Contribution vs Causation: The Hard Problem Behind Merit-Based Funding

Getting your Trinity Audio player ready...

Merit-based funding faces a fundamental problem: identifying someone’s contribution is different from establishing how much benefit that person caused. A researcher may produce a valuable theorem, dataset, or software tool without anyone being able to calculate precisely what the world would have lost without it.

A credible funding system must distinguish documented work, its demonstrated usefulness, its estimated causal impact, and the reasons for paying its creator. This distinction matters for universities, grant agencies, scientific prizes, and experimental systems such as AI Internet-Meritocracy (AIIM).

Contribution, causation, and merit are different questions

In this article, contribution means identifiable work that participates in producing, validating, communicating, or maintaining a research output. Causal impact means the difference an action makes relative to a specified alternative. Merit is an evaluative judgment about what deserves recognition or reward.

Question	What it asks	Evidence or judgment needed
Contribution	What did this person do?	Manuscripts, code changes, datasets, contribution statements
Usefulness	How did others benefit?	Documented reuse, validation, applications, time saved
Causal impact	What would have happened without it?	A defined alternative, causal assumptions, and suitable evidence
Reward	What payment is deserved?	An explicit policy connecting evidence to values
Funding impact	What difference would paying this person now make?	Evidence about what funding enables, including uncertainty

These questions overlap, but answering one does not automatically answer the others.

The CRediT contributor taxonomy illustrates the distinction. It identifies roles such as conceptualization, software, investigation, and validation. These descriptions help establish who did what; they do not calculate each person’s share of the resulting social benefit.

A contribution record documents participation. It is not, by itself, a causal valuation.

Why “what would have happened otherwise?” is difficult

Suppose a researcher develops an algorithm that a laboratory uses to accelerate its work.

There is evidence of contribution: the algorithm exists and its authorship can be checked. There may also be evidence of usefulness: the laboratory reports completing a task faster.

But several alternatives remain possible. Without that algorithm, the laboratory might have:

• abandoned the task;
• completed it later;
• used a competing algorithm;
• developed an equivalent method itself.

Each alternative implies a different estimate of causal impact.

This is why causal inference requires more than observing a successful outcome. The Nobel Prize explanation of causal research describes how carefully designed comparisons and assumptions make particular cause-and-effect conclusions possible.

Scientific credit adds another difficulty: a discovery may be unique, embedded in decades of prior work, and impossible to reproduce under an alternative history. We can still make informed judgments, but their precision must match the evidence.

The alternative must include time

“Someone else would eventually have discovered it” does not imply that the original discovery had no impact.

Advancing a discovery by five years can matter substantially. The relevant comparison may therefore be an earlier discovery versus a later one, rather than discovery versus permanent ignorance.

A funding system should specify its time horizon. Near-term applications and long-term foundational value can produce very different assessments.

Shared causation does not produce automatic shares of credit

Consider a hypothetical project that creates 100 units of benefit. It requires both a scientist’s method and a developer’s implementation. Neither produces any benefit alone.

Removing the scientist eliminates all 100 units. Removing the developer also eliminates all 100 units.

Both are necessary under these assumptions. Yet a fixed reward budget cannot give each person 100% of the same pool.

Causal dependence does not automatically determine how a joint reward should be divided.

An equal split may be defensible in this symmetric example. Real collaborations introduce differences in originality, effort, responsibility, available substitutes, and prior agreements. Deciding which differences matter requires a rule of fairness.

The opposite problem appears when two researchers independently produce interchangeable solutions. Removing either may leave the outcome unchanged because the other solution remains available. A simple removal test could assign each zero marginal impact, despite both having completed valuable work.

This is a reason to distinguish recognition of achievement from estimates of indispensability. It also explains why a funding formula can be mathematically consistent while remaining ethically contestable.

Scientific dependencies create a double-counting problem

A discovery may depend on an earlier theorem, a public dataset, a software library, laboratory equipment, and years of maintenance.

Recognizing these dependencies can improve evaluation. However, assigning the entire downstream benefit independently to every contributor would count the same benefit repeatedly.

There are at least two coherent approaches:

• Divide a defined reward pool. Multiple contributors receive shares under an explicit allocation rule.
• Recognize multiple contributions separately. Each award acknowledges valuable work without claiming to represent an exclusive fraction of total social impact.

Either approach can be defensible. Confusion arises when a system issues overlapping recognition scores and presents their sum as a measured total of newly created benefit.

A dependency graph shows relationships between outputs. It does not, on its own, measure the strength of those relationships or the availability of substitutes.

Citations and prestige cannot settle causation

A citation establishes that one work refers to another. Its meaning depends on context: it may acknowledge a crucial method, provide background, or criticize an error.

Similarly, a prestigious publication venue does not establish how much benefit an individual researcher caused.

The San Francisco Declaration on Research Assessment (DORA) recommends assessing research on its merits, avoiding journal-based metrics as substitutes for individual assessment, and considering outputs such as datasets and software.

For merit-based funding, this supports a practical principle: inspect the contribution itself and use indicators as evidence requiring interpretation.

That principle also protects less visible work. A carefully documented correction, replication, or maintenance release can deserve support even when it lacks a dramatic impact story.

Rewarding past work and enabling future work are separate goals

A retrospective award asks: What has this person already contributed?

A prospective funding decision asks: What additional activity or benefit might this payment enable?

An outstanding researcher may deserve recognition even if another payment changes little about their immediate output. A less established researcher may have fewer past achievements but be able to undertake valuable work only with support.

Neither case invalidates merit-based funding. They show why funders should explain their objectives.

A system might deliberately reward completed work to recognize fairness, sustain contributors, or create incentives for others. But the causal effect of that reward policy must be evaluated separately from the quality of the work being rewarded.

Otherwise, the system risks confusing “this person did something valuable” with “this payment will produce the greatest additional benefit.”

What this means for AI Internet-Meritocracy

AIIM’s project description presents it as an experiment in evaluating documented public contributions and using AI-assisted assessments to recommend allocations of donated funds.

Its potential value does not depend on proving that AI can reconstruct a uniquely correct alternative history of science. A more defensible research question is whether it can produce useful, fair, and reviewable funding decisions at an acceptable cost.

The following are proposed evaluation principles, not claims that AIIM already implements them:

1. Separate evidence from inference. Distinguish verified authorship and reuse from estimated downstream impact.
2. Expose assumptions. State the alternative scenario, time horizon, and treatment of substitutes behind a causal estimate.
3. Make shared-credit rules explicit. Explain how dependencies and overlapping contributions affect allocations.
4. Represent uncertainty. Avoid implying that a precise numerical score has equally precise evidential support.
5. Allow correction and appeal. Contributors need a way to challenge missing records, mistaken attribution, and disputed evaluations.
6. Test decisions against alternatives. Compare results with expert review and simple baselines, including administrative cost and access for independent researchers.

AIIM’s proposed adversarial testing addresses manipulation and governance risks. A complementary test should examine ordinary attribution errors: can the system distinguish a crucial contribution from a superficial association, and can it recognize uncertainty when the evidence is insufficient?

A defensible meaning of merit-based funding

Merit-based funding need not mean paying each person an exactly measured fraction of everything their existence caused. That standard would require unavailable evidence and a universally agreed theory of fairness.

A workable standard is to reward substantiated contributions through explicit rules, qualified impact estimates, and procedures that can be challenged and improved.

The hard problem is not only discovering who helped. It is deciding how evidence of help should become a claim on limited funds. AIIM and other funding systems become more credible when they make that decision visible.