A client asks, a year on, where a claim in your report came from. Or a colleague notices a text comes out differently from last time. Or a regulator wants to know who approved what. In all three cases the question is the same: what happened back then?
If you work with AI and record nothing, you can't answer it. The answer was in a chat window that has been closed, in a version of a model that no longer exists, and in the head of someone who has since left.
Why it doesn't happen by itself
There are two causes, and they reinforce each other.
The first is that recording feels like admin. It takes a few minutes per step, gives you nothing at the time, and in a busy week it's the first thing to go. A trail someone has to keep by hand only exists in the weeks when nothing special is going on.
The second is that you can't repeat the step. That sounds like a detail, but it's the heart of it. In 2023, Chen, Zaharia and Zou gave the same tasks to two versions of GPT-4, the March and the June 2023 versions. On telling prime numbers from composite ones, accuracy fell from 84 to 51 percent. Same task, same service, not a word of the question changed. The researchers conclude that the behaviour of such services can shift substantially within a short time, and that continuous monitoring is therefore needed.
For knowledge work that means the same request can give a different answer six months later. If you don't know which system, in which version and when something was made, you can't reproduce it, and so you can't explain it either.
Which layer this is
This is the work of Trail, the first layer of TRACES.
TRACES are the six layers on which you assess an AI system: Trail, Retrieval, Approval, Constraints, Expertise, Storage. Where a layer is missing, there is a risk.
Trail is about what's recorded: the source, the claim, who approved it, when. And about one requirement that settles the rest: the trail comes into being as a by-product of the work, not as admin someone has to keep up. A trail that builds itself is there in the busy week too.
This fits the A in SEAL. An approval belongs to the passage it refers to; if the passage is rewritten, the approval lapses. That only works if it's on record which passage was approved by whom, and when.
Five fields
You don't need to record much, as long as it's the right things. For each step with AI there are five fields.
- System. The name of the system, the model version if it's known, and the date. The date counts even when the version can't be seen: at least you know when.
- Request. What was asked, and on the basis of what material? That's a reference to the document used, or the question itself, and not a copy of data that didn't need to go in.
- Source. What does the claim that came out rest on? The passage, not just the title of the document. This is the field where Source and Edit meet.
- Change. What did a person change in the result? A text that has been rewritten after the model may no longer be covered by its source.
- Sign-off. Who checked it, with what expertise, against which passage, and when?
More fields are rarely needed. Fewer is possible, for the lightest steps. For a step in the "allowed" bin (see the article on sorting a workflow), system and request are enough. For "allowed with a check", you need all five.
An example of a trail
A made-up example, one line from a newsroom's log:
- System: language model from vendor X, version "2.1", 12 October
- Request: summary of the report "Quarterly figures, care sector", pp. 14 to 18
- Source: report p. 14, table 2; p. 17, paragraph 3
- Change: second sentence rewritten by the editor
- Sign-off: M. de Boer, 13 October, scope, certainty and time checked
A line like that reads in ten seconds. And it answers the three questions from the start: what happened, what it rested on, and who checked it.
What doesn't belong in it
A trail isn't an archive of everything put into the system. Personal data, client documents and other sensitive input don't get stored again, in a log more people can reach. Refer to the document it came from and leave it there. Rules on what may and may not go into a model belong in the work policy.
How the trail builds itself
The five fields only help if filling them in isn't an extra task. Three ways to get there, from simple to involved:
- A template in the document itself. A box at the bottom of every text with the five fields, or a comment on the passage. Whoever opens the document sees the trail.
- A fixed place for the result. The output of a step with AI always goes in the same folder or system, with the fields alongside.
- A system that does it for you. When choosing an AI system, this is the first question for the vendor: what does it record for every answer and every change, and can we export that? A system where that happens by default takes the work off a person. The vendor scan on this site takes a system through the six layers of TRACES, Trail among them.
What you can do this week
- Take a step with AI that was used last month, and try to reconstruct the trail: which system, which version, which source, who checked?
- Note which fields you could no longer fill in. Those are the fields that are missing today.
- Make a template with the five fields and put it in the document or folder where the work already lives.
- Agree that a step in the "allowed with a check" bin doesn't go ahead without the fields filled in.
- In a month, see how many lines are complete. That number is your Trail.
Source
Chen, L., Zaharia, M. & Zou, J. (2023). How is ChatGPT's behavior changing over time? arXiv:2307.09009. arxiv.org/abs/2307.09009