🤖 machine learning

Epistemology gives a Future to Complementarity in Human-AI Interactions

This paper addresses the theoretical and empirical limitations of human-AI complementarity by reframing it through an epistemological lens of computational reliabilism, arguing that it should serve as evidence of a reliable epistemic process to better guide decision-making and inform design and governance practices.

Original authors: Andrea Ferrario, Alessandro Facchini, Juan M. Durán

Published 2026-04-23

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Andrea Ferrario, Alessandro Facchini, Juan M. Durán

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Why "Working Together" Isn't Always Enough

Imagine you are trying to solve a difficult puzzle. You have a friend (a human) and a super-fast robot (an AI).

For a long time, researchers have been obsessed with a concept called Complementarity. The idea is simple: If the human and the robot work together, they should solve the puzzle better than the human alone OR the robot alone.

The paper argues that while this sounds great, treating "Complementarity" as the only gold standard for success is a mistake. It's like judging a restaurant solely by how much food is on the plate, ignoring whether the food is fresh, safe to eat, or if the chef is actually skilled.

The authors, Andrea Ferrario and colleagues, propose a new way to look at this. They suggest we stop asking, "Did they beat the score?" and start asking, "Is this a reliable way to make decisions?" They use a philosophical framework called Computational Reliabilism to explain why.

The Problem: The "Complementarity Trap"

The paper identifies four big problems with how we currently measure human-AI teamwork:

It's a Hindsight Mirror: You can only know if the team was "complementary" after you know the right answer. But in real life (like a doctor diagnosing a patient), you don't know the answer yet. You need to know now if you can trust the team.
It Ignores the "Cost": What if the human and AI team gets a slightly better score, but it takes the human 10 hours of extra work to get there? That's a bad deal. The current definition doesn't care about the time or effort wasted.
It's Too Narrow: A team could be "complementary" (better than either alone) but still be unfair, biased, or legally risky. Just because you improved the score doesn't mean you did the right thing.
It's Hard to Achieve: In real-world studies, human-AI teams often don't beat the best of the two working alone. They often just make the same mistakes or get confused.

The Solution: The "Reliability Report Card"

Instead of a simple "Pass/Fail" on whether the team beat the solo scores, the authors suggest we view the Human-AI team as a Reliable Process.

Think of it like a Flight Safety Inspection.

The Old Way (Complementarity): "Did this plane fly 10% faster than a car?" (Irrelevant if the plane crashes).
The New Way (Reliabilism): "Is this plane safe to fly?" To answer that, we check three different things (called Reliability Indicators):

1. The Engine Check (Technical Performance)

This is where "Complementarity" fits in. Did the team actually perform better than the parts alone?

Analogy: If the pilot and the autopilot work together, do they land the plane smoother than the pilot alone or the autopilot alone?
The Catch: This is just one piece of evidence. It helps, but it doesn't guarantee safety.

2. The Blueprint Check (Epistemic Standards)

Is the task even being done correctly?

Analogy: Are we using the right map? Is the "destination" we are aiming for actually the right place? Are we measuring the right things?
Example: If an AI helps a judge decide on bail, is the AI measuring "risk of re-offending" correctly, or is it just measuring "how poor the defendant is"? If the blueprint is wrong, the team is unreliable, no matter how well they work together.

3. The Crew & Rules Check (Socio-Technical Practices)

Are the humans trained? Are there rules for when to panic?

Analogy: Is the pilot certified? Is there a clear rule for what to do if the engine sputters? Is there a logbook?
Example: If a doctor uses an AI, do they have training on when to ignore the AI? Is there a system to report if the AI starts making weird mistakes?

The "Efficiency" Meter: Is the Gain Worth the Pain?

The paper introduces a new concept: Efficient Complementarity.

Imagine you are baking a cake.

Scenario A: You and a robot mix the batter. You get a cake that is 1% tastier than if you baked it alone. But, it took you 5 extra hours to argue with the robot about the flour.
Scenario B: You and a robot mix the batter. You get a cake that is 20% tastier, and it took you 5 minutes.

The old way of thinking just says, "Scenario A is complementary because you did better!"
The new way says, "Scenario A is inefficient. The cost (5 hours of arguing) outweighs the tiny gain (1% tastier). Scenario B is efficient."

The authors argue we need to measure the Net Gain: How much better is the result, minus how much effort it took to get there?

Real-World Examples from the Paper

The Dermatologist (The Good Team): A skin doctor uses an AI to spot cancer. They have training, the AI is well-tested, and they know when to double-check the AI. Sometimes they beat the AI alone, sometimes they beat the doctor alone. This is a Reliable Process because all three checks (Engine, Blueprint, Crew) are strong.
The Student (The "Almost" Team): A student uses a free AI chatbot to do math homework. The AI makes mistakes, but the student catches them. Sometimes they get the right answer together. This is Partially Reliable. The "Engine" (complementarity) works sometimes, but the "Blueprint" (the AI isn't a certified math tool) and "Crew" (no teacher oversight) are weak. It's risky to trust this for a final exam.
The Forensic Lab (The "No Gain" Team): A lab uses a voice-recognition AI. The human expert is already so good that the AI adds almost nothing. But, checking the AI takes time and slows down the court case. Here, Complementarity is low, but the process is still Reliable because the human is the expert, and the rules (Blueprint/Crew) are strict. Forcing them to "optimize for complementarity" would just waste time and money.

The Takeaway: What Should We Do?

The authors suggest we stop treating "Complementarity" as the only goal. Instead, we should use a Checklist for Trust.

When a hospital, a bank, or a government wants to use Human-AI teams, they shouldn't just ask, "Does it work better?" They should ask:

What is the protocol? (How do they talk to each other?)
What is the cost? (How much time/money does it take?)
Is it fair? (Does it hurt any specific group?)
Is it monitored? (What happens if the AI breaks?)

In short: Don't just look for the magic moment where Human + AI > Human + AI. Look for a reliable system where the human and AI work together safely, fairly, and efficiently, even if they don't always win the "scoreboard" game.

The paper ends by providing a Checklist (Table 1 in the appendix) that designers and regulators can use to ensure these teams are actually trustworthy, not just statistically "better."

1. Problem Statement

The paper addresses the theoretical and empirical limitations of Human-AI Complementarity as currently defined in the literature. Complementarity is the claim that a human-AI team outperforms both the human and the AI acting alone in decision-making tasks. While intuitively appealing, the authors identify four critical challenges preventing it from serving as a robust "gold standard" for Human-AI Interaction (HAI):

Lack of Theoretical Anchoring: Complementarity lacks a precise theoretical foundation, often relying on generic Input-Processing-Output models that fail to distinguish between different interaction types (e.g., reliance vs. active collaboration).
Post Hoc Limitations: Current metrics (Complementarity Team Performance, or CTP) are ground-truth-dependent and calculated ex post. They cannot guide real-time decision-making because ground truth is unavailable at the moment of decision.
Narrow Focus on Relative Accuracy: CTP is a binary, relative metric. A team can achieve complementarity even if absolute performance is poor, provided it is better than the individual components. This ignores critical non-epistemic desiderata such as fairness, robustness, explainability, and resource constraints.
Ignoring Magnitude-Cost Profiles: The literature abstracts away from the magnitude of the performance gain versus the cost (cognitive load, time, institutional overhead) required to achieve it. A marginal gain achieved through excessive human effort may not be practically valuable.

Consequently, empirical studies show that complementarity is difficult to achieve systematically, and its current formulation is ill-suited for high-stakes domains like healthcare and law.

2. Methodology: Computational Reliabilism (CR)

The authors propose a theoretical repositioning of complementarity using Computational Reliabilism (CR), a framework rooted in epistemology (specifically process reliabilism).

Core Concept: In CR, a belief (or prediction) is justified if it is produced by a reliable process. The unit of evaluation shifts from the AI system in isolation to the Human-AI interaction as a computational process (a Prediction-Task Human-AI Interaction, or PT-HAI).
Reliability Indicators: Reliability is not binary but graded, supported by a family of heterogeneous indicators (Type-RIs) that provide defeasible evidence of process reliability:
- Type-RI1 (Technical/Operational): Performance metrics, error patterns, and stability.
- Type-RI2 (Epistemic/Scientific): Alignment with domain standards, validity of the target construct, and uncertainty calibration.
- Type-RI3 (Socio-Technical): Governance, training, monitoring, and institutional practices that sustain reliability over time.
Reframing Complementarity: The authors treat complementarity not as a standalone success metric, but as a specific Type-RI1 indicator. Historical instances of complementarity serve as evidence that the interaction protocol yields epistemic gains beyond individual components, thereby supporting the claim that the PT-HAI is a reliable process.

3. Key Contributions

A. Theoretical Repositioning

The paper argues that complementarity should be viewed as evidential rather than constitutive.

It is not a necessary condition for reliability (a team can be reliable via strict human oversight even without complementarity).
It is not a sufficient condition (a team can show complementarity while being unsafe due to lack of governance).
Instead, it is a historical signal that helps justify the epistemic adequacy of the team's output to third parties (patients, regulators, managers).

B. Operationalizing "Efficient Complementarity"

To address the magnitude-cost challenge, the authors introduce a formal framework for Net Complementarity Gain:

Gross Gain ( $\Delta$ ): The absolute reduction in loss achieved by the team compared to the best standalone component.
Cost ( $c$ ): The interactional and organizational resources required (time, monitoring, documentation).
Efficiency Threshold ( $\lambda$ ): An institutional parameter representing the minimum acceptable rate of improvement per unit of cost.
Formula: Complementarity is efficient if $\Delta_{\tau}(D) > \lambda \cdot c(D)$ .
This allows organizations to determine if the epistemic gain justifies the operational burden.

C. Minimal Reporting Checklist

The authors propose a Justificatory Checklist (Table 1 in the Appendix) for designing and governing Human-AI systems. This checklist moves beyond simple accuracy reporting to include:

Protocol specifications (how disagreement is handled).
User competence and training.
Disaggregated performance (subgroup analysis).
Cost-benefit analysis of complementarity (Net Gain).
Lifecycle governance (updates, drift management).

4. Results and Illustrative Examples

The paper validates the framework through three case studies demonstrating how CR handles complementarity differently than traditional metrics:

The Technologist Dermatologist: High reliability is achieved because complementarity (Type-RI1) is supported by strong Type-RI2 (medical validity) and Type-RI3 (training/audits). The gain is efficient because the clinician's expertise reduces the cognitive cost over time.
The Student and the LLM: A student using an unreliable LLM achieves complementarity (detecting/correcting errors). However, the PT-HAI has low overall reliability because Type-RI2 (validation) and Type-RI3 (governance) are weak. The complementarity signal is insufficient to justify the output for high-stakes use.
The Forensic Lab: Here, complementarity gains are marginal (the human and AI mostly agree). However, the PT-HAI is highly reliable due to strict Type-RI3 constraints (legal protocols, version control). The authors argue that optimizing for marginal complementarity gains would be inefficient and harmful due to high operational costs (delays in court).

5. Significance and Implications

For Designers: Designers should stop optimizing solely for complementarity. Instead, they must design interaction protocols that make epistemic behaviors observable (e.g., structured contestation) and report the cost-efficiency of any gains.
For Regulators: Mandating complementarity as a compliance target is dangerous as it encourages optimizing a relative metric while ignoring broader reliability. Regulations (like the EU AI Act) should focus on lifecycle governance and the stability of the socio-technical process.
For Epistemology: The paper bridges the gap between performance metrics and epistemic justification. It clarifies that "trust" in AI should be replaced by justification based on process reliability, where complementarity is just one piece of evidence among many.
Practical Utility: By introducing the concept of Efficient Complementarity, the paper provides a quantitative tool for stakeholders to decide when human-AI collaboration is worth the effort, preventing the deployment of systems that offer negligible gains at high cognitive or institutional costs.

In conclusion, the paper argues that complementarity has a future in Human-AI interactions, but only if it is demoted from a "gold standard" to a component of a broader, graded reliability framework grounded in computational reliabilism.