In the 1970s, an ecologist observed that an animal foraging for food would move from one patch to another without taking all the berries, nuts, or grass in the previous patch. He determined the reason was the value of return from the first patch diminished, and the effort to move to another patch without finishing the first yielded greater value. This is the "low hanging fruit" analogy. Eric Charnov published a paper on this topic, "Optimal foraging, the marginal value theorem," in the journal Theoretical Population Biology in 1976.
When I ran incident response teams years ago, we had a point while someone was doing data collection or an investigation that we "pulled them off" the task to move on, because we knew through experience that they weren't going to find more meaningful data. We don't have this type of barometer for "foraging" for information from AI.
This applies to humans in lots of ways: within AI, the food patch is a research thread, the switching to another patch cost is the cognitive context-switch to a new query, and the depletion curve is the diminishing quality of what AI returns after the first few queries. A universal video game example is as you are farming for a resource, you will move through the space to collect items quickly (e.g., in Super Mario jumping for coins), then move on to the next room instead of taking the time to get all the coins.
In organizations, data management takes effort of cleaning data before processing; the first few cycles achieve great results but diminish over time. Or as in code debugging with AI, the first few passes find lots of things to fix, but then the yields become fewer and less impactful.
In 1999, two researchers took the marginal value theorem (MVT) concept and related it to humans gathering data. The core idea is people will use cues about the information (they called its "scent") from things like search results headers to determine expected gains of finding quality information, but will stop or switch strategies as cost raises or quality of return falls. This is called Information Foraging Theory (IFT) and was developed by Peter Pirolli and Stuart K. Card.
The opposite of MVT is Sunk Cost Fallacy, where there is cognitive bias for people to "over-graze" on a task because they choose not to switch to another method, platform, or widget. This is a fallacy because even with obvious benefits of switching, the amount of money or time they have sunk into the first option is perceived not to be worth moving on or starting over.
It is important to develop standards, personally or within an organization, for people to know when to seek assistance and when to move on. Otherwise, you will waste time and effort (even AI tokens) on tasks that are not yielding value.
There's an optimal stopping point when foraging for food or information. Most people overcorrect in both directions: under-delegate by spending too much time doing personal analysis that develops confirmation bias, or over-delegate and become too reliant on AI that forfeits their personal judgement or creates hallucinations.
One problem is the cost of switching between patches is not symmetric. Working with AI is nearly free; doing it yourself costs more (in time and effort). IFT theory helps make it easier to look up things automatically rather than reviewing separate physical books.
The root problem is the foraging quality signal is not obvious with AI. Animals know there are berries in the patch because they can see them. AI will confidently give you answers with diminishing value (or outright hallucinations), and you won't realize it. AI will continue to answer confidently, making the patch appear full when marginal value has decreased.
How do you accommodate this broken quality signal with AI? AI mimics patch fullness regardless of actual yield (of quality information), which is why AI-assisted knowledge work may not succeed in practice. With Google searches, we know the later pages are less valuable, so we don't waste time checking every link in the 12 pages of results. Humans need to externally impose signals of the quality depletion that AI doesn't reveal on its own.
A final example from my security consulting days: my ethical hacking team usually had a full week to test applications for customers. One customer who we did dozens of tests for over the years asked us to just do a three-day "quick check" and give a list of significant findings instead of a full report. She wanted to reduce testing costs, and she recognized that we found most significant findings within the first couple days. My hackers hated it, because they knew they could find more vulnerabilities if they had more time; but the customer was leveraging MVT to recognize that the value she needed was if the application was insecure or not—not to find everything wrong.
This is where AI governance needs to establish the mechanism that identifies the information value signal that AI obscures, and MVT gives you a framework to describe it. We have an opportunity to design solutions like loop or turn limiters that cap how many AI exchanges are permitted before requiring human review, validation checkpoints, or token cost thresholds to give us that signal.
