The State of Play

A living index of AI adoption across industries — where established practice meets the bleeding edge
UPDATED DAILY
← All domains

🎬 Creative & Generative Media

AI for generating and editing images, video, audio, 3D assets, and cross-media content. Mostly leading-edge with rapid advancement — image generation, music composition, and voice synthesis are approaching good practice. Video generation and 3D asset creation are progressing fast but quality and controllability gaps persist. The most active domain by momentum: over half the practices are advancing.

21 practices: 4 good practice, 15 leading edge, 2 bleeding edge

The Headline

AI-made images, voices and video are now cheap and convincing. Buyers are getting returns from tools that keep what is real, while audiences and platforms push back on AI work that replaces the finished product.

21
practices tracked
17
at the frontier
2
moving this fortnight
4,212
evidence items

The Picture

Generating images, voices and short video clips is now inexpensive and easy to get, and most marketing and creative teams already use it for drafts, variations and resizing. The organizations reporting measurable value treat AI as raw material under human direction. Examples include training and localization video, automated sports highlights, and product backgrounds built around real photographs. Where AI output has to stand alone as the finished product, results fall behind: by one industry estimate, AI-disclosed games were 30.8% of Steam releases in 2026 but took about 10% of estimated second-quarter sales. Revenue is concentrating in a few large platforms such as Adobe and ElevenLabs while smaller specialists close or sell cheaply. If you use AI to produce more variants that people review, you are in the pack. If you use it to replace finished work, you are ahead of the evidence.

This Fortnight

  • Vestiaire Collective reports that every generative image tool it tested changed product details, while edits that kept the original photo lifted item liquidity 5%. The resale marketplace randomly split listings between its in-house tool and Photoroom's edits, which excluded generative changes. Photoroom published the results and gave no sample sizes. For any image a customer will buy from, test tools on whether the product comes through unchanged, not on how good the picture looks.

  • Google made its Omni 1.2 Flash video model generally available on November 5, with 15-second clips and a 30% price cut to $0.07 a second at 720p. At that rate a 15-second clip costs about a dollar, by Google's own figures. Reviewers warn that repeated edits can erase the savings. The price of generation is no longer the constraint, so budget for the number of attempts it takes to get a clip you can use.

  • Roblox said more than 80% of its top 100 games used AI to speed up creation, and its shares fell about 11% in after-hours trading. In the same weeks, the indie game store itch.io deleted "a couple terabytes of AI slop" and added a No AI search filter. VRChat's Steam rating fell to Mostly Negative after it declined to label AI-made worlds. Producing more content with AI is not translating into engagement, and audiences increasingly want to see AI content labeled and to filter it out.

  • The English version of Ramayana: Part 1 opened in cinemas on November 5, with AI reshaping the actors' lip movements to match English dialogue. Human voice actors performed the lines, and the vendor, Brahma AI, changed only the mouths. SAG-AFTRA's newly ratified contract now requires consent before a performer's digital replica dubs them into another language. Visual dubbing has reached premium releases, so treat performer consent as a contract line item from the start.

  • In a study of 1,783 adults, including judges and attorneys, people told AI-generated media from authentic media correctly only 47.4% of the time. Legal training gave no advantage, and a separate test showed a detector's accuracy collapsing on material unlike its training data. Do not rely on staff or a single detection tool to catch fakes in payment approvals, evidence or hiring. Build verification steps into the process itself.

Coming Up

  • Adobe expects to put connectors for outside AI agents into private beta for Creative Cloud customers in early 2027. The connectors use MCP (a connector standard for AI tools) and let agents (software that acts on its own without being prompted) work inside Adobe's apps. Several of the related features are still in beta or preview, and Adobe has not given pricing for each one. Before switching them on, decide who approves agent-made assets, because review queues are already the bottleneck in most marketing teams.

  • YouTube plans to pilot automatic dubbing of livestreams in early 2027, extending a feature that is already switched on by default for eligible creators. One dubbing vendor reports that AI-only dubbed tracks held viewers 4 to 10 times less well than professionally dubbed ones across its client channels. The vendor sells professional dubbing and published no methodology. If you localize video, measure watch time by language before letting free automatic dubs replace paid ones, especially for emotional, comedic or children's content.

  • The leading US artists' case against image generators, Andersen v. Stability AI, has summary judgment set for September 29, 2027. No US court has yet ruled on whether training image generators on copyrighted art is fair use. In music, Universal and Sony have filed a second suit against Suno covering 60,202 recordings. Until these cases are decided, check what your vendor's indemnity actually covers: Adobe's, for example, excludes beta features and output that has been modified or combined.

What's Hard About This

  • Generative tools fill gaps with plausible detail, while the most valuable commercial uses depend on keeping what was actually there. In a study of a commercial pipeline that generates 15-second product ads, reviewers marked the product wrong in 60% of 562 scenes, compared with under 10% each for script, caption or sound problems. A human-in-the-loop step (a person reviews each AI output before it ships) is a permanent cost of product and brand work, not a temporary one.

  • Audiences often cannot spot AI work, yet they mark it down when they know or suspect it. One field experiment found that disclosing AI use cut ad click-through by 31.5%. EU rules enforced since August 2 require exactly those labels. Put generated content where it holds up even when labeled, such as training, product detail and variation testing, rather than in brand storytelling.

  • Consent, not capability, now sets the cost of using voices and likenesses. In a survey by the US National Association of Voice Actors, 21% of voice actors had knowingly lost work to synthetic voices, while only 13% had ever agreed to a clone. Courts and unions from Munich to Hollywood are tightening the rules. Choose vendors that can document consent for every voice and face, because the legal exposure stays with you.

2012View this domain's timeline →Today

Practices in this Domain (21)

PRACTICETIERTREND
3D asset, scene & texture generationLEADING EDGE— Steady
AI-driven video editing & post-productionLEADING EDGE— Steady
Audio production — editing, podcasts & sound designLEADING EDGE— Steady
Avatar generation & personalised media at scaleLEADING EDGE— Steady
Brand asset generation & variationGOOD PRACTICE— Steady
Content authenticity — deepfake detection & provenanceLEADING EDGE— Steady
Image editing — inpainting, outpainting & extensionLEADING EDGE— Steady
Image editing — style transfer & artistic transformationLEADING EDGE— Steady
Image generation — photorealistic & illustrativeGOOD PRACTICE— Steady
Image generation — product visualisation & mockupsGOOD PRACTICE— Steady
Image processing — upscaling, restoration & compositingLEADING EDGE— Steady
Interactive content — game, AR/VR & environment generationLEADING EDGE↘ Slowing
Lip sync & video dubbingLEADING EDGE— Steady
Motion capture & pose estimation for content productionLEADING EDGE↘ Slowing
Multimodal content generationGOOD PRACTICE— Steady
Music generation — background & ambientLEADING EDGE— Steady
Music generation — full compositionBLEEDING EDGE— Steady
Text-to-speech — natural voice synthesisLEADING EDGE— Steady
Text-to-speech — voice cloning & custom voicesLEADING EDGE— Steady
Video generation — long-form narrative & explainerBLEEDING EDGE— Steady
Video generation — short-formLEADING EDGE— Steady
Read the full technical briefing (1,630 words) →

Where AI Stands in Creative & Generative Media

Creative and generative media is the domain where capability has most clearly run ahead of any settled idea of what it is for. Generating images, voices and short video clips is now a commodity. Leaderboard positions change hands within weeks, and in a study of 1,783 adults, including judges and attorneys, participants told AI-generated media from authentic media only 47.4% of the time. Revenue has concentrated in a few firms: Adobe's AI-first ARR is above $650M, the FT reports ElevenLabs pacing at $600M ARR on a $22bn valuation, Runway reports $200M ARR and Kling is put at a run rate near $500M. Where AI does bounded, rule-bound work at volume, the momentum is real. WSC Sports reports, through its own unaudited figures, 16 million highlights from more than 450,000 broadcasts in 2025. Synthesia has passed $100M ARR, largely on training and localisation video. Stem separation and speech cleanup now ship as standard features in Premiere and iZotope.

Progress stalls wherever the output has to stand alone as finished work in front of a paying audience. Sanxingdui, China's first AI feature cleared for cinemas, recovered roughly 3.9% of a budget above 10M yuan, by one account drawing on AFP. Deezer receives about 90,000 AI tracks a day, but they draw only 1-3% of streams. AI-disclosed games made up 30.8% of Steam releases in 2026 but took about 10% of estimated sales in the second quarter. In an Ipsos study from May 2026, human-made ads beat fully AI-made ones by 14% on short-term and 17% on long-term predicted effectiveness. The pattern is consistent. Value goes to AI used as raw material, tooling and variation under human direction, not to AI as a replacement for the finished product.

The domain's defining failure is fidelity, not capability. Generative methods fill gaps with plausible detail, while the most lucrative uses, including product imagery, restoration, dubbing and brand assets, depend on keeping what was actually there. In Photoroom's benchmark, base models kept a product correct only 29% of the time. A second pressure is consolidation. The platforms that own distribution are absorbing generation as a metered feature: Adobe bought Topaz for about $340m, and Premiere, YouTube, Google Ads and Roblox Studio all bundle generation. Specialists are folding. Beatoven closed its consumer service in August, Move AI was sold out of administration for £525,000 after raising $17.5m, and OpenAI closed Sora. Rights remain unsettled, with a Munich court ruling against Suno and UMG and Sony suing it again over 60,202 recordings, so adopters still carry the legal exposure themselves.

What's New, 2026-10-29 to 2026-11-12

The clearest finding this fortnight is that buyers are earning returns from fidelity, not from generation. In a randomised split published by Photoroom, Vestiaire Collective reports that every generative tool it tested altered product details. Photoroom's edits, which preserve the original product, lifted item liquidity by 5%; no sample sizes were given. A study of a commercial pipeline that generates 15-second product ads found reviewers marking the product wrong in 60% of 562 scenes. The FDA's catalogue now lists sFRC, open-source software that flags hallucinated patches in AI-restored CT and MRI scans, in the same weeks that Adobe made Topaz-powered Generative Upscale generally available in Lightroom 10.0. Verification also looked weak: the 1,783-adult study found legal training no help in spotting AI media, and one detector's AUC fell from 0.98 to 0.66 once tested outside its training data. Releases continued. Google made Omni 1.2 Flash generally available on 5 November with 3-15 second clips and, by its own figures, a 30% price cut to $0.07 a second at 720p. ElevenLabs opened Professional Voice Clone fine-tuning on v4 to every workspace. Premiere gained beta Generate Music and Generate Soundscape tools. YouTube's Replace Song began writing AI instrumental replacements for creators outside the EEA, and the English version of Ramayana: Part 1 opened in cinemas on 5 November with Brahma AI's lip-sync.

The pushback was just as concrete. Itch.io deleted "a couple terabytes of AI slop" and added a No AI search filter. VRChat surveyed more than 21,000 users, then declined to label AI-made worlds, and its Steam rating fell to Mostly Negative. Roblox said Build reached more than 500,000 users and 70,000 published games in five months, and that over 80% of its top 100 games used AI, yet its shares fell about 11% after hours. Embark replaced flat synthetic voices in ARC Raiders with human actors. Korean producer Audien made about 380 AI audiobooks against 230 human-narrated ones from January to October, but human-narrated fiction draws roughly ten times the interest. In a randomised trial, 140 urology residents rated AI-avatar lectures below traditional ones, although exam scores did not differ. On cost, Studio Meta-K's founder put the token bill alone for a 100-minute AI film above KRW500 million. On rights, SAG-AFTRA's newly ratified deal requires consent before a replica dubs a performer into another language, and putative class actions filed in October allege, so far without proof, that Grok extracted facial geometry from uploaded photos. None of this changes the overall picture; it reinforces it.

Key Tensions

  • Plausible detail versus faithful detail. Generative models fill in what is likely, not what was there, and the most valuable commercial uses punish exactly that. Adobe's Generative Upscale has altered embedded text and line art, Seedream treats a marked region as a hint and not a protected mask, and REI pulled an ad after Meta's auto-enrolled tool gave a bicycle a second set of handlebars. Buyers who measure outcomes, from Vestiaire to the FDA's sFRC tool, are paying for preservation, not invention.

  • Supply that outruns attention. Generation has made output almost free, and audiences have not grown with it. Deezer receives about 90,000 AI tracks a day that draw 1-3% of streams, 36% of October's Steam Next Fest demos carried an AI label but only two reached the top 25, and Roblox's creation boom has coincided with three straight quarters of falling daily active users. More content brings more competition for discovery, not more engagement.

  • Platforms absorb, specialists fold. The firms that own distribution are turning generation into a feature billed by credits: Adobe now owns Topaz, Premiere routes across rival video and music models, Google Ads tests free dubbing and YouTube writes replacement soundtracks itself. Standalone vendors without that distribution are closing or selling cheaply, among them Beatoven's consumer service, Move AI and Sora. Credit metering also passes the cost of re-rolls and failed generations to the user.

  • Indistinguishable yet penalised. Audiences cannot reliably spot AI work, yet they mark it down once they know or suspect. Disclosure cut ad click-through by 31.5%, the Columbia and Taboola study found AI ads matched human ones on click-through except when viewers perceived them as AI, and generative AI made up 56.2% of D&AD's AI entries but only 44.7% of AI Pencil winners. Disclosure rules such as the EU AI Act's Article 50, enforced since 2 August, require the very labels that carry the penalty.

  • Consent is the real gating cost for voice and likeness. Law on voice and likeness is moving faster than the tooling. In NAVA's survey, 21% of voice actors had knowingly lost work to synthetic voices while only 13% had consented to a clone, and Indian voice artists report collapsing dubbing work in a market with no voice-protection law. Vendors increasingly sell consent as part of the product: WellSaid offers only enterprise voices built from paid, consenting actors, and Fish Audio's licence bars cloning real people without consent.

Top 10 Evidence Items

  1. How Vestiaire Collective increased item liquidity by 5% with accurate listing photos from Photoroom (case-study) — Concrete randomised-split evidence that fidelity-preserving editing, not generation, is what actually moves a commercial outcome. https://www.photoroom.com/customer-stories/vestiaire-collective
  2. Microsoft Word - Cover Letter for the Advisory Committee (research-paper) — The headline verification statistic the briefing leans on to show audiences and even legal professionals cannot reliably spot AI media. https://www.uscourts.gov/sites/default/files/document/26-ev-18_suggestion_from_justin_sevier_henry_zhuhao_wang_and_gino_cerrone_rule_707_and_901c.pdf
  3. Why China's First AI-Generated Film Failed at the Box Office — and What's Next for AI Video Production? (opinion) — The clearest case of generated output failing to stand alone as a finished product in front of a paying audience. https://dev.to/alirezaai/why-chinas-first-ai-generated-film-failed-at-the-box-office-and-whats-next-for-ai-video-1m7e
  4. Multimodal Judges Flag, Humans Decide: A Review-and-Edit Framework for AI-Generated Product Video Ads (research-paper) — Quantifies exactly how often a generative ad pipeline gets the product wrong, grounding the fidelity-versus-plausibility tension in a real production system. https://arxiv.org/html/2611.13178
  5. What's new in Lightroom (product-ga) — Shows the acquired Topaz capability being absorbed as a metered feature inside a platform Adobe already owns, the consolidation pattern named in the tensions. https://helpx.adobe.com/mena_ar/lightroom/desktop/introduction/whats-new.html
  6. Indie Game Storefront Itch Deletes a 'Couple of Terabytes of AI Slop,' Adds 'No AI' Filter to Search Results (news-coverage) — A distribution platform actively pushing back against AI output rather than absorbing it, the sharpest form of the audience-penalty tension. https://www.ign.com/articles/itch-deletes-a-couple-of-terabytes-of-ai-slop
  7. Roblox says more than 80% of its top 100 games used AI to accelerate creation (news-coverage) — Captures the indistinguishable-yet-penalised tension directly: heavy AI adoption in top games coinciding with a market reaction against it. https://gamesbeat.com/roblox-says-more-than-80-of-its-top-100-games-used-ai-to-accelerate-creation/
  8. Human review in AI dubbing belongs wherever a real voice or face is copied (opinion) — Shows consent becoming a contractual gating cost for voice and likeness rather than a solved legal question. https://www.mba-training.com/en/blog/ai-dubbing-voice-rights-review
  9. AI Could Make Films Costlier Than the 35mm Era, Korean Pioneers Warn (news-coverage) — Undercuts the cost-saving narrative around generative film production with a concrete token-bill figure exceeding traditional costs. https://variety.com/2026/film/markets-festivals/ai-films-costlier-35mm-era-korean-pioneers-warn-1236942217/
  10. Defending Against Executive Deepfake Fraud: A Hybrid Detection Framework and Governance Strategy for Enterprise Security (research-paper) — Demonstrates that detection tooling itself is unreliable once tested outside its training distribution, reinforcing that verification is failing on both sides. https://www.mdpi.com/2305-6703/6/4/46/pdf