My illustration entitled: “The Open Visor” – A colossal mechanical face opens like a visor, revealing a human standing behind it, directing its address to a vast audience.
Imagine asking an artificial intelligence system to prepare a reply to a supplier. You want clearer language, a more persuasive argument and a professional tone. The system produces a polished message.
In the middle of that message is a sentence: “We accept your price and can guarantee delivery by Friday.”
The wording is fluent. The promise may be one you never intended to make.
This is where assistance with expression can become an exercise of authority. A system that helps someone find words is doing something different from a system that decides what that person will promise, disclose, accept or endorse.
For Unceded Intelligence, the distinction is fundamental. Intelligence should expand human capability while leaving legitimate human authority unceded. When a machine speaks on someone’s behalf, its ability to sound convincing must remain subordinate to the mandate it has actually received.
A system’s ability to speak in my voice does not give it authority to decide my position.
Four functions that should remain distinguishable
AI assistance with communication can involve at least four different functions: drafting, recommending, communicating and committing.
Drafting produces possible words. The person remains free to reject, revise or ignore them.
Recommending proposes a position or course of action. The system may suggest accepting an offer, requesting a concession or choosing a particular response.
Communicating transmits something to another party. Information leaves the private space in which it could still be reconsidered.
Committing expresses an undertaking, acceptance, assurance or other position on which another person may rely.
These functions can overlap. Sending a message can disclose private information, create expectations and make a commitment in a single action. They should not be treated as a simple sequence in which permission for an earlier stage automatically authorizes everything that follows.
A person can authorize drafting without authorizing transmission. They can authorize transmission of a question without authorizing acceptance of the answer. They can permit a negotiation within agreed limits without permitting the system to conclude it.
The design of the service should preserve these distinctions as clearly as its language describes them.
A draft is a proposal to the person
A draft belongs first in a space where its intended author can consider it.
The fact that a system has generated a sentence does not mean that the person has adopted it. A suggested apology is not yet the person’s apology. A proposed endorsement is not yet their endorsement. A possible negotiating position is not yet their offer.
This matters because polished language can make unsettled decisions appear settled. A paragraph may combine facts supplied by the person with assumptions introduced by the system, giving both the same confident tone.
A useful drafting assistant should distinguish what it has been told from what it has inferred or proposed. If it inserts a delivery date, a concession or a claim about the person’s experience, that addition should be visible for review.
It should also preserve uncertainty. “I will consider this” and “I agree to this” perform different acts. Replacing the former with the latter is a change in substance, even if the revision sounds more decisive.
Helping someone express an intention should not quietly manufacture the intention being expressed.
Recommendations can choose a position before anyone notices
The authority question begins before a message is sent.
A system asked to recommend a reply may frame one option as obvious, hide an important trade-off or optimise for a goal the person has not endorsed. It can thereby influence the position that later appears to be the user’s own considered choice.
Imagine a system helping someone respond to a complaint. It might recommend the answer most likely to end the exchange quickly. The person may instead value repairing the relationship, acknowledging a mistake or understanding what happened.
The fastest resolution and the most appropriate response need not be the same.
A recommendation should therefore make its important assumptions visible. Where purposes conflict, the person should be able to see the alternatives and decide which consideration matters.
Material conflicts of interest also deserve disclosure. A service recommending a provider from which it receives a benefit should not present that recommendation as though the user’s interests were its only consideration.
Unceded Intelligence asks that advice strengthen the person’s judgment rather than replace it through an arrangement the person cannot meaningfully inspect.
Transmission is itself a consequential action
It is tempting to treat sending as a minor final step after the difficult work of composing. But transmission changes who can see, interpret, retain and act upon the message.
A draft can remain private. A sent message enters a relationship.
Even a message that makes no promise can reveal a negotiation strategy, expose personal information or change another person’s understanding of the sender. The recipient, timing, attachments and channel can matter as much as the words.
Permission to prepare a response should therefore be separate from permission to deliver it. Approval to send one version should not silently cover a materially revised version, an additional attachment or a wider audience.
This principle also applies to translation and summarisation. A person approving a message in one language needs a reasonable basis for understanding its meaning in another. A summary used for approval should preserve the qualifications that matter.
The content authorized should remain the content transmitted, subject only to changes that genuinely leave its meaning and relevant consequences intact.
Commitment requires authority over the substance
Some statements do more than convey information. They undertake to do something, accept a condition or invite another person to rely on a position.
“Your proposal has been received” differs from “Your proposal is accepted.” “Tuesday may be possible” differs from “Tuesday is confirmed.” “I can ask about a refund” differs from “A refund will be provided.”
These differences should not disappear under a general instruction to handle correspondence.
A system that makes a commitment needs authority covering the substance of that commitment. Permission to negotiate a price does not necessarily include permission to accept one. Permission to arrange a meeting does not include permission to promise attendance by people whose schedules the user does not control.
Nor does the system’s prediction that a person would probably agree supply the missing mandate.
The standard should be especially clear where a statement concerns money, confidentiality, public endorsement, deadlines or obligations involving other people.
The question is not simply whether the sentence is well written. It is whether the person authorized what the sentence does.
My illustration “The Open Visor” work-in-progress. The art represents that the representative’s impressive presence never obscures whose voice it carries.
Negotiation needs boundaries, including a stopping point
Negotiation makes the distinction between capability and authority particularly visible.
A capable representative might compare offers, identify concessions, ask questions and propose alternatives. Its usefulness does not establish how much of that activity it has permission to undertake.
Suppose a small business authorizes a system to seek a lower price from a supplier. The mandate might permit questions about volume discounts and delivery options while reserving acceptance for the owner.
Alternatively, the owner might authorize acceptance within a specified price range, provided that quality, delivery and payment conditions remain unchanged.
Both arrangements can preserve human authority. The difference is what has been deliberately delegated.
A price limit alone may be insufficient. A cheaper offer could come with an unacceptable deadline, an exclusivity condition or permission to use the customer’s name publicly. Those terms change the substance of the bargain.
A negotiation mandate should therefore identify permitted concessions, protected conditions, information that may be disclosed, duration and the circumstances requiring further approval.
When an offer falls outside those limits, the system should return it for consideration. It should not invent a broader mandate to avoid losing an opportunity.
The recipient needs to understand the representative’s role
Representation involves someone on the other side of the conversation.
That person may need to know whether they are receiving a tentative proposal, a message approved by the sender or a response generated under a standing instruction.
The appropriate disclosure depends on what matters to the exchange. Assistance with spelling does not create the same issue as an automated negotiation conducted without immediate human review.
Where identity, personal participation or authority matters, the representative’s role and limits should be clear. A system should not imply that a person has personally checked a fact, experienced an event or considered a proposal when that has not happened.
Nor should it imitate intimacy to manufacture the impression of attention or agreement that the person never supplied.
A recipient should have an appropriate route to seek confirmation, challenge an apparent commitment or reach someone with authority to resolve uncertainty.
Calling a message “automated” is insufficient if the surrounding behaviour still invites reliance beyond the system’s mandate. Disclosure and conduct should tell the same story.
Human approval must concern the actual decision
Requiring approval before every message can look like a complete solution. It is not enough if the approval process obscures what matters.
Imagine a long message presented with a single button asking whether it “looks good.” Buried inside are a new assurance, a changed price and an additional recipient.
The person may approve the tone while overlooking the commitment.
Meaningful review should make consequential elements easy to identify: what is being promised, which terms are accepted, what information is disclosed and who receives it.
Changes introduced after review should trigger further attention when they alter those elements. Approval should attach to a sufficiently clear action, rather than to an open-ended process that can keep revising itself.
Standing authority remains possible. A person may allow routine acknowledgements or scheduling messages within defined conditions. The system need not interrupt for every ordinary step.
The aim is to concentrate human judgment where the communication could exceed the established mandate.
Revocation and accountability must survive delegation
A representative’s authority should be revisable.
A person should be able to pause outgoing communication, narrow the permitted subject matter or end a negotiation mandate. Queued messages and permissions passed to connected services should be addressed when the mandate changes.
Where another party is relying on continuing representation, an appropriate notice may be needed so that the change is understood on both sides.
Withdrawal does not make messages already received disappear. Nor should it be used to deny commitments a person knowingly authorized merely because the outcome later became inconvenient.
At the same time, a person should have a meaningful way to challenge actions taken outside the mandate. The existence of an account or a general permission to assist should not settle every dispute about what was authorized.
Accessible records can help establish what instruction applied, what was approved and what was sent. Those records should be limited and protected; accountability does not justify collecting every private detail indefinitely.
The user, service provider and deploying institution may each bear responsibilities according to the powers they exercised and the failures involved. Responsibility should be examined rather than displaced into the phrase “the AI said it.”
Assisted expression can remain human expression
None of this requires people to write every word unaided.
A person may have a clear intention but struggle to express it. They may need translation, help organising an argument or a more accessible way to participate in a conversation. Assistance can enlarge their voice and their practical independence.
The relevant question is whether the resulting communication remains connected to an intention and mandate the person can understand, endorse and revise.
Someone can adopt words they did not personally compose. What matters is that adoption is real, and that authority over consequential representation remains where it legitimately belongs.
I want intelligent systems that help people enter conversations they might otherwise be excluded from, examine options they might otherwise miss and express positions they might otherwise struggle to communicate.
That ambition becomes stronger when the boundary between expression and commitment is respected.
An AI may help me find my words, examine my options and carry my message. Its ability to represent me does not give it an unrestricted right to commit me.