Three years into the generative boom, the question for working artists is no longer whether AI can produce a passable painting, song, or short story — it clearly can. The question is what survives, who gets paid, and which tools are actually worth the subscription. In 2026, the answers are starting to harden into law, into union contract language, and into numbers you can quote in a client meeting. This is the working-creative field report on what changed, what cost it cost you, and what to actually use.
Five things shifted between January 2024 and now. The US Copyright Office’s January 2025 Part 2 report ended the “is a prompt enough?” debate by saying no. The D.C. Circuit confirmed in March 2025 that purely AI-generated work gets no copyright at all, and the Supreme Court declined to revisit that ruling in March 2026. The Writers Guild of America’s 2023 contract terms are now in force, SAG-AFTRA’s 2025 Commercials contract added the strongest AI clauses any union has negotiated, and Adobe is the only major model vendor offering contractual IP indemnification. The Anthropic Bartz settlement, finalized July 20, 2026, set the first hard price for AI training data on copyrighted text: roughly $3,000 per work. And CISAC’s 2025 Global Collections report put a number on the creator-income risk: generative AI could divert 24% of music creator royalties — €4 billion — by 2028 if licensing frameworks are not built.
What follows is a working-creative’s field guide to the post-2024 landscape: which tools actually ship and at what price, what the new copyright rules mean in practice, what the union contracts changed and what they did not, where the money is and is not flowing, and the 12 questions you should answer before starting your next paid project.
The 2026 creative tool landscape: what is actually shipping and what costs what
The tool landscape has consolidated around four domains. None of them is cheap at the professional tier, and none of them is indemnified unless you are specifically on an Adobe Enterprise plan.
Image generation. Three tools are doing real work. Midjourney V7 (alpha April 4, 2025, fully rolled out by mid-year, ~40% faster) is still the aesthetic ceiling — better bodies, better hands, more cinematic composition, with the caveat that Midjourney offers no IP indemnification and its training-data provenance is opaque. The independent Midjourney V7 review lines up with that assessment. Adobe Firefly Image Model 5 (released October 2025, GA 2026, 4MP native resolution) is the only mainstream image generator trained exclusively on licensed Adobe Stock content and public-domain material; the paid tiers include contractual IP indemnification, with Enterprise VIP covering up to $10,000 per output. Stable Diffusion 3.5 remains the open-source option for self-hosted pipelines, but the same no-indemnification problem applies. GPT-image-1 (OpenAI’s native image model, late 2024) is the most accessible and the least style-distinctive, but it is now wired into ChatGPT, the API, and Microsoft 365 Copilot. For a side-by-side assessment of the current image-generation options, see our Best AI Image Generators We’ve Tested for 2026.
Video generation. OpenAI Sora 2 (released September 30, 2025) is the first system that OpenAI is willing to call “the GPT-3.5 moment for video.” It ships with native synchronized audio (dialogue, sound effects, ambient noise), meaningfully better physics, a Cameo feature for inserting real people after a short verification recording, and Storyboard for multi-prompt timelines. Pro users get 1080p at 20-25 seconds. Runway Gen-4 and Google Veo 3 are the two production-grade competitors. The complication: Sora 2 has no IP indemnification, and the December 11, 2025 Disney-OpenAI $1 billion partnership means any new Sora generation may now incorporate elements from over 200 Disney, Pixar, Marvel, and Lucasfilm characters — which makes any non-Disney-branded Sora output a careful IP review situation. Pricing sits at roughly $0.10 per second of output.
Audio generation. Suno v5 (September 23, 2025) is the dominant music generator, with a v5.5 update that added Voices, Custom Models (upload six of your own tracks to train a personalized version), and My Taste for mood-based prompting. The internal ELO jump from v4.5+ to v5 was 1,208 to 1,293 — measurable progress, but a Verge review found the vocals still lack the imperfections that carry emotional weight, and Suno quietly dropped its “emotionally rich / human-like emotional depth” marketing language. ElevenLabs v3 (alpha June 2025, GA March 2026) is the expressive-TTS story of the year: inline audio tags like [whispers], [sighs], [laughs], [shouts] let you direct emotion in the script, with 70+ language support and a multi-speaker dialogue mode. Udio remains the main alternative for music.
Text generation. Claude, GPT-5, and Gemini are the three frontier models writing-room professionals actually use. The WGA’s 2023 contract is the only union framework in the world that explicitly governs how screenwriters can and cannot use them. The multimodal landscape in 2026 blurs the boundaries: image, audio, and text models increasingly ship as one integrated stack rather than separate tools.
What actually changed for working creatives since 2024
The most-cited single number in the Stanford HAI AI Index 2025 for this conversation is the Anthropic team’s analysis of Claude.AI conversation data: arts, design, entertainment, sports, and media occupations account for 10.3% of all AI interactions, second only to computer and mathematical occupations at 37.2%. That is the demand side. On the supply side, the same Index tracks a 26% decline in illustrator demand and a 36% decline in translator demand in the past 18 months, citing sector surveys.
Where the displacement is showing up fastest: stock photography (commission volume at Getty and Shutterstock is down meaningfully year-over-year, with both companies pivoting toward AI-image distribution), concept art for film and game (mid-tier studios downsized in late 2024 and early 2025), book-cover illustration (Penguin Random House, HarperCollins, and several mid-tier imprints have visibly reduced the number of human illustrators per title), and session-music work for ad and UGC production (Suno v5 outputs are now meeting “good enough for the brief” bars that used to require a real session player). The workflow side is shifting too: the AI workflow automation tools that 12-18 months ago were sold to ops teams are now landing in creative-team stacks. The CISAC 2025 Global Collections Report shows that total music royalties are still growing (€12.59 billion in 2024, +7.2% year over year) but the share going to human creators is at structural risk over the next 24 months — see the CISAC/PMP Strategy study for the projected 24% music-royalty diversion by 2028.
What is not happening yet, contrary to the 2024 hype: working novelists are not being replaced. The WGA contract drew a hard line on scripted material (more on that below), and the long-form book market is still human-authored. Voice actors are not being replaced, but they are being re-priced by the SAG-AFTRA digital-replica framework. The 200,000-customer-support-agents productivity study cited in the AI Index — 10% to 45% productivity gains, depending on the role — is mostly white-collar service work, not creative work, but the broader pattern is that the lower the proportion of judgment-and-craft in the deliverable, the faster the substitution.
Copyright, finally: the 2025-2026 rules you can actually rely on
Three legal milestones from 2025 to early 2026 are now the foundation. Read them once, and you will know what you can and cannot protect.
1. Thaler v. Perlmutter (D.C. Circuit, March 2025; cert denied March 2026). Stephen Thaler wanted copyright on an image his AI system, DABUS, had produced with zero human input. The D.C. Circuit said no: copyright requires human authorship. The Supreme Court declined to take the case on March 2, 2026, leaving the rule in place as settled law. The opinion is careful to note that it does not prevent copyright on work made with AI assistance — only on work where there was no meaningful human creative input.
2. US Copyright Office Part 2 report (January 29, 2025). After reviewing more than 10,000 public comments, the Copyright Office staked out three rules: purely AI-generated material is not copyrightable; prompting alone, however iterative, is not sufficient human authorship; and copyright protects only the elements where human authorship shows up in the selection, coordination, arrangement, or modification of AI output. The Zarya of the Dawn graphic novel decision (February 2023) is still the canonical example: text and arrangement were protected, Midjourney images were not.
3. Anthropic Bartz settlement (final approval July 20, 2026). Judge Araceli Martínez-Olguín approved the largest copyright class-action settlement in US history at $1.5 billion. The underlying June 2025 ruling was a split decision: training on lawfully-acquired books was held to be fair use (“spectacularly transformative,” Judge William Alsup wrote in a related proceeding), but Anthropic’s downloading of books from pirate libraries (LibGen, PiLiMi) was not. The settlement requires Anthropic to destroy all the downloaded works and pays approximately $3,000 per work to the class of authors. About 91% of authors and publishers filed claims; 350 opted out, and 28 of those opt-outs have filed a separate copyright suit requesting a jury trial. The big open question — the New York Times v. OpenAI case, which widened in July 2026 when the Times, the Daily News, the Center for Investigative Reporting, and Ziff Davis asked for sanctions over an alleged internal “Project Giraffe” detection system — is still active.
What this means in practice. If you use Midjourney or Sora 2 for a base image or clip, then make significant human-authored changes in Photoshop, After Effects, or Resolve, the human-authored parts of the final file are protectable; the raw AI portions are not. If you sell an unmodified AI output, you are selling into the public domain — anyone can copy and resell it. If you train a model on copyrighted material without a license, you are exposed to the same kind of claim the Bartz authors just settled. The privacy and data-collection questions raised by the AI tool logging story are now the second front of the same fight. The rules are narrow but they are finally clear.
The union wins: WGA 2023 and SAG-AFTRA 2025 AI clauses, what they actually do
Two union contracts now define the floor for AI use in American scripted and on-camera work. Both are short, both are enforceable, and both are already in production.
Writers Guild of America 2023 MBA (effective September 25, 2023 through May 1, 2026). The four operative AI clauses, from the WGA’s own summary: (1) AI cannot write or rewrite literary material, and AI-generated material is not “source material” under the MBA — meaning it cannot be used to undermine a writer’s credit or separated rights. (2) A writer can choose to use AI when performing writing services if the company consents, but the company cannot require a writer to use AI. (3) The company must disclose to the writer if any materials given to the writer have been generated by AI or incorporate AI-generated material. (4) The WGA reserves the right to assert that exploitation of writers’ material to train AI is prohibited by the MBA or other law. In plain English: a studio cannot hand a writer an AI draft and pay them a rewrite fee, and a writer who uses AI in their own process must have the company’s consent and must be told when the company is feeding them AI-generated source.
SAG-AFTRA 2023 TV/Theatrical and 2025 Commercials contracts. The 2023 deal was the first to bake AI protections into the core agreement, defining employment-based and independently-created digital replicas and requiring explicit, signed consent before a replica is used in scenes the actor did not perform. The 2025 Commercials contract, effective April 1, 2025, is the strictest AI framework in any major union contract: use of a digital replica to generate a performance triggers a 1.5x session fee plus full applicable holding and use fees, and digital replicas cannot be used to circumvent the use of background actors. California AB 2602 reinforces both contracts by making vague “any and all AI uses” language in talent contracts unenforceable. The union wins are binding in any US scripted production, any US commercial, and (via the 2023 deal) most streaming series shot under a SAG-AFTRA signatory.
What the contracts do not do: they do not stop voice actors from being cloned with their own consent at renegotiated rates, and they do not stop a non-union freelancer from being undercut by a studio that ships AI-narrated ads in markets where SAG-AFTRA has no jurisdiction. The 2025 Commercials digital-replica framework also does not require that the consent specify the project in advance — a union performer can sign a consent that allows the use in multiple projects under the same compensation scheme, which is why a number of high-profile voice actors have been signing those blanket deals. The structural fix for the unorganized and non-US worker is still not in place.
Where the money is flowing: royalty data, indemnification, and the creator-payout gap
Three data points describe the 2026 money picture. The first is the good news: CISAC’s 2025 Global Collections Report found that global music creator royalties grew 7.2% in 2024 to a record €12.59 billion, with digital growth +9.6% (live and background +9.6%, audiovisual +5.2%, visual arts +1.9%). The music creator pie is bigger than it has ever been.
The second is the structural risk. The same CISAC, working with PMP Strategy, projects that unlicensed generative AI could divert up to 25% of creator royalties (€8.5 billion annually) by 2028 if left unregulated, with the music segment alone at 24% / €4 billion and the audiovisual segment at 21% / €4.5 billion. The CISAC report is not a vendor projection; it is the international federation of 228 author societies in 111 countries, working from actual collections data. The “60% of music streaming platform revenues” projection for 2028 — meaning AI-generated music on the same platforms where human creators currently earn their royalties — is the headline number. The implementation question is whether the streaming platforms will require AI music to be labeled, royalty-poolable, and excluded from royalty-bearing consumption. None of the major platforms has committed to that yet.
The third is the new pricing band for AI training. The Anthropic Bartz settlement, at approximately $3,000 per work for 506,194 works on the class Works List, is the first hard price for AI training on copyrighted text. That is four times the $750 statutory minimum for ordinary infringement and 15 times the $200 statutory minimum for innocent infringement. It is also an order of magnitude more than the per-work payment proposed in the Google Books settlement that the Southern District of New York rejected in 2011 for releasing future claims. The Bartz number is the new floor for class-action settlement negotiations in any text-training case. The 28 authors who opted out are testing whether they can get more in a jury trial.
Image and video are still unpriced. Adobe Firefly is the only major model vendor with explicit IP indemnification on paid plans — Enterprise VIP covers up to $10,000 per output, with full legal defense. Midjourney, Sora 2, Runway, Veo, Suno, ElevenLabs, Stable Diffusion, and Flux all explicitly disclaim indemnification in their terms. The capital side of the same picture is that OpenAI, Anthropic, and the rest of the frontier-model tier are valued at a scale that makes the training-data settlement line a rounding error — Anthropic paid $1.5B out of a multi-hundred-billion-dollar valuation. The structural fix has to be licensing frameworks, not litigation. Operator-style AI agents are the next layer down the same stack. The practical effect: if a freelancer delivers a Midjourney image to a client and the client gets sued for infringement, the freelancer is on their own.
The tools actually worth paying for in 2026, by creative discipline
Stop paying for tools you are not using at the top of your license. Stop paying for the bleeding edge of the bleeding edge. Here is the working-creative short list, by discipline.
Commercially safe client work (image). Adobe Firefly Image Model 5, on a paid plan. The indemnification is the only one on the market, and it converts AI from a liability into an asset on your client invoice. The aesthetic ceiling is below Midjourney V7 on cinematic composition, but that gap closes every release cycle. Use Firefly for the client deliverable; keep Midjourney for mood boards and concept exploration.
Peak aesthetic image work. Midjourney V7 for any deliverable where the image itself is the value (book covers, key art, indie game splash screens where the client understands the indemnification is theirs to buy). Treat the prompt-to-output as a draft; take the output into Photoshop for the human-authored layer (selection, masking, retouching, color grading) that will be the copyrightable part of the file.
Cinematic video. Sora 2 if you are doing pre-visualization, concept shorts, or short-form social and you have a clearance workflow for the December 2025 Disney-deal universe of characters. Runway Gen-4 if you want a more conservative training-data profile. Veo 3 if you are on the Google Cloud stack. None of them indemnifies. The SAG-AFTRA 2025 Commercials 1.5x session fee applies if any part of the deliverable uses a digital replica of a union performer — budget for it in the quote.
Voice and audiobook. ElevenLabs v3 for the text-to-speech work where the new audio tags (e.g., [whispers], [laughs], [sighs]) actually carry the performance. Use the Professional Voice Clone flow with signed consent and a contract that spells out the use. If the client is a publisher, the union consent clauses in the SAG-AFTRA 2025 Commercials contract will be on the publisher’s standard contract anyway.
Music for ad, UGC, podcast, indie game. Suno v5.5 with Custom Models (upload six of your own tracks to train a personalized version) for the in-house score work. Udio as a backup. Reserve paid session musicians for the deliverable that ships to a broadcast partner with a music-supervision requirement, which most still are. Suno’s internal ELO jump from v4.5+ to v5 (1,208 → 1,293) is real progress, and the cleaner mixes and more complex arrangements are noticeable, but the model still serves up vocals that sound like Mumford & Sons by default and refuses to reproduce lo-fi or “off” performance even when you prompt for it.
Writing assistance. Claude or GPT-5 for outlining, research synthesis, and first-draft ideation. The WGA contract requires you to disclose AI use to the company, and the company must consent. If you are writing under a WGA contract, the consent conversation has to happen before you open the model. If you are writing for a non-union client, check the contract for an AI clause — most 2025-and-later publishing contracts have one. The craft side of the same question is how to actually prompt well when the model is the instrument.
What survives: the human-AI workflows that are working in 2026
The workflows that are paying in 2026 are not "AI does the work." They are "AI drafts, you direct, you sign, you ship." The director-orchestrated tools — Sora 2 Storyboard, Suno v5 Custom Models, Midjourney V7 with consistent-character features — are the ones that turn the model from a vending machine into an instrument.
Five workflows that are working, with caveats:
1. Director-orchestrated video. Sora 2's Storyboard feature lets a director chain multiple prompts into a multi-clip timeline, drag and drop transitions, and set pacing. The Cameo feature (you record a short verification clip, the model can then insert you into any generated scene) is the first digital-replica tool that has consent and compensation baked in. Use the Storyboard for the rough cut, use Cameo for the personal-replica shots, and finish in a real NLE for the human-authored color and edit.
2. Concept art for indie film and game. Firefly base + Photoshop refinement, signed by the human artist. The Firefly indemnification covers the AI portion; the Photoshop refinement creates the human-authored layer that is protectable. This is now the standard workflow for small-to-mid-budget game and indie film.
3. Book-cover illustration. Firefly for the base, Photoshop for the human layer, signed off by the human illustrator. The publisher buys the rights to the human-authored layer; the AI base is not protectable but the integrated work is. This is a workable legal position as long as the publisher is willing to accept it (most of the major imprints are, with disclosure).
4. Composer + AI scratch tracks. Use Suno v5.5 to generate the reference tracks for client review, then record the final with the human composer and session musicians. The scratch tracks are not the deliverable; they are the rough cut. This compresses the client-iteration cycle from weeks to days, and the human composer still gets the final fee.
5. Writer + LLM outlining and research. Use Claude or GPT-5 for the research synthesis and structural outline, then write the draft yourself. The WGA disclosure clause is satisfied if you tell the company you used AI for outlining. The actual literary material is still yours. Holly Herndon and Mat Dryhurst's Spawning and Have I Been Trained work is the more ambitious version: build the consent layer into the data itself, with opt-out and revenue-sharing for the creators whose work trains the model. For working teams scaling this further, the OpenAI Agents SDK and related production-agent tooling are the path from single-creator workflows to multi-creator pipelines.
Your decision tree: 12 questions a working creative should answer before the next project
Print this out. Pin it above your desk. Work through it before you start any paid project that uses AI.
- Will I be paid for the deliverable or for the rights? (If for the rights, the AI portion of the deliverable is not protectable — price accordingly.)
- Who indemnifies me if my output is challenged? (If the answer is "nobody," price the risk in or use a Firefly-on-paid-plan pipeline.)
- Am I required to disclose AI use under a union contract, client contract, or platform ToS? (WGA: yes, with company consent. SAG-AFTRA 2025 Commercials: yes for digital replicas. Most 2025-and-later publishing contracts: yes.)
- Did I license the training data behind my tool, or am I inheriting its IP risk? (If unknown, the answer is "I am inheriting the risk.")
- What fraction of the final deliverable is AI-generated, and what is the human-touched layer that will be the protectable part? (Aim for 30-50% human-authored, even if it is a Photoshop layer, a color grade, a sound mix.)
- Does the client want exclusive rights, or non-exclusive? (Exclusive rights on AI portions is impossible — they are public domain — and any contract that grants exclusive rights on an AI deliverable is unenforceable on the AI portion.)
- If the deliverable is voice, did the talent sign a digital-replica consent that names the project, the platforms, and the duration? (SAG-AFTRA 2025 Commercials standard.)
- If the deliverable is music, will the platform (Spotify, Apple, TikTok, YouTube) require AI-content labeling under their 2025 ToS updates? (Most have rolled this out in 2026; check before delivery.)
- If the deliverable is video, does any of the Sora 2 / Runway / Veo output incorporate a Disney, Pixar, Marvel, or Lucasfilm character? (December 2025 Disney-OpenAI $1B partnership; non-Disney-branded use requires careful IP review.)
- What is the rollback if the model is later found to have been trained on data it should not have been? (For text, the Bartz precedent is your floor; for image and video, the answer is currently "you absorb the risk.")
- Is my country/jurisdiction a signatory to a generative-AI training framework that requires license? (EU AI Act Article 53(1)(c) on training-data transparency is the operative rule for any model shipped in the EU; Japan, Korea, and the UK have parallel frameworks.)
- Am I building a human-authored layer that compounds, or am I generating disposable AI drafts? (The first builds a career; the second is a race to the bottom.)
Frequently asked questions
Can I copyright an image I made with Midjourney?
Purely AI-generated images get no copyright protection in the US — the Thaler v. Perlmutter ruling (D.C. Circuit, March 2025; cert denied March 2026) confirmed that zero human authorship means zero copyright. The US Copyright Office's January 2025 Part 2 report went further: prompting alone is not enough; you need to have determined sufficient expressive elements yourself. If you use Midjourney for a base, then make significant human-authored changes in Photoshop (selection, coordination, arrangement, modification), the human-authored parts can be protected, but the raw AI output cannot.
Did the Anthropic $1.5 billion settlement mean AI training on books is now legal?
No — it means the opposite is now the legal baseline. Judge Araceli Martínez-Olguín's June 2025 Bartz ruling held that training on lawfully-acquired books is fair use, but Anthropic's downloading of books from pirate libraries (LibGen, PiLiMi) was not. The July 20, 2026 settlement of $1.5 billion (roughly 4x the $750 statutory minimum per work) covers only the piracy. Future training cases (NYT v. OpenAI, the 28 remaining opt-out authors) are still active and the precedent is still narrowing.
Can a studio use my voice to train an AI without my consent?
Not if you are SAG-AFTRA union, and increasingly not if you work in California regardless. The 2023 TV/Theatrical and 2025 Commercials contracts both require explicit, signed consent before a digital replica is created. The 2025 Commercials contract adds a 1.5x session fee trigger when a replica is used. California AB 2602 (effective 2025) made vague "any and all AI uses" language in talent contracts unenforceable. Outside union and California, the answer is still legally murky — ElevenLabs' own Professional Voice Clone flow now requires signed consent per the updated Eleven v3 documentation.
What AI image generator should I use for client work?
Adobe Firefly, if the client is paying and the deliverable ships in a regulated or brand-sensitive context. It is the only mainstream image generator with explicit IP indemnification on paid plans (Enterprise VIP covers up to $10,000 per output per incident). Midjourney V7 produces better-looking work but offers no contractual defense. Sora 2 has the highest quality video but no indemnification and the December 2025 Disney deal (200+ copyrighted characters) makes any new Sora output a careful IP review. The cheap answer is to combine: Firefly for the deliverable, Midjourney for mood boards, and your own Photoshop pass for the human-authored layer.
What is actually happening to creator income in 2026?
Three data points, in order. (1) CISAC's Global Collections Report 2025: 2024 global music creator royalties hit €12.59 billion (+7.2% year over year) — that part is healthy. (2) CISAC's PMP Strategy study projects GenAI could divert 24% of music creator royalties (€4 billion) and 21% of audiovisual (€4.5 billion) by 2028 if licensing frameworks are not established. (3) The Anthropic Bartz settlement (July 2026) set the first hard price for AI training data at approximately $3,000 per work. Net: the music pie is still growing but the slice going to human creators is at structural risk over the next 24 months; books and journalism are now priced in the $3,000-per-work band; image and video are still unpriced.
What to do this week
If you are a working artist, writer, or musician in 2026, three things. First, audit your last six paid projects and count how many of them used AI in any portion of the deliverable. Second, for the next paid project, run the 12-question decision tree in the section above before you open the model. Third, pick one tool in your discipline that is contractually defensible (Firefly for image, Suno v5.5 with Custom Models for music, ElevenLabs v3 with signed consent for voice) and make it your default. The director-orchestrated, human-authored-layer workflow is the one that pays. The vending-machine workflow is the one that gets undercut.