Sometimes the breadcrumbs are the story

In Hansel and Gretel, the breadcrumbs are not background detail. Their disappearance changes the story.

On the children's first journey into the forest, Hansel leaves white pebbles and follows them home. On the second, he scatters breadcrumbs instead. Birds eat them, the route home disappears, and the story moves deeper into the forest (Brothers Grimm, Hansel and Gretel).

The breadcrumbs are not merely a record of how the children travelled. Their presence, loss and consequence are part of the story itself.

I keep returning to that distinction in my own work. The questions, evidence, assumptions and alternatives behind a conclusion can look like working material to remove once the answer is polished. Sometimes they are where the understanding lives.

I notice this when I return to some of my handwritten notes.

The uncertainty is still visible. A question sits beside a conclusion. An arrow points to an alternative. A sentence trails off because I had not yet worked out what I meant. The page holds what I understood at that moment, including the parts that were unfinished.

A digital document can be much cleaner and less revealing. I can revise the uncertainty away. With AI, I can turn a rough collection of observations into coherent language in minutes. That is genuinely useful. It also makes it easy to keep only the destination.

The problem is not that every abandoned sentence deserves to survive. Most working material is disposable. The problem is losing the part of the journey that explains why the conclusion made sense: the original question, the evidence available at the time, the important assumptions and the alternatives that were rejected.

When those connections disappear, the final artefact has to stand on its own. We begin judging whether it is clear, consistent and complete rather than whether it answers the problem we meant to solve.

Polish is not evidence

AI did not create the tendency to take ease for truth. In an experiment with 235 undergraduates, Rolf Reber and Norbert Schwarz showed participants short factual statements in colours that made them easier or harder to read against a white background. The more visible statements were judged as true more often (Reber and Schwarz, 1999).

That experiment concerned colour contrast and trivia, not AI-generated reasoning. It does not show that polished prose deceives us. It supports a narrower caution: qualities of the presentation can influence a truth judgement even though they add no evidence to the statement.

Explanations are not a complete safeguard either. In a prospective randomised study, 220 physicians interpreted chest radiographs with simulated AI advice. The form of the explanation affected diagnostic performance and trust. Physicians placed greater “simple trust” in local, feature-based explanations than in global, example-based explanations regardless of whether the advice was correct (Prinster and colleagues, 2024).

That is a specialised medical task with simulated advice, so it should not be generalised to ordinary writing or software work. What it usefully demonstrates is that the presence and form of an explanation can change reliance. An explanation may help someone reason; it is not evidence that the conclusion being explained is correct.

The risk I see in my own work is simpler. AI can make a conclusion fluent before I have made its foundations visible. The result looks ready for review, but the reviewer is being shown the answer without the material needed to challenge how it was reached.

Correct against itself

A requirement can be clear, testable and competently implemented. Tests can pass because they accurately reflect that requirement. None of this establishes that the requirement addresses the right problem or respects the wider constraints of the system.

The work can be correct against itself and still wrong for what it was meant to achieve.

This happens before the snowball effect I described in When AI velocity outruns feedback. That note is about what happens once a weak decision is reused as context. This concern is earlier: the reasoning can disappear before the decision begins to travel.

This is where the loss of the journey matters. The final decision becomes the only remaining frame of reference. A review can ask whether its parts agree with one another, but not whether the whole thing has drifted from the observation that began the work.

Keep enough of the journey

The answer is not to preserve every prompt, draft and conversation forever. Raw capture without synthesis becomes difficult to use, and indiscriminate retention creates a different kind of obscurity.

I want a proportionate connection between three things:

• the original observation or problem
• the reasoning that materially shaped the decision
• the decision that applies now

For routine and reversible work, that may be a sentence linking the decision to the problem and its decisive assumption. For consequential or hard-to-reverse work, it may also need the strongest rejected alternative, the unresolved uncertainty and the condition that would cause the decision to be reconsidered.

Those records do not have to live in one document. The current decision should tell us what applies now. The material behind it should let us understand why it became reasonable.

AI can help with this. It can identify assumptions, compare a proposal with its sources and produce a concise account of what changed. But the summary should remain connected to the material it summarises. Otherwise AI has not preserved the reasoning; it has produced another polished destination.

This level of traceability is not necessary everywhere. A low-risk copy edit or an easily reversed experiment should not carry the same record as a decision that changes a product rule, architecture or commitment. The amount of journey worth retaining should rise with the consequence of losing it.

A finished answer tells us where we arrived. Correctness asks whether we can still connect that place to the problem that set us moving.

Arrival is not evidence that we reached the right destination.

Further reading

These books offer related lenses rather than direct proof of the argument. They are included as a reading list to examine, not as endorsements of every claim they contain.

• The Reflective Practitioner, Donald A. Schön—how practitioners frame uncertain problems and develop knowledge through reflection in action
• The Design of Design, Frederick P. Brooks—design processes, constraints and the value of representing a design's trajectory and rationale