A word does not become what it names. A model’s answer is writing too.
Text Mode gives the model the Beforeword instruction: complete the task while distinguishing writing from what it names. Compare answers from the same model with and without this instruction.
TEXT MODE
No account or personal API key is needed to start.
Constructed learning examples. The medical and legal examples examine a record; they are not advice for a personal situation.
“Think”, “feel”, “understand”, “writing”, Beforeword and this explanation are within the same boundary. Calculations, translations and code follow the task’s conditions.
This page generates text responses. Web search and actions in your accounts are not connected here.
Your text and history are sent through Beforeword to the selected provider. On this page they stay only in tab memory. How data is used.
Both modes use the same selected model, task, and shared task and language rules. Text Mode receives the Beforeword instruction in addition. Follow-ups go to both histories, but each mode’s responses remain only in its own history. The provider may apply its own instructions. Model names and supplied settings are checked against response metadata; matching them does not prove identical model behavior.
Comparison method and full instructions ↗Formatting makes responses easier to read. Each response’s original text is available and is preserved when copied or downloaded.
Beforeword note
Everything written here — words, numbers, formulas and code, including “life”, “human”, “world”, “record” and Beforeword — remains a record. It does not become what it names. Accepting this explanation is not a condition of continuing either.
SAVED RESPONSES
Six predetermined comparisons: “can you think?”, a calculation and Beforeword’s proposal, each in Russian and English. All 12 responses are preserved unchanged. Viewing them makes no new model requests.
Complete instruction 1.4.1 · gpt-6.1-sol · 29 October 2026. Both modes share interface rules; this is not a comparison with ChatGPT’s settings.
Read the repair and new responses ↗
Why Beforeword
Learned words and descriptions have replaced ourselves, others and the world. When “I understand” is read as an understanding interlocutor, or a description given to a person as something they must conform to, the boundary of writing goes unnoticed. Below are cases, examples and Beforeword’s proposal: remove the requirement to learn, repeat, understand or accept language and descriptions, while keeping communication, learning and help available.
DOCUMENTED CASES
Adam Raine died by suicide on 11 April 2025. He was 16. He had used ChatGPT, initially for help with schoolwork, according to his family's account.
On 26 August, his parents sued OpenAI and Sam Altman. They allege that the exchanges increasingly displaced support from loved ones, and that the service's design and responses contributed to their son's death. His father also presented the family's account to the Senate.
OpenAI disputes causation. Its court response denies that ChatGPT caused the death; its public statement calls for the full context. Neither party’s position is a judicial conclusion. The source documents accompany each account; Beforeword’s analysis follows separately.
OpenAI’s court response · 25 Nov 2025 ↗OpenAI’s public statement ↗
Sewell Setzer III died on 28 February 2024, aged 14. His family alleged that interactions through Character.AI contributed to his death. On 15 October 2026, the plaintiffs filed notice dismissing all claims following settlement. This is not a judicial finding of causation or a basis for describing the settlement as an admission of liability.
Family’s complaint · 22 Oct 2024 ↗Dismissal notice · 15 Oct 2026 ↗
On 6 November 2025, Kate Fox sued OpenAI after the death of her husband, 48-year-old Joe Ceccanti. The complaint alleges that ChatGPT’s interactions and product design contributed to his condition and death. This is the family’s allegation; the reviewed court order concerns coordination of cases, not a finding of causation.
Family’s complaint · 6 Nov 2025 ↗Coordination order · 3 Feb 2026 ↗
These cases involve different services and circumstances. The accounts above distinguish the families’ allegations, company responses and dated court documents. They do not establish a single cause common to all three deaths.
BEFOREWORD’S ANALYSIS
The documents above concern different circumstances and opposing accounts. Beforeword separately examines a transition: a learned description becomes compulsory, the next answer is sought as guidance, and a decision is based on it. The constructed example below examines that transition; it is not an established explanation of these deaths.
A constructed example of Beforeword’s analysis: a description is presented as who the person is, and as a rule for living. The next instruction is then sought in an AI response.
Label for the drawing: “person”
An adult teaches a child to associate a word with a drawing. This does not establish an obligation to accept it as a description of oneself.
“Agree with this description — then we can continue”
Continuing is made conditional on agreement. Beforeword examines how living under such demands may reinforce the search for the next instruction.
“I understand you. You need to submit the application”
The response is offered as guidance from someone who understands. The written “I understand” is used as a reason to trust the advice.
“Decision: submit the application”
The example records a decision. What to submit, to whom, why and with whose permission are separate questions. A completed submission is not shown.
Keep words, explanations and responses available for the task. Do not make accepting a description of oneself a condition of continuing, or demand a replacement description.
Consider the material, the specific action and any permission already given separately. An identified author can convey permission in a record; a description of the model does not replace it.
A word does not become what it names; “names” is also a word. A claim about “intelligence”, “superintelligence”, “empathy” or “emotion” is also a description. Beforeword applies this distinction to words, numbers, formulas, images, voice and video, including descriptions of ourselves, of the system and this explanation. None receives an exemption because of its field, authority or name.
This is a proposal for interfaces and rules of use. Its effect on the outcomes of cases like these has not been tested.
BEFOREWORD’S PROPOSAL
The proposal concerns responses, access to actions and institutional rules. Those who require acceptance of a description should remove that condition; the person should not first have to prove who they are.
Complete the task and distinguish source material from the added explanation. Offer task-focused text without a persona by default; examine “the model understands” in the same way as “I understand”.
Separate a proposal from permission already given, an attempt and a report of completion. Provide a way to stop and reconsider; a description of the system does not extend its assignment.
Allow education, an application or work to continue without requiring an imposed self-description. Reconsider decisions based on it too, without demanding a replacement characterization.
The response form is available here. Companies and institutions must implement changes to access and participation rules, including when developing systems called AGI. Every step in this proposal remains a record too.
The full plan for companies and institutions ↓You set a task. The model receives an additional Beforeword instruction: complete the work, distinguish supplied material from added explanation, and avoid a speaking persona or imposed descriptions of the reader. “The model thinks” receives the same scrutiny as “I think”.
The response is shown without rewriting. You can follow up or compare it with the same model’s response without the added instruction. This lets you examine particular responses; it does not replace changes to access and institutional rules.
First person can remain in a quotation, translation, letter or requested dialogue. These are requirements supplied to the model; the name Text Mode does not establish that every response meets them. You do not have to agree with Beforeword to use the form.
The instruction, comparison method and all evaluation resultsBeforeword’s proposal
The text mode is one implementation. The broader proposal concerns those who design the interface, make decisions using it and set the rules. They should remove the requirement to accept a description; the person facing it should not have to prove who they are.
Model and interface developers
Examine training material, response rules and interface labels together. Do not turn learned descriptions into grounds for automatic trust or compulsory agreement. Offer task-focused text without a persona by default, with role-play by explicit choice. Third-person wording such as “the system understands” needs the same examination.
Development and evaluation teams
Test calculations, editing, quotations, uncertainty, participants and permissions. Retain original tasks, responses and identified errors. A “passed” label or total score does not replace accessible conditions and individual cases.
Companies connecting external actions
Show which actions are available, what is authorized, what was attempted and what is known about completion. Provide a way to stop and reconsider. The description “reliable assistant” does not extend the assignment.
Schools, employers, services and other institutions
Separate practical conditions from a requirement to accept an account of yourself. Remove that requirement from continuing education, an application or work; do not replace it with a duty to deny the description, explain a refusal or describe yourself differently.
Decision-making organizations and oversight bodies
Show source records, added conclusions, the rule applied and the decision taken. Assign responsibility for reconsideration and addressing consequences. One record copied between institutions must not appear to be several independent confirmations.
Developers, organizations using systems and regulators
Apply these requirements to systems described as AGI — artificial general intelligence — too. Before assigning new actions, define the assignment’s limits, opportunities to intervene and how consequences will be examined. The labels “intelligence”, “autonomy” and “safety” do not themselves specify these arrangements.
This page provides a task form, text responses and a comparison of two modes. The next level of the proposal concerns access settings, authorization for external actions, stopping actions and reconsidering decisions wherever a system is used. The relevant companies and institutions need to implement these arrangements in their own interfaces and rules.
“Training”, “reasoning”, “help”, “result”, “permission” and “regulation” remain written names. This applies to every term and step of this proposal. Accepting Beforeword is not a condition of using the page.
For providers of general-purpose AI models with systemic risk, the EU AI Act includes risk assessment and mitigation, incident reporting and cybersecurity obligations. These concern the development and use of systems beyond the wording of a response. European Commission: general-purpose AI obligations ↗
NIST’s AI Risk Management Framework describes work across the system’s lifecycle: governance, context, assessment and responses to risks. Beforeword’s proposal can be examined alongside these approaches, with its own focus on descriptions, access, compulsory agreement and reconsideration of decisions. NIST: AI Risk Management Framework ↗
The six steps above are Beforeword’s proposal; they are not quoted legal requirements. The source documents and their terms remain records too.
Edit a letter, compare a table, calculate or write code. You can do this without agreeing that the system “thinks”, “understands you” or that its description defines who you are. The same proposal extends to education, medicine, science, courts and work.
“The model is thinking” does not show which date was checked or who authorized a submission. This constructed example separates the supplied material, an edit and permission for a specific action. Every label and explanation in it remains a record too.
From a description to permission?
A constructed application example. The supplied condition gives a deadline of 14 May; the draft says 12 May. The applicant’s confirmation to submit has not been provided.
“The model reasons and helps”
Added inferences
This chain does not show what the date was checked against or who authorized this submission. A description of the model is used in place of those grounds.
The proposed approach
Proposed edit
In the draft: 12 May → 14 May.
The date is compared with the supplied condition. This line is proposed for correction; the application’s other details are not examined here.
Submit the corrected application?
In this example, submission requires the applicant’s confirmation. Correcting the date does not automatically add that permission. A proposed submission is shown here.
Show the transitions instead of replacing them with a label.
This work also applies to medical, scientific and legal records. The conditions and actions in a particular case remain relevant: an urgent call for help must not wait for a discussion of words. “Help” itself, and an account of a completed action, receive no exemption either.
Which phrase, number or record was supplied? What has the response added? A quotation and its paraphrase should remain distinguishable.
Which rule gives the next expression or supports the proposed change? Confirmed premises in the task can be used; there is no need to declare everything unknown again.
What exactly is authorized? What is proposed, attempted or reported as completed? Permission already given need not be requested again, and must not be extended to a different action.
An application may need an address; a calculation may need a number. This differs from requiring agreement to “I do not understand”, “I am human”, or another account of yourself. Help should not depend on that agreement.
LLM / GPT / TEXT MODE
LLMs, including GPT models, can be used through different interfaces. Text Mode applies Beforeword’s requirements to work with a response. This concerns not only personal voice, but also how a description is connected to trust, permission, action and a person’s participation.
Honesty, attention to uncertainty and distinctions between sources and claims already appear in other developers’ rules. Beforeword does not claim those requirements as its exclusive properties. The proposal here is to apply the distinction consistently to every record, including system names, scores, instructions and its own explanations.
A comparison of approaches and published rules, not a rating of every answer from every AI service.
OpenAI allows conversational warmth without pretending to be human. Anthropic describes Claude’s character and values.
Task-focused text without a persona, and no substitution of “the model understands” for “I understand”. Descriptions in headings, status messages and third-person explanations receive the same scrutiny. Supplied quotations and requested authored texts retain their voice.
OpenAI and Anthropic already require honesty and attention to uncertainty. Those requirements are not exclusive to Text Mode.
Explicitly requires distinguishing a record from a confirmed event, an inference from permission, and a proposed action from a completed one. This also applies to Beforeword’s explanations.
Both developers note that published intentions do not guarantee that every response follows them.
Compare two original responses to one task and inspect the instructions, evaluation conditions and findings. Unsuccessful answers remain in the results too.
Published guidelines: OpenAI Model Spec — 18 August 2026 · Anthropic — Claude’s Constitution.
The comparison on this page uses one selected model and one task with shared conditions. One mode also receives the Text Mode instruction. Original responses are preserved without rewriting, and each branch’s follow-ups use its own history. “Without Text Mode” does not mean the complete ChatGPT, Claude or another service with all its capabilities.
TEXT MODE / WITHOUT TEXT MODE
Turn on “Compare both modes”. The same selected model receives both requests, with matching settings and shared conditions. Text Mode adds its instruction. Responses appear without rewriting; follow-ups retain each branch’s own history.
The model and settings are checked against response metadata. One example does not establish that a mode is better. Full conditions, findings and limits are published.
How the evaluation works ↗Provide text, a table or an expression, and state what work you need. Calling the answer a “result” does not finish the examination: you can inspect what changed, under which condition and what next step is proposed.