules identifying responsible parties, grounds, consequences and how they can be changed.
- How to check
- Address objections on their merits and show what changed. The label “independent review” does not supply its participants, material or findings.
16
Use the published instruction
Full instruction · 1.4.2
Instruction checks: answers and limits
Copy the complete published instruction into a new conversation with the AI you use. Then add your task, source material and relevant context. You can request a translation, calculation, code, a direct answer or an examination of an added explanation.
This section contains the full 1.4.2 instruction. Historical 1.4.1 evaluations are preserved separately; their results do not describe the new version. Adding the instruction does not change model training or the provider’s product for everyone, and does not guarantee compliance with every requirement.
Full instruction 1.4.2
17
Continue with a particular example
Start with one response, requirement or decision. Show the original record, the added explanation and the proposed action. This plan can also be challenged, including its aims, criteria and proposed limits on requirements.
- HUMAN: learned descriptions of ourselves, others and the world.
- Self-description and Decisions Based on Records: assessments and conditions of participation.
- No Authority in the Output: grounds for moving from an answer to an action.
Continue from a concrete case
Instructions, connections and further material
Links to services and documentation, current 1.4.2 instruction files, and separately preserved historical reports.
- ChatGPT
- Claude
- Gemini
- Grok
- DeepSeek
- Qwen
- Mistral Le Chat
- Perplexity
- Microsoft Copilot
- Download the full instruction
- Download the instruction of up to 5,000 characters
- Download the short instruction
- ChatGPT settings
- Connecting to Codex
- Projects in Claude
- Connecting to Claude
- Where Claude connections are supported
- Gemini help: settings
- Gemini help: connections
- About Grok
- Connecting DeepSeek to an app
- Qwen chat
- Connecting Qwen to an app
- Instructions in Mistral Vibe
- Skills in Mistral Vibe
- Projects in Perplexity
- Data settings in Copilot
- Memory settings in Microsoft 365 Copilot
- Response settings in Microsoft 365 Copilot
- Download for Codex
- Download for Claude
- Download the Beforeword skill
- Download connection materials
- How Beforeword is built
- An earlier comparison of instructions
- Details of the download files
- Data for checking downloaded files
- Previous review: tasks, responses and limits
- Connection guide
- Changelog
- Report a problem
