The objection to AI in forensic work is usually stated as a question about the tool. It is really a question about the record.
An engineer who cannot say which parts of an opinion were formed by a person has a problem that no model quality fixes, and one that no disclaimer cures. An engineer who can say it, and can show the record, is in a stronger position than a colleague working entirely by hand who kept no record of anything.
These are different acts and they carry different exposure. Drafting is turning established findings into prose. Determining is deciding what the evidence means. A tool that drafts is a word processor with better manners. A tool that determines has formed part of an opinion that a licensed engineer will seal.
Most practices have never written that line down, which means every engineer draws it privately and differently, and nobody could state the firm's position if asked under oath. Writing it down is most of the work.
A causation opinion is strong in proportion to what it rules out. Naming a mechanism is the easy half. The half that survives challenge is the record of the candidates that were considered and eliminated, each with the evidence that eliminated it.
This was already the right method before any of this. The problem was cost: running an exhaustive elimination on every file takes more hours than the fee supports, so it gets compressed into narrative and the record of what was ruled out is never written. That compression is where opinions become vulnerable.
Machine assistance changes the economics of that specific step. Exhaustive elimination becomes affordable to run every time, and what was considered gets written down rather than remembered.
Framed this way, "so a machine formed your opinion?" has an answer. Framed the other way, it does not.
A usable record answers four questions about any determination, months later, without relying on anyone's memory:
A log that can be edited after the fact by the system that wrote it is not evidence of anything. The record has to be append-only and tamper-evident, and the reasoning system must not be able to rewrite the rules it is judged against. That constraint is not a nicety. A system that can quietly change its own standard cannot be audited, and an opinion that cannot be audited is an opinion waiting to be excluded.
Assume opposing counsel knows AI was involved, because in time they will ask as a matter of routine. The attack is not "you used a computer." The attacks that land are narrower:
Each of those is answerable with a record and unanswerable without one. Firms that write the rules first answer them in a sentence. Firms that adopted quietly answer them badly, in a deposition, with the file open.
Quiet adoption without a record is the actual danger, and it is already happening in most practices. The taboo exists because nobody has written the rules, not because the rules are impossible to write. A firm that has a policy, applies it uniformly, and keeps the record is in a defensible position today.
The AI Readiness Review is one hour with a firm's engineering staff and a written policy at the end of it.