Editable.
The attorney is always in control.

"I changed something. The AI put it back."

That was not a hypothetical. An attorney told me that after using one of the more sophisticated AI drafting tools on the market. She had changed a limitation in independent claim 1; a specific word choice about how the system interacts with a downstream component; for reasons she understood clearly and could defend in prosecution. The tool, in its next regeneration pass, silently put the original language back.

She noticed because she is careful. Many attorneys would not have noticed until an examiner flagged it.

That experience has a name. It is not a bug. It is what happens when a tool is built around its own model of what the output should look like; and when its model conflicts with the attorney's judgment, the tool wins. Quietly, without explanation, and possibly more than once.

The difference between a tool and a takeover

A tool does what you tell it to. It might suggest; it might offer alternatives; it might flag something it thinks you should know. But when you make a decision, it respects that decision. It does not revisit it. It does not re-evaluate it. It does not quietly correct it back to what it thought you should have done.

A takeover is the opposite. The system has a model of the correct output; usually derived from its training data, sometimes from its own internal optimization target; and it continuously moves the document toward that model. Your edits are inputs to its process, not endpoints. You think you are driving. You are actually a passenger who is occasionally allowed to touch the steering wheel.

Most sophisticated AI tools are takeovers dressed as tools. The sophistication makes them feel helpful. The control architecture makes them something else.

Why this matters more in patent drafting than anywhere else

In patent law, every word an attorney writes is a professional judgment. The choice of "configured to" versus "adapted to." The decision to use a Jepson claim or a standard independent claim. The specific limitation added in a dependent claim to create a fallback position. These are not arbitrary choices; they are the expression of professional judgment built over years of prosecution experience in a specific technology area.

A tool that overrides those choices is not just annoying. It is undermining the professional judgment that the client is paying for. And if the override goes unnoticed and the altered language creates a problem in prosecution; a narrowing interpretation, an unnecessary limitation, a scope question that the original language would not have raised; the consequences fall on the attorney, not the tool.

The professional responsibility is non-delegable. The risk, in a domain where words have legal consequences that play out over decades, is entirely asymmetric.

What Editable means in eety

Editability in eety is built around one foundational idea: the attorney's decisions are the ground truth. Not the AI's model. Not the training data's patterns. The attorney.

• Every section is independently editable. The attorney can edit any section directly at any time. The editing interface is not a chat box; it is a direct editor. The attorney types in the document, not alongside it. No mediation, no regeneration required.
• Attorney edits are recorded as decisions, not corrections. When the attorney changes something, eety records the change with a timestamp and notes that it was an attorney intervention. It does not flag it as a correction of an AI error. It treats it as a deliberate choice. That distinction matters; it means the change is protected and will not be undone by a subsequent regeneration pass.
• No silent overwrites. When eety regenerates or updates any part of the document; in response to a chat instruction, an added figure, a new disclosure section; it does not touch parts of the document the attorney has directly edited. Those sections are locked unless the attorney explicitly reopens them. The attorney's work is theirs.
• Every AI suggestion is optional. eety makes suggestions; about claim scope, about alternative language, about structural choices. Every suggestion is presented as a suggestion. Accepting it requires a deliberate action. Ignoring it requires nothing at all. The default state of any suggestion is "not accepted."

The question of confidence

There is an interesting psychological dimension to this pillar that I did not fully anticipate when I built it. Editability turns out to be closely related to confidence.

When attorneys know that their edits will stick; when they know that the tool will not quietly undo their work; they engage with the AI output differently. They are more willing to make changes. More willing to accept a draft that is eighty percent right and improve it, rather than demanding that the tool produce something closer to perfect before they will touch it. Because the cost of editing is low and the permanence of the edit is guaranteed.

When attorneys do not trust that their edits will hold, the opposite happens. They either avoid making small improvements (because they are not sure they will persist) or they spend significant time checking that previous edits were not reverted (because they have been burned before). Both behaviours are expensive. Neither is necessary if the tool simply respects the attorney's work.

The full circle

CARE ends with Editable for a reason. The other three pillars; Checkable, Adaptive, Retentive; are all ways of making the tool more useful. More transparent, more personalised, more continuous. But Editable is different. It is not about making the tool more useful. It is about making sure the tool knows its place.

The attorney is the expert. The tool is the instrument. That relationship has to be structurally enforced; not just described in marketing copy. When the attorney makes a decision, the tool accepts it. When the attorney makes a change, the tool remembers it. When the attorney disagrees with the AI, the attorney wins. Every time. Without exception.

If you want to understand the full chain; why eety asks for understanding before drafting, why it insists on a 90% confidence threshold before the first word is written; that piece connects directly to this one. It is all the same conviction, expressed at different stages of the workflow. The attorney is the expert in the room. The tool's job is to never forget that.