Segments covered: news and magazine publishers, broadcasters and streaming services, and film and television production and post-production, with notes on music and games where their constraints differ. Linked workflows and technology: speech systems, computer vision, catalog enrichment, project knowledge retrieval, retrieval, support replies. Linked roles: support leaders, finance leaders, operations leaders, engineering leaders.
Media companies sell attention, access, and rights. A story or an episode earns money when it reaches an audience, when advertisers pay to sit beside it, and when someone licenses it again years later. The same asset is a bundle of permissions: who wrote it, who appears in it, what music is under it, and what each contract allows. AI can make describing, searching, captioning, and localizing a library cheaper, and it raises questions about who owns what a model produces and whose work and likeness a model may use.
Sourced findings are cited inline with their date and limits. Recommendations about sequencing, ownership, and controls are proposed practice, not findings. The company in the sequencing section and every number in the localization example are invented. No media company was interviewed and no deployment observed. Nothing here is legal advice; where rights or labor terms decide a project, the question goes to counsel and to whoever administers the relevant agreements.
How the business makes money
Media companies combine a few revenue lines in different proportions. The lines behave differently when AI lowers the cost of producing or packaging content:
| Revenue line | Who pays | What AI could change | What it cannot change by itself |
|---|---|---|---|
| Advertising | Advertisers, for audience reach next to content | Volume of packaged content, speed of publishing, searchability | Audience demand, ad prices, and whether advertisers accept the content |
| Subscriptions and circulation | Readers, viewers, listeners | Service cost, accessibility, retention tooling | Whether people value the product enough to keep paying |
| Licensing and syndication | Other outlets, platforms, AI developers | Discoverability of the catalog, localization cost, clearance speed | Whether the company holds the rights it would be licensing |
| Production fees and commissions | Studios, networks, streamers paying producers | Post-production and localization cost per hour | Contract terms, union rules, and the commissioner's budget |
| Events, merchandise, other | Attendees, buyers | Marketing and support cost | Core demand |
The main costs are people (editorial, production, post-production, sales), content acquisition and talent, and distribution. The library is a distinct asset: past articles, photos, footage, and recordings whose value depends on being found, cleared, and packaged.
For scale, the US Bureau of Labor Statistics' Current Employment Statistics survey puts August 2026 employment (preliminary, seasonally adjusted) at 328.7 thousand in motion picture and sound recording industries (NAICS 512), 233.1 thousand in newspaper, periodical, book, and directory publishers (NAICS 5131), and 331.1 thousand in broadcasting and content providers (NAICS 516). BLS CES5051200001, CES5051310001, CES5051600001 These are payroll counts. They exclude the self-employed and freelancers, and they say nothing about AI.
Media employment is concentrated in Manhattan. In the BLS Quarterly Census of Employment and Wages for 2025, private employers in New York County (Manhattan) employed 39,938 people in motion picture and sound recording, 11.0% of the US total; 32,282 in newspaper, periodical, book, and directory publishing, 13.0%; and 44,299 in broadcasting and content providers, 13.0%. Manhattan's share of all US private employment was 1.7%. These shares are our arithmetic on the New York County and national annual-average files. They count jobs by establishment location, cover Manhattan only, and exclude freelancers.
This guide's lens is the company that owns content and runs editorial or production work on it: a publisher, a broadcaster or streamer, or a production company. Music companies share the rights questions but have their own licensing structures, and game studios are closer to software development; neither is analyzed. Social platforms are outside the lens.
Where AI could touch the value chain
| Process | Capability | Canonical chapter | Benefit mechanism | Benefit location |
|---|---|---|---|---|
| Transcription, captioning, and subtitling | Speech recognition, translation | Speech systems; pattern from meeting follow-up | Lower cost per program hour; more of the library made accessible | Cost of delivery; revenue only with a buyer |
| Archive and asset tagging | Image and video recognition, extraction | Computer vision; metadata pattern from catalog enrichment | Catalog-maintenance capacity; findable assets | Operating expense; licensing revenue if assets sell |
| Archive and research search | Retrieve, summarize | Project knowledge retrieval, retrieval | Research time for journalists and producers | Cost of delivery |
| Subscriber and audience support | Retrieve, classify, generate drafts | Support replies | Service capacity | Operating expense |
| Rights and clearances records | Records and rules; contract extraction later | Ordinary software first; then document extraction | Fewer uncleared uses; faster licensing | Risk; licensing revenue |
| Generating scripts, articles, images, synthetic voices or performers, dubbing in a synthetic voice | Generate | Not yet covered | Production cost or new formats | Contested; see constraints |
| Recommendation and personalization | Predict, rank | Not yet covered | Engagement and retention | Revenue |
| Ad sales, yield, and audience targeting | Predict, optimize | Not yet covered | Ad revenue | Revenue |
How the existing chapters transfer, and what differs. The speech chapter's evaluation method (test on your own speakers, audio, and vocabulary; score names and numbers separately) carries over directly. What differs is the standard: broadcast captions are held to a rule (see constraints), and a subtitle file delivered to a licensee is checked against that licensee's specification. Live news captions differ again, because errors reach viewers before anyone can correct them.
Archive tagging follows the catalog chapter's pattern: a system proposes attributes linked to evidence, and a steward approves. Here the attributes are people, places, events, and subjects. Identifying a person in a photo is a matter for counsel before it is a matter for a model. The computer-vision chapter's warning about changing inputs applies: a model tuned on recent digital footage can fail on scanned film or old prints.
Archive search follows the knowledge-retrieval chapter: show each user only what they may see, cite the source, and deny unlabeled material by default. In media the label that matters is reuse permission: a producer may view footage that cannot air without a license.
Subscriber support follows the software support chapter almost unchanged; billing disputes, cancellations with retention offers, and complaints about content stay with their existing procedures.
Not AI first. Rights and clearances are a records problem. A register of who owns each asset, what music and talent are in it, and what each license allows comes before any model that reads contracts; without it, every other row inherits the uncertainty.
Constraints that shape every project
Copyright in AI output. The US Copyright Office's January 2025 report on copyrightability concluded that copyright does not extend to purely AI-generated material or to material where there is insufficient human control over the expressive elements, that prompts alone do not provide sufficient control with current generally available technology, and that human expression perceptible in an output, and creative selection, arrangement, or modification of outputs, remain protected. Using AI to assist rather than stand in for human creativity does not affect protection. Sufficiency is decided case by case. Copyright Office, Part 2: Copyrightability The Office's registration guidance, effective March 16, 2023, tells applicants not to list an AI tool as an author and to exclude AI-generated content that is more than de minimis from the claim. Registration guidance, 88 Fed. Reg. 16190 For a catalog business, the consequence is record-keeping: knowing which parts of a work people made is part of knowing what you own.
Training on copyrighted works. The Office's Part 3 report on generative AI training was released on May 9, 2025 as a pre-publication version; the Office's AI page says a final version will follow without substantive changes expected. It concluded that several stages of model development implicate copyright owners' rights and that fair use cannot be prejudged; various training uses are likely transformative, but commercial use of vast troves of works to produce expressive content that competes with them in existing markets, especially through illegal access, goes beyond established fair use boundaries. Copyright Office, Part 3 (pre-publication), status It is agency analysis, not a court ruling, and it bears on two questions for counsel: what vendors' models were trained on, and on what terms to license your own catalog.
Writers' agreements. The Writers Guild of America summarizes its Minimum Basic Agreement AI terms: AI is not a writer and its output is not literary material; AI-generated material handed to a writer does not count as assigned or source material for pay and credit; a writer may use AI with the company's consent but cannot be required to; companies must disclose AI-generated material they give writers; and the Guild reserves the right to assert that training on writers' work is prohibited. WGA, Artificial Intelligence The 2026 MBA, ratified April 24, 2026 for May 2, 2026 to May 1, 2030, preserves the 2023 protections according to the Guild's FAQ, and adds written notice to the Guild when a company licenses writers' work to train a commercial generative AI system, after which the Guild may request discussion, potentially including compensation. WGA 2026 MBA FAQ, ratification These are union summaries, not contract text.
Performers' agreements. We could not retrieve SAG-AFTRA's own summary or memorandum pages, so this guide states no provision of its TV/Theatrical contracts. Any project touching a performer's voice, face, or performance, including dubbing in a cloned voice, needs the current agreement read by counsel and labor relations before design starts.
Likeness and digital replicas. The Copyright Office's July 2024 report on digital replicas found existing laws insufficient against unauthorized realistic replicas and recommended a new federal law; that is a recommendation to Congress, not a statute. Copyright Office, Part 1: Digital Replicas In New York, the governor signed two bills on December 11, 2025: one requiring disclosure when advertisements include AI-generated synthetic performers, and one requiring consent from heirs or executors before commercial use of a deceased person's name, image, or likeness. Governor's release, December 2025 A June 9, 2026 release states the advertising disclosure requirement is in effect. Governor's release, June 2026 We have read the releases, not the statutes; definitions, exemptions, and penalties are questions for counsel.
Caption quality. FCC rules require closed captions on televised programming to be accurate, synchronous, complete, and appropriately placed; accuracy means matching the spoken words in order, without substituting words for proper names and places, and without paraphrasing except where time constraints require it. 47 CFR 79.1 That is the evaluation target for a broadcast captioning trial; whether it reaches podcasts or streaming-only programs is a question for counsel.
Provenance. The C2PA specification (version 2.4, April 2026) defines a manifest, called a Content Credential, as signed provenance information bound to an asset by hashes or by fingerprints and watermarks. The specification says it does not judge whether provenance is good or bad, only whether the assertions validate, are well formed, and are untampered. C2PA specification 2.4 Content Credentials can record that a tool was used; they do not make a picture true. We make no claim about their adoption.
Contracts on the catalog. Agreements with wire services, stock libraries, freelancers, and talent may predate AI. Whether they permit machine processing, third-party indexing, or licensing for training is a per-contract question; treat unrecorded rights as unknown, not as permitted.
Editorial standards. Corrections, sourcing, and byline rules apply to anything published, however it was produced.
The sector's economics lens
For your finance lead
The decisive question for a media company is the counterfactual: was this work going to be done and paid for, or does AI make it possible for the first time? A cost saving and a new revenue bet need different evidence. The example below is invented, and its costs are not market rates.
An invented library owner holds 600 hours of programming without Spanish subtitles. Human subtitling costs $1,200 per program hour: $720,000 in total. A machine draft costs $100 per hour, and a human subtitler reviewing and correcting it costs $600 per hour as planned, so the assisted cost is $700 per hour, or $420,000.
| Case | Cost of the work | Compared with human subtitling | What it means |
|---|---|---|---|
| Human subtitling | $720,000 | — | Baseline, if the work is committed |
| Assisted, review as planned ($700 per hour) | $420,000 | $300,000 saved | A real saving only if the $720,000 was going to be spent |
| Assisted, review slower ($1,000 per hour) | $600,000 | $120,000 saved | Review effort decides the result |
| Work not otherwise committed, $250,000 licensing offer | $420,000 | Not a saving | Offer minus cost: −$170,000 |
Three points follow from the arithmetic, not from any measurement. First, the saving exists only against spending that would have happened; otherwise the $420,000 is new cost and the subtitles must earn at least that much. Second, review effort carries the result: moving review from $600 to $900 per hour cuts the saving from $300,000 to $120,000, so measure correction time on your own programs before a volume commitment. Third, rejected deliveries to a licensee are a cost the table leaves out.
Avoid two forms of double counting. Do not count both a localization saving and the licensing revenue from the same hours as gains against the same baseline: one assumes the work was committed, the other that it was not. And catalog-tagging capacity, valued as staff time in the catalog chapter, is not also licensing revenue unless a sale can be traced to an asset that was findable only because of the new tags. The finance leaders guide sets out how to test whether a mechanism reaches cash, and the economics chapter distinguishes benefit location from accounting classification. This example is a decision model, not accounting advice; how a company capitalizes or expenses content and localization costs is a question for its accountants.
What is measured about adoption and outcomes
For your analyst
Adoption. The US Census Bureau's Business Trends and Outlook Survey asks employer businesses every two weeks whether they used AI "in any of its business functions" in the last two weeks. For the cycle collected September 21 to October 4, 2026 (reference period September 7–20; published October 8, 2026), 50.2% of businesses in the Information sector (NAICS 51) said yes, with a standard error of 1.43 points, against 23.8% of all businesses. BTOS sector data, national data The Information sector is broader than media: it includes software publishers, telecommunications, and data processing and hosting.
The subsector file splits it further. In the same cycle, 34.6% of businesses in motion picture and sound recording (NAICS 512) said yes (standard error 4.60 points), 64.3% in publishing industries (NAICS 513, standard error 3.88), and 36.6% in broadcasting and content providers (NAICS 516, standard error 12.21). Expected use in the next six months was 38.8%, 64.7%, and 39.7%. BTOS subsector data Read these with care:
- NAICS 513 is not news publishing. Under the 2022 classification it includes software publishers alongside newspaper, periodical, book, and directory publishers, and the file does not separate them. Its high figure cannot be attributed to newsrooms.
- NAICS 516 includes more than broadcasters. It covers radio and TV broadcasting along with streaming distribution services, social networks, and other content providers.
- The errors are wide. With a standard error above 12 points, NAICS 516's swings across five cycles (34.7%, 40.8%, 27.8%, 44.1%, 36.6%) are within sampling noise.
- Definitions limit the reading. The estimates are shares of firms, weighted to the business population, not of employees or revenue. BTOS methodology, version 5 The question was reworded in November 2025, so earlier figures are not comparable. Census Bureau, May 2026 Subsectors 513 and 516 exist in the file only from cycle 202616. BTOS is an experimental product, and use "in any business function" can mean one person using a writing tool.
Share of businesses using AI, by sector
- This guide’s sector
- Other sectors in the guide
- All businesses
Show the data
| Group | Estimate | Standard error | 90% interval |
|---|---|---|---|
| Information | 50.2% | 1.43 pts | 47.8–52.6% |
| Motion picture and sound recording industries (NAICS 512) | 34.6% | 4.60 pts | 27.0–42.2% |
| Publishing industries, including software publishers (NAICS 513) | 64.3% | 3.88 pts | 57.9–70.7% |
| Broadcasting and content providers (NAICS 516) | 36.6% | 12.21 pts | 16.5–56.7% |
| Finance and insurance | 40.1% | 1.86 pts | 37.0–43.2% |
| Professional, scientific, and technical services | 42.7% | 1.01 pts | 41.0–44.4% |
| Health care and social assistance | 24.8% | 1.01 pts | 23.1–26.5% |
| Wholesale trade | 21.6% | 1.46 pts | 19.2–24.0% |
| All employer businesses | 23.8% | 0.41 pts | 23.1–24.5% |
What AI-using firms say they did. The 2025–26 BTOS AI supplement (reference period November 2025 to January 2026) is published at sector level only. Among AI-using Information-sector firms, 59.5% said they used AI to supplement an employee's task and 16.8% to perform a task previously done by an employee. BTOS AI supplement tables These are self-reports across the whole sector, not measured outcomes in media.
Outcomes. We found no attributable measurement of how AI use changes captioning cost, archive revenue, newsroom productivity, or production budgets across media companies. Vendor case studies exist; none was used here, because none we found states a baseline, population, and independent check. The Neuronify AI hiring index measures job postings at New York AI companies; it is not evidence of adoption or outcomes in media.
Who is involved
Editors and executive producers own what is published or aired and its corrections. Post-production supervisors own caption, subtitle, and delivery quality. A rights and clearances lead owns the register of who owns what. Counsel owns copyright, likeness, and disclosure questions. Labor relations owns the union agreements and is the route to the guilds when notice or bargaining is required. Archivists own metadata and its vocabulary. Audience teams own support, with support leaders as the closest role guide. Finance leaders own whether a benefit reaches cash. Operations leaders own the end-to-end process, and engineering leaders own integration with asset-management systems, access control, and logging. This guide does not yet include role guides for editors, rights leads, or archivists.
Craft and labor are a decision, not a forecast. Captioners, subtitlers, researchers, and archivists hold the judgment a review step depends on. If assistance takes over first drafts, decide who reviews, how reviewers keep their skills, and whether an agreement requires notice or consultation.
Expertise specific to this industry. A general AI builder lacks the rights and labor knowledge this guide's constraints turn on: clearing catalog contracts for machine processing or training, recording which parts of a work people made, reading the guild agreements' AI terms (the WGA's consent, disclosure, and training-notice provisions, and the current SAG-AFTRA agreement for any voice or likeness work), and meeting the FCC caption accuracy standard in a broadcast captioning trial. This guide places that expertise in-house, with the rights and clearances lead, labor relations, and post-production supervisors, with counsel as adviser on copyright, likeness, and disclosure. See Staffing an AI pilot and Choosing and working with an AI partner.
Sequencing: a first move and what follows
The following is a fictional teaching example. Halvorsen Street Media is an invented New York company of about 400 people with three magazines, a podcast network, a small video unit, and a photo and print archive going back four decades. Its numbers and decisions exist only for this example.
First, record what it owns. Halvorsen builds a rights register for the podcast network and the photo archive: for each asset, who made it, under what contract, what third-party material it contains, and whether reuse, third-party processing, and licensing are permitted. Assets whose contracts cannot be found are marked unknown and excluded. No tool is selected yet.
First move: transcripts and captions for its own podcasts and video. Halvorsen chooses transcription of shows it fully owns, following the speech systems chapter, because the output can be checked against the audio, guests signed release forms, and a transcript mistake is visible and correctable. An editor reviews every transcript before publication. The trial measures errors on names and numbers and correction time per hour of audio; success means a cost per hour below the current contractor's, on Halvorsen's own audio.
Next, if that works: archive tagging, then archive search. Tagging comes second, in the photo archive only, with an archivist approving every proposed tag as the catalog chapter proposes; person identification is excluded pending counsel's review. Search comes third, for newsroom research, showing reuse status beside each result as the knowledge retrieval chapter recommends. Subscriber support drafting (support chapter) can run in parallel because it touches no archive rights.
Counterexample: synthetic voice reads wait. The podcast team proposes audio versions of articles read in a cloned voice of a popular host. Halvorsen declines for now. The host's contract predates voice cloning, no union agreement has been reviewed, freelance writers' contracts do not address audio adaptation, and there is no disclosure practice for synthetic voices. A contract review, a negotiated consent and compensation term with the host, and a disclosure standard come first. The constraint is permission, not capability.
Questions for operators, investors, and vendors
- What share of the library has a recorded owner, contract, and reuse permission?
- For each proposed use, was the work going to be paid for anyway? If not, what revenue is expected and from whom?
- Which union agreements cover the people whose work or likeness is involved, and what do they require?
- Who reviews AI-produced captions, subtitles, tags, and answers, and how is their correction time measured?
- For vendors: is our content used to train or improve shared models, and what contract term says so?
- For vendors: does your system preserve provenance records through our edit and export chain?
- For investors: does a claimed AI advantage show up in cost per program hour, catalog revenue, or audience retention, or only in demonstrations?
When to wait
Wait when the library's rights are unrecorded, when the work was never going to be funded and no buyer is named, or when a project touches a writer's or performer's work or likeness before counsel and labor relations have read the applicable agreements. Wait on live news captioning until batch captioning has been measured on your own audio. The simpler investments pay off without AI: a rights register, a controlled metadata vocabulary, a clean asset-management system, and contract templates that address AI for future work.
Evidence still needed and related reading
Obtain operator accounts of subtitle correction time per program hour, tagging accuracy by era and format, and licensing revenue traceable to newly findable assets. Read SAG-AFTRA's current agreement (we could not retrieve it), the WGA 2026 memorandum of agreement, and the New York statutes behind the governor's releases. Check whether the final Part 3 report has been published. A dedicated workflow chapter on media transcription and indexing is planned but not yet written. No media company has been interviewed and no deployment observed for this chapter.
Continue with speech systems, computer vision, catalog enrichment, project knowledge retrieval, support replies, and finance leaders.
Hiring in New York: open roles at NYC AI companies building generative media and voice tools, read daily from the companies' own job boards. These companies sell AI into the industry; they show which skills vendors are hiring for, not what firms in the industry are hiring.