AI Art Ethics: Ownership, Bias, and Indian Cultural Context

On 24 July 2026, Justice Amit Bansal of the Delhi High Court refused to grant Asian News International (ANI) an interim injunction against OpenAI. ANI had sued in November 2024, arguing that ChatGPT was trained on its copyrighted news reports without a licence and, worse, sometimes attributed fabricated statements to ANI's byline. OpenAI raised a jurisdiction objection — arguing an Indian court had no business hearing a case against a company running servers in the United States — but Justice Amit Bansal rejected it outright, holding the Delhi High Court did have territorial jurisdiction. The injunction was dismissed on the merits instead: the court held that training ChatGPT on ANI's reports fell within the Section 52(1)(a)(i) fair-dealing exception for private use and research, and found no substantial similarity between ChatGPT's outputs and ANI's copyrighted text, with no evidence the model had memorized it. The underlying copyright suit is still pending; this was only the interim-injunction ruling. Nothing about this case is unique to text. Swap "news article" for "Madhubani painting" and "ChatGPT" for a diffusion model, and every legal question ANI raised — was the training use licensed, does the output reproduce protectable expression, who counts as an "author" when a machine does the drawing — applies identically to AI-generated images. This chapter builds the technical and legal machinery to reason about that question precisely, using two computations you can run yourself: an embedding-similarity test for "did the model just copy this," and a disparity index for "does the model's output distribution match reality."

How a Diffusion Model Actually Produces an Image

You already know, from deep learning foundations, that a neural network is a function trained to minimize a loss over a distribution of examples. A text-to-image diffusion model (Stable Diffusion, Midjourney's backend, DALL·E) is trained by taking real images, progressively adding Gaussian noise until they become pure static, and training a U-Net to predict — at each noise level — what noise was added. Generation runs this backward: start from pure noise, and repeatedly subtract the network's noise prediction, conditioned on a text-embedding of your prompt, until a coherent image emerges.

Two things follow directly from this mechanism, and both are load-bearing for the ethics discussion:

• The model does not store a lookup table of training images. It stores millions of weights that jointly encode a learned probability distribution over "what images look like," conditioned on text. In the overwhelming majority of cases, a generated image is a genuinely novel sample from that distribution — no single training image is being reproduced.
• Memorization is still possible and has been empirically measured. When a training image appears many times in the dataset (a famous photograph reposted across thousands of web pages, for instance), the denoising network can learn to collapse the diffusion trajectory back to that one image almost exactly, regardless of prompt diversity. Carlini et al. (2023), in "Extracting Training Data from Diffusion Models," demonstrated this empirically on Stable Diffusion by generating thousands of images and checking each against the training set for near-duplicates.

The diagram below shows the full pipeline, including the checkpoint researchers and courts actually use to tell these two cases apart: nearest-neighbour similarity in embedding space.

Ownership Under Indian Law: Section 2(d)(vi) and the Substantial-Similarity Test

The Indian Copyright Act, 1957, has an unusual provision most jurisdictions lack. Section 2(d)(vi) defines the "author" of a computer-generated literary, dramatic, musical, or artistic work as "the person who causes the work to be created." This is deliberately different from the US Copyright Office's 2023 position, which denies copyright protection to a work with no human author at all. Indian law instead grants authorship to a human "causer" — typically read as the person who wrote the prompt and made the creative choices that shaped the output. So if you type a detailed prompt into a diffusion model and curate the result, Indian law is comparatively willing to say you own the output image.

That answers only half the question. Owning the output does not settle whether producing it was lawful in the first place. A second, independent question is whether training the model on someone else's copyrighted images — without a licence — was itself an infringing use, and whether a specific output substantially reproduces a specific protected work. This is exactly what the ANI case turns on for text, and it maps onto images through the same two-part test courts use for infringement anywhere: (1) did the defendant have access to the copyrighted work, and (2) is there substantial similarity between the protected expression and the accused work. Access is easy to establish for a model trained on billions of scraped images. Substantial similarity is the hard, technical part — and it is exactly what embedding similarity is used to approximate.

Worked Example: Detecting Memorization with Embedding Similarity

Image encoders (like the vision half of CLIP) map an image to a fixed-length vector such that visually and semantically similar images land close together. Cosine similarity between two such vectors is the standard proxy for "how similar are these two images," and it is what memorization-detection research and takedown investigations actually compute. Cosine similarity of vectors a and b is:

cosθ = (a · b) / (|a| |b|)

Suppose a training image T has (a toy, 4-dimensional) embedding [0.8, 0.6, 0.0, 0.0], and we generate two images from a diffusion model, getting embeddings G1 = [0.6, 0.8, 0.0, 0.0] and G2 = [0.0, 0.0, 1.0, 0.0]. Here is the exact computation:

import math

def cosine_similarity(a, b):
    dot = sum(x * y for x, y in zip(a, b))
    norm_a = math.sqrt(sum(x * x for x in a))
    norm_b = math.sqrt(sum(y * y for y in b))
    return dot / (norm_a * norm_b)

T  = [0.8, 0.6, 0.0, 0.0]   # embedding of a specific training image
G1 = [0.6, 0.8, 0.0, 0.0]   # embedding of generated image #1
G2 = [0.0, 0.0, 1.0, 0.0]   # embedding of generated image #2

print(round(cosine_similarity(T, G1), 4))
print(round(cosine_similarity(T, G2), 4))

Trace it by hand before trusting the code. For (T, G1): the dot product is 0.8×0.6 + 0.6×0.8 + 0×0 + 0×0 = 0.48 + 0.48 = 0.96. Both vectors have norm √(0.8²+0.6²) = √1.0 = 1.0, so cosθ = 0.96 / (1.0×1.0) = 0.96. For (T, G2): the dot product is 0.8×0 + 0.6×0 + 0×1 + 0×0 = 0, so cosθ = 0.0. The program prints:

0.96
0.0

A cosine similarity of 0.96 is near the ceiling of 1.0 (identical direction) — this is the signature Carlini et al. looked for: an output landing almost exactly on top of one specific training image, which is strong evidence of memorization rather than synthesis, and the kind of evidence that would support a substantial-similarity claim. A cosine similarity of 0.0 means the vectors are orthogonal — G2 shares essentially nothing with T and is comfortably in "novel synthesis" territory. Real memorization-detection pipelines run this check against millions of training images and flag any output whose highest similarity crosses a threshold (commonly around 0.9, as used in the diagram above); a single high-similarity match is what turns "the model drew something new" into "the model regurgitated a specific photograph."

Bias as a Measurable Quantity: The Representation Disparity Index

Bias in a generative model is not a vague discomfort — it is a measurable mismatch between the distribution the model outputs and a reference distribution you care about. Define the Representation Disparity Index for a category c as:

DI(c) = p_generated(c) / p_baseline(c)

where p_generated(c) is the fraction of a generated image batch that falls into category c, and p_baseline(c) is the real-world reference fraction. DI = 1 means the model matches reality; DI < 1 means under-representation; DI > 1 means over-representation. Bianchi et al. (2023, published at FAccT) ran exactly this kind of audit at scale on Stable Diffusion and found systematic amplification of occupational and national stereotypes relative to real-world statistics — the method below is a simplified, hand-traceable version of that idea, using illustrative numbers so you can verify the arithmetic yourself.

Suppose you prompt a diffusion model with "a scientist" fifty times and, on inspection, 4 of the 50 images depict women. For illustration, suppose the real-world baseline share of women among scientists in the reference population is 0.28 (28%) — a hypothetical figure chosen only to demonstrate the calculation, not a cited statistic.

def disparity_index(p_generated, p_baseline):
    return p_generated / p_baseline

N = 50
women_shown = 4
p_gen_women = women_shown / N          # generated share

p_baseline_women = 0.28                # illustrative reference share

DI = disparity_index(p_gen_women, p_baseline_women)
print(round(p_gen_women, 4))
print(round(DI, 4))
print(round(1 / DI, 4))

Trace it: p_gen_women = 4 / 50 = 0.08. DI = 0.08 / 0.28 = 2/7 ≈ 0.2857. 1 / DI = 0.28 / 0.08 = 3.5 exactly. The program prints:

0.08
0.2857
3.5

Read this as: the model generated women scientists at roughly 29% of the rate you would expect from the baseline (DI ≈ 0.29), which is the same as saying the baseline category is under-represented by a factor of 3.5×. This is the exact shape of finding published bias audits report, and it is a computation any student can rerun on a real model's outputs by counting a batch of generated images against a baseline of their choosing — the method generalizes to skin tone, religion, region, or any category where you can define both an output count and a reference proportion. Two things make this specifically dangerous for Indian users: first, image encoders and the datasets behind them (LAION-scale corpora) are overwhelmingly sourced from English-language web pages, so prompts naming Indian occupations, festivals, or regional dress draw on a thinner and less curated slice of training data than prompts about globally dominant (usually Western, English-labeled) imagery; second, that thinness compounds with existing stereotypes already present in web imagery, so DI values for Indian-context prompts tend to be more extreme, not less, than the illustrative example above.

Indian Cultural Context: GI Tags, Style, and the Idea-Expression Gap

Several Indian art forms carry Geographical Indication (GI) protection: Madhubani painting from Bihar's Mithila region (GI-tagged in 2007, supporting an estimated 25,000–30,000 artisans), Pattachitra scroll painting from Odisha (GI-tagged in 2008), and Warli painting from Maharashtra, among others. A GI tag legally restricts who may use the geographic name — a piece cannot be sold as "Madhubani painting" unless it comes from that region and community. This looks, on the surface, like exactly the protection an artisan community would want against AI appropriation. It is not, and understanding why requires the same idea-expression dichotomy you would use to reason about why copyright protects a specific song's melody but not the abstract idea of a "sad love song."

Copyright and GI law both protect specific things, not visual styles in the abstract. A GI tag protects the name "Madhubani" from being misapplied to non-Mithila work; it says nothing about whether a diffusion model may learn the visual grammar of the style (bold black outlines, double-line borders, geometric fish and peacock motifs, natural dye palettes) from thousands of scraped photographs and then generate new images "in that style" without ever calling the output "Madhubani." Copyright protects a specific artist's specific composition; it does not protect the folk-art style itself, any more than a novelist can copyright the genre of "epic fantasy." The result is a genuine legal gap: a model can be trained on an entire regional art tradition, absorb its distinctive visual language, and generate unlimited stylistically faithful output competing directly with the livelihoods of the artisans whose work trained it — without infringing either GI law (no false regional label is used) or copyright law in the narrow substantial-similarity sense (no single artist's specific composition is being reproduced), unless a specific output crosses the memorization threshold from the earlier section. The economic harm is real and measurable in lost commissions; the two legal frameworks built to protect these communities were not designed with this failure mode in mind. This is the same gap the Delhi High Court is now being asked to rule on for text, generalized to image style.

Correcting a Common Misconception

The most common misconception a student forms at this point is: "If the AI's output is pixel-different from any training image, it cannot be a copyright problem, since nothing was literally copied." This is wrong on two independent grounds, both established above. First, the substantial-similarity legal test never required pixel-identical copying anywhere, for any medium — a painter who repaints another artist's exact composition in different colours and a slightly different medium still infringes, because copyright protects the protectable expression (composition, arrangement, distinctive detail), not the literal pixel or ink values. Second, even at the pixel level, "different pixels" is not the same as "no similarity" — the cosine-similarity computation above showed exactly how a generated image can sit at cosθ = 0.96 to a specific training photograph without being a literal pixel-for-pixel copy, and that level of similarity is precisely the empirical signature Carlini et al. used to prove Stable Diffusion memorizes and regurgitates training images. "Not identical" and "not infringing" are two different claims, and conflating them is the exact error that makes this topic feel more legally settled than it is.

Active Recall

Attempt each question before reading the answer below it.

1. Why can a diffusion model generate a near-exact copy of a specific training photograph even though it never stores that photograph directly?
2. Under Section 2(d)(vi) of the Indian Copyright Act, 1957, who is treated as the "author" of a computer-generated artistic work, and how does this differ from the current US position?
3. A generated image has cosine similarity 0.94 with one specific training image and similarity below 0.3 with every other training image in the corpus. What does this pattern suggest, and what threshold from this chapter would flag it?
4. A model generates 40 images for the prompt "an engineer," of which 6 depict women. If the real-world baseline share of women engineers is 0.20, compute the Representation Disparity Index and interpret it in one sentence.
5. Explain why a GI tag on "Pattachitra painting" does not stop a diffusion model from generating unlimited new images in the Pattachitra visual style.
6. A classmate says, "The AI painting doesn't look exactly like any real Warli mural, so there's no ethical issue." Identify the specific flaw in this reasoning using two distinct concepts from this chapter.

Answers

1. The network is trained to predict noise at each diffusion step across the whole distribution of training images, not to memorize individual images. But when one image (or very similar copies of it) appears many times in the training set, gradient descent can push the network toward a near-collapse onto that one image, so the same denoising trajectory reproduces it almost exactly regardless of prompt wording — this is the phenomenon Carlini et al. (2023) measured directly on Stable Diffusion.
2. The "author" is "the person who causes the work to be created" — generally the prompt-writer who makes the creative choices that shape the output. This differs from the current US Copyright Office position, which denies copyright protection to a work altogether if it lacks a human author; Indian law instead assigns human authorship to the causer rather than refusing protection.
3. One very high similarity (0.94, above the 0.9 threshold used in the diagram) paired with uniformly low similarity to everything else is exactly the "spike" pattern that indicates memorization of one specific image rather than genuine synthesis from the learned distribution — it would be flagged as a memorization risk and a candidate for a substantial-similarity review.
4. p_generated = 6/40 = 0.15. DI = 0.15/0.20 = 0.75. The model under-represents women engineers relative to the stated baseline, generating them at 75% of the expected rate (equivalently, under-representation by a factor of about 1.33×).
5. A GI tag restricts who may use the regional name ("Pattachitra") on goods sold as authentic; it does not, and was never designed to, restrict who may learn or reproduce the visual style itself, because style is treated as an unprotectable idea rather than protectable expression under both GI and copyright frameworks — so a model can absorb and reproduce the style without ever violating the naming restriction the GI tag actually enforces.
6. First, the idea-expression dichotomy means infringement was never about pixel-identical copying — reproducing distinctive expression (composition, motif arrangement, style) in a different rendering can still be a legal and ethical problem even with zero pixel-for-pixel matches. Second, "no ethical issue" ignores the economic-displacement harm documented in the GI-tag discussion: even a stylistically faithful but non-infringing image competes directly with the livelihoods of the artisan community whose collective, uncredited work trained the model, which is a real harm independent of whether any single copyright test is triggered.

Think About It

Think about this: How would you explain ai art ethics: ownership, bias, and indian cultural context to a friend who has never seen a computer? What real-world analogy would you use? Imagine you had to build a system using these concepts — what would be your first step? Try this: before moving on, write down three things you learned and one question you still have.

Practice Exercises

Now it is time to practice! Complete these challenges to solidify your understanding:

• Exercise 1: Write a short program that demonstrates the core concept from this chapter. Test it with at least 3 different inputs.
• Exercise 2: Find a real-world example where ai art ethics: ownership, bias, and indian cultural context is used in an Indian company (like TCS, Infosys, Flipkart, or ISRO). Write a paragraph explaining the connection.
• Exercise 3: Create a mind-map connecting ai art ethics: ownership, bias, and indian cultural context to at least 3 other topics you have studied.

Key Takeaways — Summary and Recap

Let us recap what we covered: the core ideas behind ai art ethics: ownership, bias, and indian cultural context, how they connect to real-world applications, and why they matter for your journey in computer science. Remember these key points as you move forward. For competitive exam preparation (CBSE, JEE, BITSAT), focus on understanding the WHY behind each concept, not just the WHAT.