Why the AI Treats the Latest Memo as the Truth

When you give a chat-based AI several documents that disagree, it often doesn't tell you they disagree. In practice, these tools tend to treat a later document as a correction or update of the earlier ones. Then they give you one clean answer.

That habit makes sense for a news story or a software manual, where the latest version usually is the right one. A matter file works differently. A memo written in June does not outrank a contract signed in January just because it came later.

Your prompt can make this worse. Requests like these push the tool toward recency:

• "What is the current delivery date?"
• "Give me the latest position on payment terms."

The tool hears "use the newest thing you can find." The result is that a recent memo's version of the facts can replace the signed contract's version in a summary, and the reader gets no notice.

Evidentiary weight is how much a source deserves to be relied on, based on what kind of document it is and how well it is supported. It is not the same as how recent the document is.

Here is a typical case. In a supply dispute, the supplier's operations manager testified at deposition that delivery was due on 1 March. Months later, someone wrote an internal memo summarizing the deal, and the memo says delivery was due on 15 March.

You ask the AI to build a chronology, which is a dated list of the key events in the matter. The entry reads:

15 March: Delivery due.

The chronology doesn't mention the deposition or 1 March, and it says nothing about a disagreement. Anyone reading it would assume the date is settled.

Notice what makes this hard to catch. The entry looks normal. It has a date, an event and a tidy format. It doesn't look wrong. You only find the problem if you already know about the deposition, or if you go back to the sources.

This is the pattern to watch for. One confident version of a fact, when the file holds more than one version, with no source named. If a chronology entry or summary sentence has no source reference attached, you can't tell whether the AI chose between sources.

Many documents in a matter file talk about other documents. An email summarizes what the contract requires. A memo recaps what a witness said. A case summary paraphrases a transcript.

Each of these is a secondary account: a later description of what a source says. The primary source is the thing itself, such as the signed agreement or the transcript of sworn testimony.

A secondary account does not change the primary source. If the signed contract says payment is due in 30 days and a later email says "the contract gives them 45 days," the contract still says 30. The email is wrong, or it refers to something else, or the writer misremembered. Either way, it doesn't rewrite the contract.

AI tools often miss this distinction. To the tool, both documents are just text that mentions a payment term, and the later one looks like an update. You need to know which kind of document you're looking at, because that decides whether a disagreement is a real change in the record or a description that doesn't match its source.

In the delivery example, the memo describing the deal is a secondary account. Its 15 March date doesn't replace what the deposition transcript records.

Some newer documents can change the record. The clearest example is an executed amendment, a later written change to a contract that the parties have signed.

Suppose the supply agreement set delivery for 1 March, and both parties later signed an amendment moving it to 15 March. This is not just a memo describing the deal. It may actually change the deal.

Even then, whether the amendment works is not automatic. Whether it is effective can depend on:

• the contract's own modification terms, such as a clause requiring changes to be in writing and signed by both parties
• the governing law, meaning the law that applies to the contract

Those are legal questions, so they go to the attorney to confirm. The AI should not decide them.

For the AI, the right move is the same either way. It should show both the original term and the amendment, note that the amendment may change the term, and flag that the attorney needs to confirm effectiveness. It should not quietly drop the original date.

So the question to ask about any newer document is whether it is the kind of document that could change the record, or only a comment on it.

An allegation is a claim one party makes that has not been established. Allegations show up in demand letters, complaints, pleadings and correspondence between counsel.

They also get repeated. A demand letter says the supplier "knowingly shipped defective goods." The complaint says it again. A later letter from opposing counsel says it a third time. By the end of the file, the claim appears in several recent documents.

An AI tool can mistake that repetition for confirmation. It sees the claim stated plainly, several times, in the newest documents, and it may write in a summary:

The supplier knowingly shipped defective goods.

That sentence turns an allegation into an established fact. Neither repetition nor recency makes a claim true. Three documents from the same party saying the same thing are still one party's position.

The safer wording keeps the claim attached to who made it and where:

The buyer alleges in its demand letter and complaint that the supplier knowingly shipped defective goods.

This matters well beyond chronologies. An allegation that slips into a summary as fact can travel into a draft, where it is much harder to spot.

Every time the AI picks one version of a fact over another, it makes a choice. The problem is not that a choice gets made. Sorting out conflicting evidence is part of legal work. The problem is a choice made quietly, where nobody can see it.

Compare two outputs:

• A chronology that shows 1 March and 15 March side by side, each with its source. The attorney sees the conflict and decides what to do.
• A chronology that says 15 March. The attorney never learns there was a conflict, so they never weigh it.

The first output may look messier, but it is safer. A conflict that is shown can be resolved properly. A conflict that is hidden carries forward into the summary, the research notes and the draft, and it may only come up when the other side points to the deposition.

Deciding which evidence wins is the attorney's call, not the AI's. The AI's job is to put both versions in front of the attorney, each with its source and its weight. That is the goal of this course: make every such choice visible.

In the next lesson, you'll set out how evidentiary weight is judged in your matter so the AI has a standard to apply.