🤖 AI

Beyond Pairwise Comparisons: A Distributional Test of Distinctiveness for Machine-Generated Works in Intellectual Property Law

This paper proposes a distributional, two-sample test based on maximum mean discrepancy to determine if machine-generated works are statistically distinct from human-created ones, demonstrating that generative models produce semantically human-like yet stochastically unique outputs that challenge the view of them as mere regurgitators of training data.

Original authors: Anirban Mukherjee, Hannah Hanwen Chang

Published 2026-01-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anirban Mukherjee, Hannah Hanwen Chang

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 Problem: The "One-by-One" vs. The "Infinite Ocean"

Imagine you are a judge trying to decide if a new painting is truly original or just a copy of old ones.

How the law usually works:
Currently, the law acts like a detective looking at two specific photos side-by-side. They ask, "Does this specific new painting look exactly like that specific old painting?" If the answer is no, they might say it's original. This works fine if you are comparing two human artists who have made a few hundred paintings each.

The AI Problem:
Now, imagine an AI artist that can paint an infinite number of pictures. It doesn't just have a fixed portfolio; it has a "generative process" that can create endless variations.

If you only compare one AI picture to one human picture, you might get lucky and find they look different.
But if you pick a different AI picture, it might look identical to a human one.
Because the AI can make infinite pictures, checking them "one by one" is impossible. It's like trying to find a specific grain of sand on a beach by looking at one grain at a time. You'll never know if the whole beach is different.

Furthermore, humans are bad at spotting the difference. Studies show that even experts can only tell AI art from human art about 58% of the time (barely better than flipping a coin). So, the human "judge" is failing.

The Solution: The "Fingerprint of the Process"

The authors propose a new way to look at the problem. Instead of comparing Item A to Item B, they compare the fingerprint of the whole process that made Item A against the fingerprint of the process that made Item B.

Think of it like this:

Old Way: Comparing two individual fingerprints to see if they match.
New Way: Analyzing the soil where the fingerprints were found. If one set of prints came from a muddy forest and the other from a sandy beach, you know the process of making the prints is different, even if two specific prints look similar.

How It Works: The "Statistical Taste Test"

The authors use a mathematical tool called Maximum Mean Discrepancy (MMD). Here is a simple way to understand it:

Translation: First, they translate images (or text) into a "language" that computers understand, called embeddings. Imagine turning a painting into a long list of numbers that describe its style, colors, and mood, rather than just its pixels.
The Cloud: They take a small group of AI paintings and a small group of human paintings and turn them into two "clouds" of points in a multi-dimensional space.
The Test: They ask a statistical question: "Are these two clouds coming from the same source?"
- If the AI is just regurgitating (copying and pasting) old training data, its cloud will sit right on top of the human cloud.
- If the AI is interpolating (mixing and matching ideas to create something new), its cloud will be in a different spot, even if it's still in the same neighborhood.

The Surprising Discovery: The "Perceptual Paradox"

The paper found something very interesting, which they call a Perceptual Paradox:

What Humans See: Humans look at an AI painting and a human painting and say, "They look the same." (They can't tell the difference).
What the Math Sees: The math looks at the distribution of the AI paintings and says, "These are statistically different from the human paintings."

The Analogy:
Imagine a human chef and a robot chef.

The human chef makes 1,000 unique soups.
The robot chef tastes the human soups and learns the recipe.
If the robot just regurgitates, it serves you the exact same 1,000 soups.
If the robot interpolates, it creates 1,000 new soups that taste very similar to the human ones, but if you analyzed the chemical composition of the entire batch, you'd find the robot's soups have a slightly different "flavor profile" because it's mixing ingredients in new ways.

The paper proves that AI is doing the second thing. It isn't just copying; it is creating a new, distinct "flavor profile" that is mathematically different from humans, even though it tastes (looks) the same to our senses.

Why This Matters for Law

It's Fast and Cheap: You don't need to see thousands of pictures. The math can tell the difference with as few as 5 to 10 images per group. This is great for court cases where you can't get access to the AI's secret training data.
It's Objective: It removes the human bias. Since humans can't reliably tell the difference, this tool gives the judge a "ruler" that actually works.
It Proves Creativity (Sort of): The results suggest that AI isn't just a "photocopier" of the internet. It is a "semantic interpolator." It learns the rules of art and creates new combinations within those rules. This challenges the idea that AI is just stealing and repeating.

Summary

The paper argues that to judge AI in court, we need to stop looking at single pictures and start looking at the statistical pattern of the whole group. They built a math test that can spot the difference between human and AI art with high accuracy, even when humans can't. This proves that AI creates a unique, distinct "distribution" of work, suggesting it is doing something more complex than just copying and pasting.

Technical Summary: Beyond Pairwise Comparisons: A Distributional Test of Distinctiveness for Machine-Generated Works in Intellectual Property Law

1. Problem Statement

Intellectual property (IP) law relies on a shared empirical premise: to warrant protection (novelty, originality, or distinctiveness), a work must be meaningfully distinct from a relevant reference class. However, current legal and technical practices suffer from a fundamental unit-of-analysis mismatch.

Legal Practice: Courts typically assess distinctiveness using item-level, pairwise comparisons (e.g., comparing a single painting to a specific prior art reference) mediated by human proxies ("ordinary observer" or "person having ordinary skill in the art").
The Challenge of Generative AI: Generative models are stochastic processes with effectively unbounded output spaces. Assessing a finite sample of outputs against a reference class yields an incomplete assessment. Furthermore, human evaluators struggle to distinguish AI-generated art from human art (achieving only ~58% accuracy), rendering the "ordinary observer" proxy unreliable.
Theoretical Impasse: The debate between "regurgitation" (AI merely memorizing training data) and "interpolation" (AI generating novel combinations) cannot be resolved by item-level metrics. Verbatim checks miss stylistic mimicry, while pairwise semantic similarity lacks a principled statistical basis for determining if a process is distinct from a distribution.

2. Methodology

The authors propose a distributional alternative: a two-sample hypothesis test based on Maximum Mean Discrepancy (MMD) computed on semantic embeddings. This shifts the unit of analysis from the individual item to the creative process.

2.1 Core Framework

Semantic Embedding: Non-numerical data (images, text) are mapped into high-dimensional vector spaces using pre-trained foundation models (e.g., CLIP for images, SentenceTransformers for text). This captures semantic relationships rather than surface form.
Kernel Mean Embeddings (KME): The probability distributions of the creative processes (Reference Class $P$ vs. Candidate Process $Q$ ) are mapped into a Reproducing Kernel Hilbert Space (RKHS).
MMD Calculation: The squared MMD ( $\text{MMD}^2$ $MMD^{2}$ ) is computed as the distance between the KMEs of the two distributions. This quantifies the statistical divergence between the processes.
- Formula: $\text{MMD}^2(P, Q) = \|\mu_P - \mu_Q\|_H^2$ .
- The estimator uses a Gaussian Radial Basis Function (RBF) kernel, which is "characteristic," ensuring that $\text{MMD}^2 = 0$ if and only if $P = Q$ .
Hypothesis Testing: To provide legal admissibility, the raw MMD score is converted into a p-value via a permutation test.
- Null Hypothesis ( $H_0$ ): $P = Q$ (The processes are identical; the candidate is not distinct).
- Alternative Hypothesis ( $H_1$ ): $P \neq Q$ (The processes are distinct).
- The test determines if the observed divergence exceeds what would be expected by chance, allowing fact-finders to set a standard of proof (significance level $\alpha$ ).

2.2 Stability and Robustness

Approximate Embeddings: The authors prove (Proposition 2) that MMD is Lipschitz-continuous with respect to bounded kernel perturbations. Even with dimensionality reduction (e.g., UMAP) or noise, the metric remains stable.
Sample Efficiency: The method is designed to be effective with small sample sizes, crucial for legal contexts where large datasets may be unavailable.

3. Key Contributions

Shift from Item to Process: The framework resolves the "infinite cardinality" problem of generative AI by treating distinctiveness as a property of the underlying distribution rather than a pairwise comparison of specific outputs.
Training-Free and Sample-Efficient Implementation: The method requires no task-specific training or access to proprietary model weights. It can detect distributional differences with as few as 5–10 images or 7–20 texts per group.
Identification of "Interpolative Distinctiveness": The paper provides empirical evidence that generative models do not merely regurgitate training data. Instead, they produce outputs that are semantically human-like but stochastically distinct, occupying unique topological regions within the learned latent space.

4. Experimental Results

4.1 Validation on MNIST (Controlled Images)

Setup: Compared distributions of handwritten digits (0–9) using a LeNet-5 embedding.
Findings: The method achieved >95% statistical power (rejection of $H_0$ ) with as few as 6 samples per distribution. It successfully distinguished visually similar digits (e.g., 3 vs. 5) where raw pixel comparisons failed.
Robustness: The metric remained stable under dimensionality reduction and input perturbations (noise, watermarks), failing to reject the null hypothesis for perturbed versions of the same image.

4.2 Validation on Patent Abstracts (Text)

Setup: Compared patent abstracts across three International Patent Classification (IPC) sections: Human Necessities (A), Chemistry (C), and Electricity (H).
Findings: The method reliably distinguished between unrelated technical fields (e.g., Chemistry vs. Electricity) with 7–15 samples per group. Split-half controls (comparing subsets of the same field) yielded non-significant results, confirming the method does not detect spurious differences.

4.3 AI-Generated Art (Real-World Images)

Dataset: AI-ArtBench (Human art vs. Stable Diffusion and Latent Diffusion outputs across 10 artistic styles).
The Perceptual Paradox: While human evaluators distinguish AI from human art with only ~58% accuracy, the MMD test detected robust distributional distinctiveness across all styles ( $p < 0.01$ ).
Magnitude of Distinctiveness: The distance between Human and AI distributions was comparable to the distance between historically distinct human art movements (e.g., Art Nouveau vs. Ukiyo-e).
Memorization Audit: Nearest-neighbor audits (CLIP, SSIM, LPIPS) found no systematic memorization. The few flagged instances were visualized as stylistic convergences (e.g., shared color palettes) rather than content replication.
Evolution of Models: As models evolved from Latent Diffusion (2021) to FLUX-Krea (2025), distributional distinctiveness increased (MMD scores rose), suggesting models are becoming better interpolators, not better regurgitators.
Embedding Sensitivity: Distinctiveness was highest in CLIP (AI-naive), intermediate in DreamSim (AI-aware but human-calibrated), and lowest in the Stable Diffusion VAE (AI-native). This confirms that AI outputs are genuine outliers in human semantic space but fit the latent manifold of the generative process.

5. Significance and Claims

The paper claims to offer a principled, quantitative framework for IP adjudication that addresses the limitations of current human-centric and pairwise methods.

Legal Admissibility: By converting distributional distances into p-values via permutation testing, the method provides a testable empirical evidence with known error rates, satisfying criteria like Daubert.
Bifurcated Inquiry: The authors propose a dual approach:
1. MMD Test: To adjudicate the nature of the generative process (is it a distinct creative process or a mere resampler?).
2. Memorization Audits: To detect specific, rare instances of item-level infringement (verbatim copying).
Implications for IP Doctrines:
- Copyright: Provides evidence of "probabilistic originality" and independent creation, countering the "regurgitation" critique.
- Patent: Offers a metric for "generative distance" relevant to non-obviousness.
- Trademark: Can quantify the distinctiveness of a brand's visual footprint against a market distribution.
Limitations: The authors acknowledge that the method depends on the quality of embeddings (which may inherit biases) and representative sampling. It measures the dominant mode of the process and does not rule out rare memorization events, necessitating the bifurcated approach.

In summary, the paper argues that generative models function as semantic interpolators within a learned latent space, producing outputs that are statistically distinct from their training distributions, even when they appear perceptually similar to human creations. This distributional distinctiveness challenges the view of AI as a mere tool for regurgitation and suggests a need for IP law to evolve from intuition-based pairwise tests to distributional evidence.