THE RUNT OF THE LITTER JUST INHERITED THE IP ARGUMENT

The runt and the inheritance

Vector AI was the runt of the generative litter. Turns out it was quietly growing teeth.

Whilst the rest of the gen-AI pack—video, image, audio—was fighting to be top dog in an increasingly crowded space, the vector world is becoming more acutely capable of producing content that can be safe to use, become human-made, and is non-destructive by nature.

That's not a small claim. And it's one that the broader conversation about AI-generated content and intellectual property has almost entirely missed—not because the evidence isn't there, but because vector has spent the last two years being the tool people reached for last.

That's changing. The family has expanded, the tools have matured, and sitting underneath the noise about image rights, video scraping, and model training ethics is a quieter and more interesting argument: that the vector pipeline is the one place in generative AI where craft and legal defensibility are pointing in the same direction.

1. Why vector is structurally different

2. The copyright question

3. The training ethics caveat

4. The landscape—vector AI in 2026

5. The one to watch (Figma + Weavy)

6. The fragile audit trail

1 / Why vector is structurally different

Let's be precise about something the industry has been deliberately vague on.

The majority of generative AI models are not creative tools. They are personalised stock generators. Sophisticated ones—capable of extraordinary aesthetic output, responsive to reference, trainable to a style, scalable to a brief. But stock generators nonetheless. The creativity is in the curation and direction. The making is done by the model.

Yes, some node-based workflows sit editor nodes between input and output—between the prompt, the LoRA, the reference image and the final result. But these are shallow interventions by the standards of traditional creative tooling. The depth simply isn't there. You are directing output, not authoring form.

This matters because the two things the industry conflates—creative production and personalised stock generation—have completely different relationships to intellectual property, authorship, and ownership. If you generate an image, a video, or an audio track and ship it largely as the model produced it, you haven't made something. You've retrieved something. The distinction is not philosophical. It has legal and professional consequences that most practitioners are only beginning to feel.

Vector is the exception—and the exception is structural, not aesthetic.

A vector file is not a picture. It's a set of instructions. Paths, anchor points, bezier handles, shape relationships—all described mathematically, all editable at the level of individual nodes. When an AI model generates a vector asset, it produces a starting set of those instructions. Plausible geometry. A rough formal argument for what the shape could be.

What it isn't is finished.

To take an AI-generated vector from output to production asset, a designer works at path level. Redraws curves. Restructures anchor relationships. Rebuilds shapes the model approximated but didn't resolve. That process isn't touching up someone else's work—it's rewriting the mathematical description of the form. Move enough handles, restructure enough paths, and you haven't filtered existing content. You've authored new geometry.

Here's the condition the rest of this article depends on: that work has to happen. An unedited AI vector output has no more legal defensibility than any other unedited AI output. The pipeline doesn't protect you. The craft does. If you're generating and shipping without editing, you're not using a creative tool—you're using a stock generator with a different file format.

The single-prompt crowd isn't using these tools as creative instruments. They're using them as clients. Skilled clients, sometimes. But clients—bypassing the creative process they're emulating rather than participating in it.

The rest of us are doing something genuinely different. And it's worth understanding exactly what that is.

2 / The copyright question

Here is what the law currently says about AI-generated content and copyright: not enough.

Not through negligence—through deliberate caution. Legislators and courts across the US, UK, and EU are moving carefully in a space where the technology is outpacing the framework, and the decisions being made now will set precedents that last decades. Nobody wants to get it wrong quickly.

What has emerged from the cases decided so far is a principle rather than a rule. Copyright protection requires sufficient human creative expression. The work must reflect human authorship in a meaningful way. AI-generated content, in isolation, doesn't qualify.

What nobody has defined is the percentage. There is no threshold. No point at which a work crosses from AI-generated to human-authored. "Sufficient" is doing all the work, and courts have been deliberately vague about what it means—partly because any fixed number would be immediately gamed, and partly because creative authorship has never been a quantitative question.

This ambiguity is further complicated by the fact that the legal framework is not uniform. It varies by jurisdiction—and in some cases, significantly.

In the United States, federal copyright law applies nationally but judicial interpretation varies by circuit. The Copyright Office guidance following Zarya of the Dawn established that purely AI-generated content is not copyrightable, and that human-authored elements within or on top of AI output may be—but this is administrative guidance, not legislation, and no federal circuit court has ruled definitively on AI authorship. The cases building on it are still working through the system.

The United Kingdom sits in an unusual position. The Copyright, Designs and Patents Act 1988 contains a provision—Section 9(3)—that offers protection for computer-generated works without a human author, crediting authorship to "the person by whom the arrangements necessary for the creation of the work are undertaken." Written in 1988 for a completely different technological context, its application to generative AI is untested and actively contested. It theoretically offers more protection than the US position, but nobody has tested that theory in court against a generative AI output.

The European Union is furthest behind legislatively. The AI Act addresses risk classification and governance but does not resolve copyright ownership of AI outputs directly. Copyright implications are being worked out in parallel through existing directives, and member state implementation will vary. The EU position is, for now, the least settled of the three.

The practical consequence for anyone working commercially—and most commercial creative work crosses jurisdictions—is that a piece of work that clears the copyright bar in one territory may not in another. This is not a theoretical problem. It is a live one.

Which is precisely where the vector argument becomes most interesting.

Path-level editing is not an aesthetic intervention. It is a formal one. When a designer restructures the geometry of an AI-generated vector—redraws a curve, rebuilds a shape, resolves what the model approximated—they are making decisions that have no pixel equivalent. Each anchor point moved is a discrete authorial choice about the mathematical description of a form. Cumulatively, those choices constitute something courts can actually evaluate as human creative expression, because they are traceable, specific, and non-trivial.

Compare that to the standard post-processing workflow for a generated image. Colour grade, crop, composite. These are legitimate creative decisions—but they operate on top of content the model produced. The underlying pixel data is intact. The human contribution is directorial. In vector, the human contribution is structural. The difference is not one of degree. It is one of kind.

This doesn't guarantee copyright protection. Nothing in this space comes with guarantees. But the vector pipeline is uniquely positioned to generate the kind of evidence that a court evaluating "sufficient human creative expression" would find meaningful—because the work itself is the record. Every path edit is documented in the file. The transformation is legible in the geometry.

Across every jurisdiction currently grappling with this question—however differently they frame it—the more legible your human authorship is in the work itself, the stronger your position. Path-level transformation is the most defensible evidence you can produce regardless of which legal framework is evaluating it.

That's not a legal strategy. It's a natural consequence of how the format works.

References for this section listed at the end of the article. Correct as of April 2026.

3 / The training ethics caveat

None of the above resolves the question of how these models were trained.

The output argument and the training argument are separate conversations, and conflating them—which the industry does constantly—serves nobody well. You can make a defensible claim of human authorship on a heavily reworked vector asset and still be working with a model trained on data it had no right to use. One doesn't cancel the other.

The honest position is this: training data provenance across most of the vector AI landscape is opaque at best and contested at worst. With the exception of Adobe Firefly—trained on licensed Adobe Stock content, openly licensed material, and public domain—most tools in this space use mixed or undisclosed training data. Some are more transparent than others. None outside Adobe offer contractual IP indemnification.

That matters. It means the ethical question about whose creative work trained these models remains open, regardless of what you do with the output. Creatives who are rightly protective of their own work being scraped without consent should hold that concern consistently—including when it applies to the tools they're reaching for.

What the training ethics argument doesn't do is determine the status of your output. Those are upstream and downstream of each other. Acknowledge the former. Work carefully on the latter. Don't use one to dismiss the other.

The vector pipeline gives you a meaningful path to ownership of what you make. It doesn't give you a clean conscience about where the model learned to make it. Both things are true simultaneously—and any serious practitioner in this space should be able to hold them both.

4 / The landscape—vector AI in 2026

The vector AI space has expanded considerably in the last twelve months. Understanding it requires separating what these tools actually do from what they claim to do—because the categories matter, and conflating them leads to the wrong tool for the wrong job.

Four distinct categories exist: native vector generation, raster-to-vector conversion, multi-angle perspective generation, and SVG code generation. Each sits differently on the commercial safety spectrum and demands a different kind of human intervention to produce defensible output.

Native vector generation

Tools that generate true SVG paths from a text prompt or image input. The only category producing genuine starting geometry for path-level authorship.

Recraft (V4) recraft.ai | Paid plans from ~$15/month. Free tier has no commercial licence. The most fully featured native vector generator currently available. Produces clean paths, organised layers, and genuine SVG output across a wide stylistic range. Brand style system allows reference-based consistency across a project. Also exports Lottie. Training data provenance is mixed and not fully disclosed—commercial licence on paid plans, no IP indemnification. The strongest pure generation tool in the space, with the caveat that provenance transparency lags behind its output quality.

Adobe Illustrator (Firefly) adobe.com/products/illustrator | From $22.99/month single-app. The only tool in this entire breakdown offering contractual IP indemnification—available to enterprise customers on qualifying plans. Trained on licensed Adobe Stock content, openly licensed material, and public domain. C2PA content credentials embedded automatically. Stylistic range narrower than Recraft. The right answer when commercial provenance is a procurement or legal requirement, not just a preference.

QuiverAI (Arrow 1.0) quiver.ai | Pricing not yet publicly confirmed. A16z-backed, launched February 2026. The differentiator is a natural language editing loop—generate, then refine through conversation with the model rather than through path editing alone. Vector-native from the ground up. Too early to make strong commercial safety claims. Training data provenance undisclosed. Worth watching closely—the trajectory is strong and the backing is serious.

Raster-to-vector conversion

Tools that trace or interpret existing raster images and produce editable vector output. Not generation—translation. The starting point is your image, not a prompt.

Figma Vectorize figma.com | Included in paid Figma plans with AI credits enabled. Launched February 2026. Converts raster images—sketches, scanned drawings, PNG icons—into editable vector layers directly inside Figma. The zero tool-hop is the USP: sketch to vector to design system without leaving the canvas. Not a generation tool. Training data provenance undisclosed. No indemnification. The right tool when your source already exists and your destination is Figma.

Adobe Illustrator Image Trace / Photoshop Included in existing CC subscription. Zero additional cost, zero credits. The tools people overlook because they've been there for years. Still the most reliable raster-to-vector option for complex or photographic source material. No AI credits, no provenance questions, no commercial safety ambiguity—you own your source, you own the trace. Underused precisely because they're not new.

Adobe Express adobe.com/express | Free tier available. Included in CC plans. Image-to-SVG conversion for simple logos and icons. Firefly-powered where AI is involved, so the same provenance and indemnification terms apply. Limited in scope—not a generation tool and not suited to complex source material. Useful for non-designer collaborators already in the Adobe ecosystem.

Multi-angle perspective generation

A category of one, currently. Not generation from scratch—transformation of existing vector art into multiple viewpoints.

Adobe Turntable adobe.com/products/illustrator | Included in Illustrator subscription. Generally available from March 2026. Takes a single flat 2D vector illustration and generates up to 74 editable views—full 360-degree rotation and vertical tilt. Each angle remains a fully editable vector. Not a true 3D model—it performs 2D-to-2D image translation using a raster intermediary internally, which is why OBJ export isn't possible. Works best on illustrations with clear outlines and recognisable real-world geometry. Abstract inputs hallucinate. The After Effects pipeline is direct. For character turnarounds and multi-angle asset production it saves days, not hours. Covered by Firefly IP indemnification on qualifying plans.

SVG code generation

LLMs and vibe coding tools generating vector as code rather than as image output. Structurally the cleanest IP position in the breakdown—and the most overlooked.

Claude, GPT-4o, Cursor and equivalents claude.ai / chat.openai.com / cursor.com | Marginal cost only within existing subscriptions. Increasingly capable at generating SVG code directly from natural language prompts. The IP argument here is structurally distinct from image model output—you are generating syntax, not reproducing training images. No credits, no provenance questions in the image-model sense, fully transparent output you can read, edit, and version control. Particularly strong for icons, UI assets, geometric and systematic illustration, data visualisation, and animated or responsive SVG. Artistic range for complex organic illustration is still limited. The right tool when the output lives in a code pipeline or when you want complete structural transparency from the start.

A note on aggregator platforms

Krea (krea.ai) deserves specific mention as both a platform in its own right and the primary aggregator context in which several of the above models appear.

As a standalone platform, Krea offers real-time image generation via its Realtime Canvas—sub-50 millisecond generative feedback as you draw or prompt, which is genuinely unmatched for ideation speed. It aggregates 64+ models including Flux, Ideogram, Veo, Runway, and hosts the Quiver-powered vectoriser. Commercial licence from the Basic plan ($10/month). No indemnification.

The broader aggregator point applies here and throughout: when AI models are accessed through aggregator platforms rather than natively, they rarely carry their full feature set. Style systems, output format options, parameter controls, and model version access are frequently stripped back or absent. Recraft accessed through Leonardo is not Recraft. QuiverAI accessed through Krea is not QuiverAI. Aggregators are useful for discovery and cross-model comparison. For production use of a specific model, go native.

5 / The one to watch—Figma Weave

weave.figma.com | Standalone product, separate billing from Figma Design.

Formerly Weavy, acquired by Figma in October 2025. A node-based AI media pipeline for image, video, animation, motion design, and VFX generation—not specifically a vector tool today, but worth examining here because the architectural logic is directly relevant to where the vector pipeline is heading.

Figma is predominantly a vector environment by nature. It's where paths live, where design systems are built, where the output of the vector workflow ultimately needs to arrive. Weave brings a node-based generative AI pipeline into the same organisation—non-destructive by architecture, multi-model by design, with outputs that branch and feed forward rather than terminate.

Put those two things together properly and you have something that doesn't currently exist in one place: generative AI input, node-based non-destructive iteration, native vector output, path-level editability, and a design system environment that the finished assets flow directly into. The entire argument this article makes—that human authorship at path level is the mechanism that shifts the IP position—gets a purpose-built home.

Every other tool in this breakdown requires you to leave somewhere to go somewhere else. Recraft to Illustrator. QuiverAI through Krea into your tool chain. Figma Vectorize back into the design file. A fully integrated Figma Weave closes that loop entirely.

Full Figma platform integration is announced for later in 2026. The execution will determine whether the promise holds. But as a structural proposition, it's the most coherent vision of what a serious vector AI pipeline could look like—and the combination most worth watching.

6 / The fragile audit trail

Adobe embeds C2PA content credentials into every Firefly-generated asset. The metadata records the model used, the prompt, the edit history. It's presented as provenance—a chain of custody from generation to delivery.

It's a good idea with a significant structural problem.

The credentials live in metadata. And metadata doesn't survive a real production pipeline.

Screenshot the output. Export via a tool that doesn't support C2PA—which is most of them. Convert the file format. Flatten the layers. Upload to a platform that strips metadata on ingest—which is almost every CMS and social platform by default. Any of these, often all of them in sequence, and the credentials are gone. Silently, without warning, leaving no trace of their absence.

This matters for two reasons that point in opposite directions.

The first is that the audit trail Adobe describes is fragile by design—not tamper-evident in any meaningful sense, but metadata that either survives a workflow or doesn't. For anyone relying on C2PA as a chain-of-custody mechanism, the honest assessment is that it will fail before the asset reaches its final destination in most real-world production contexts.

The second cuts the other way. The absence of C2PA credentials proves nothing. You cannot distinguish between "this file passed through Figma and credentials weren't preserved" and "someone deliberately stripped them." Intentional removal is undetectable. Which means the system creates a false sense of security in both directions—for those trying to establish provenance and for those trying to identify when it's been obscured.

The CAI and C2PA working group are developing harder binding mechanisms—cryptographic signing tied to file content itself rather than attached metadata. That's the right direction. It's not there yet.

For now, the most defensible provenance record isn't an embedded credential. It's your own documented process. Version history. Project files. Screenshots of the generation pipeline and the editing stages that followed. The file itself—because in vector, as this article has argued, the transformation is legible in the geometry. The path edits are the evidence.

C2PA is better than nothing at point of generation. It is not a reliable chain of custody across a production pipeline. Don't treat it as one.

Glossary

C2PA—Content Credentials / Coalition for Content Provenance and Authenticity. What it is, what it's supposed to do.

IP indemnification—what it actually means in contractual terms, distinct from a commercial licence.

Native vector generation—as used in this article, distinct from raster-to-vector conversion.

Training data provenance—what it means and why it matters differently from output rights.

LoRA—for readers who know gen-AI but not the technical side.

Generative credits—platform-specific credit systems, since the article references them across multiple tools.

Aggregator platform—as distinct from a native tool, since the article makes a specific argument about the difference.

Possibly dotLottie / Lottie given Recraft exports it—but that might be one too many given Lottie isn't in the article proper.

References

Zarya of the Dawn (US Copyright Office, 2023) Kris Kashtanova / Midjourney case. The Copyright Office ruled that AI-generated images within the graphic novel were not copyrightable, but that the human-authored text and arrangement were. The foundational US case on partial human authorship of AI-assisted works. copyright.gov/docs/zarya-of-the-dawn.pdf

US Copyright Office guidance on AI-generated works (2023–2024) A series of guidance documents establishing that works produced solely by AI without human creative control are not eligible for copyright registration. Confirms the "sufficient human creative expression" standard without defining a threshold. copyright.gov/ai

Thaler v. Perlmutter (US District Court, DC, 2023) Stephen Thaler's DABUS case. Court ruled that copyright requires human authorship and that an AI cannot be listed as an author. Reinforces the human creative expression requirement.

UK Copyright, Designs and Patents Act 1988—Section 9(3) The provision covering computer-generated works. Authorship attributed to "the person by whom the arrangements necessary for the creation of the work are undertaken." Drafted for pre-generative AI context. Application to LLM and diffusion model outputs untested in UK courts as of 2026. legislation.gov.uk/ukpga/1988/48

UK Intellectual Property Office—AI and IP consultation The IPO has consulted on whether the Section 9(3) provision remains fit for purpose in the context of generative AI. No legislative change enacted as of 2026. gov.uk/government/consultations/artificial-intelligence-and-ip

EU AI Act (Regulation 2024/1689) Addresses risk classification and governance of AI systems. Does not resolve copyright ownership of AI-generated outputs directly. Copyright implications deferred to existing EU copyright directives and member state implementation. eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689

EU Copyright in the Digital Single Market Directive (2019/790) The existing EU framework most relevant to AI training data and output. Article 4 text and data mining exception and its interaction with generative AI outputs is actively debated across member states. eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32019L0790