Additive Bias¶
Core Idea¶
When asked to improve a system, agents reliably reach first for adding a component and systematically underweight removing one — a direction-asymmetric search distribution that over-samples the "add" region of transformation-space even when subtraction is better.
How would you explain it like I'm…
Add-A-Block Habit
When something is wrong, most people try to ADD something to fix it instead of taking something away. If a tower wobbles, we want to glue on more blocks, when really we should pull out the crooked one. Taking things away is just as good a fix, but our brains forget to try it. Additive Bias is reaching to add a piece when removing one would work better.
Pile-It-On Bias
When people are asked to improve something, they almost always think of adding a part, a rule, a step, or a feature, and they almost never think of removing one. This happens even when removing would clearly make things better. It shows up everywhere: in companies, in laws, in machines, even in nature. Over a long time, things just keep piling up, because adding is the move we reach for and removing is the move we forget. Removals are rare and usually need a special push, like a cleanup day or a rule that automatically expires.
The Add-Don't-Remove Bias
Additive Bias is the pattern that when an agent is asked to improve a system, it reliably reaches first for adding a component and underweights the option of removing one. The asymmetry holds even when subtractive changes are clearly better, and it recurs at organizational, legal, technical, and biological scales, where additions vastly outnumber removals over a system's life. The structural claim is a direction-asymmetric search distribution: when generating candidate changes, agents draw far more from the add region than the remove region of the space of possible changes, even when the better solutions sit in the remove region. This is not a one-off mistake but a bias in the search itself, persisting because its causes recur: additions are visible while removals are not, adding earns clearer credit than removing, and it is harder to reason about an absence than a presence. The result is a steady accretion signature, where components pile up faster than they are removed.
Additive Bias is the pattern that when an agent, whether a person, team, organization, evolving lineage, or regulator, is asked to improve a system, the agent reliably reaches first for adding a component (a feature, rule, step, piece, person, or layer) and reliably underweights the alternative of removing an existing one. The asymmetry holds even when subtractive transformations are demonstrably better, and it recurs across organizational, legal, technical, and biological scales, where the ratio of additions to removals over a system's lifespan is strikingly skewed regardless of which would optimize the stated objective. The structural commitment is a direction-asymmetric search distribution: when generating candidate transformations of an existing system, agents draw far more heavily from the add-a-component region than from the remove-a-component region of transformation-space, even when the latter holds better solutions. This is not a one-off error but a bias in the search distribution itself, persisting across substrates because its generative factors recur: the visibility of additions versus the invisibility of removals, the credit-assignment asymmetry between adding and removing, and the cognitive difficulty of reasoning about counterfactual absence versus counterfactual presence. The mechanism produces a characteristic accretion signature in long-lived systems: components accumulate faster than they are removed, the system grows in size and complexity along a trajectory tied more to iteration count than to purpose, and removals, when they happen, are infrequent, deliberate, and require special machinery such as sunset clauses, refactoring sprints, or deliberate purges.
Broad Use¶
- Cognitive problem-solving: across building, grid, and itinerary tasks, subjects offered both moves overlook the better subtractive one even when explicitly permitted.
- Software engineering: codebases accrete features and dependencies because deletion is rare; removing what additions did not earn needs dedicated refactoring sprints.
- Law and regulation: statutes and codes grow monotonically because repeal is far costlier than enactment.
- Bureaucracy: forms, sign-offs, and approval layers accumulate after perceived failures and rarely retire once the failures cease.
- Product design: "feature creep," where pressure to add user-visible features is sharp and pressure to remove them weak.
- Curriculum design: content, courses, and requirements accrete that pruning exercises rarely reverse.
- Biology: vestigial structures and genomic accretion, where adding under selection is incrementally easier than coordinated removal.
Clarity¶
It defeats the inference that accreted complexity must be earning its keep: absence-of-removal reflects an under-visited search region, not demonstrated necessity.
Manages Complexity¶
It collapses a catalogue of domain-specific bloat into one diagnostic — identify the search distribution and recognize the remove-region is underweighted unless deliberate machinery weights it.
Abstract Reasoning¶
A system left to additive-biased agents settles above its purpose-optimal complexity by a margin set by the bias strength, and the baseline shift makes each generation perceive the accreted state as the system, compounding the skew.
Knowledge Transfer¶
- Software → governance: a refactor-only sprint is the same move as a regulatory sunset review or a curriculum-pruning week.
- Regulation → product: sunset clauses and one-in-X-out rules import into roadmaps and process governance.
- Organizations: zero-based budgeting exports the "default-off" logic — justify every existing resource rather than treating it as baseline.
Example¶
Subjects asked to stabilize a Lego bridge overwhelmingly add a supporting block when removing one piece reaches the goal more cheaply — and a prompt that names subtraction as an option recovers the better solution.
Relationships to Other Primes¶
Parents (1) — more general patterns this builds on
- Additive Bias is a kind of Bias — Additive bias is a specific member of the bias family: a direction-asymmetric search distribution over transformations that over-samples 'add' and under-visits 'remove'. The 0.9991 sim to bias is genus-to-species — bias is the PARENT, NOT a reparent target.
Path to root: Additive Bias → Bias
Not to Be Confused With¶
- Additive Bias is not Bias because additive bias is the specific skew toward adding rather than removing in a transformation search, whereas bias is the umbrella of any systematic deviation.
- Additive Bias is not Scope Creep because additive bias is the deeper search-distribution asymmetry that produces growth across all substrates, whereas scope creep is requirements growing within a single project.
- Additive Bias is not Parsimony (Occam's Razor) because additive bias is a descriptive fact about how agents actually search, whereas parsimony is a normative preference for simpler explanations.