IPTC metadata is descriptive and rights information written inside an image file according to standards maintained by the International Press Telecommunications Council (IPTC), a London-based body founded by news organisations in 1965. The fields record who took or made a picture, what it shows, who must be credited, who owns the copyright and, since 2023, whether a generative artificial intelligence (AI) model produced it. They exist so that this information travels with the file rather than sitting on a web page the image can leave behind. For advertisers and retailers, a newsroom convention has become a compliance signal: Google, Meta and Pinterest read it to decide when a picture gets an AI label.

Two containers, one vocabulary

The IPTC Photo Metadata Standard defines properties; separate formats carry them. The older carrier is the Information Interchange Model (IIM), a binary structure the IPTC developed in 1990. In a JPEG (Joint Photographic Experts Group) file, IIM data sits in an APP13 segment marked "Photoshop 3.0", inside an Adobe image resource block numbered 0x0404, because Photoshop popularised embedding it. The newer carrier is the Extensible Metadata Platform (XMP), an Extensible Markup Language (XML) format Adobe introduced in September 2001 that became International Organization for Standardization (ISO) standard 16684-1 in 2012. XMP occupies an APP1 segment, next to the camera's Exchangeable image file format (Exif) data.

At version 2025.1, the standard splits into two schemas. IPTC Core holds 25 properties, most inherited from IIM and writable in both formats. IPTC Extension holds 41 newer properties that, according to the IPTC user guide, exist only in XMP. Software that discards XMP but keeps IIM therefore preserves a photographer's name while deleting every Extension field.

A handful of fields do most of the work. Creator maps to dc:creator in XMP and to dataset 2:80 in IIM, capped at 32 bytes. Copyright Notice is dc:rights or 2:116, capped at 128 bytes. Credit Line is photoshop:Credit or 2:110. Description, the caption, allows 2,000 bytes in IIM, and each keyword 64. Digital Source Type (Iptc4xmpExt:DigitalSourceType) takes a single value from a controlled vocabulary, expressed as a web address. A camera original is digitalCapture. An image from a text-to-image model should carry http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia. A photograph altered by inpainting or outpainting takes compositeWithTrainedAlgorithmicMedia, a composite containing any synthetic element takes compositeSynthetic, and output from a formula with no training data takes algorithmicMedia.

Writers include news agencies, stock libraries, photographers, editing software and AI generators. Readers include asset management systems, Google Images, Merchant Center and Meta's upload pipeline. The mechanism is thin. One command in ExifTool, an open-source utility the specification cites, writes the tag into a JPEG. The same command with an empty value deletes it. Nothing in the file records that either change happened.

From wire photos to generative AI

A few years after 1990, Adobe adopted a subset of IIM for Photoshop, which put IPTC fields into millions of photographs. The IPTC and Adobe then redefined those fields for XMP as IPTC Core 1.0, approved on October 8, 2004. The Extension schema followed on July 2, 2008, adding granular rights fields from the PLUS (Picture Licensing Universal System) Coalition.

Version 2017.1 added Web Statement of Rights, a link to licence terms, and 2021.1 added alt text for accessibility. Version 2023.1, published in October 2023, added Data Mining, a PLUS property whose values include "Prohibited for Generative AI/ML training" and "Prohibited except for search engine indexing".

The AI turn began with the vocabulary. The IPTC published a draft expansion of Digital Source Type on April 20, 2022, then recommended trainedAlgorithmicMedia for generated media on May 9, 2023. Google said the next day that its own generated images would carry the field. "We'll ensure every one of our AI-generated images has that metadata," Sundar Pichai, Google's chief executive, said, according to the IPTC. Midjourney and Shutterstock committed to the same marker. The inpainting term arrived on July 27, 2023. In October 2024 the IPTC relabelled the two main terms "Created using Generative AI" and "Edited with Generative AI", wording Google uses. Version 2025.1, approved on October 16, 2025 and published in November, added four fields: AI System Used, AI System Version Used, AI Prompt Information and AI Prompt Writer Name.

Where advertisers meet it

Search was the first commercial use. Google began reading IPTC rights metadata in 2018 and added a Licensable badge to Google Images in August 2020, which requires Web Statement of Rights or the equivalent schema.org license property. Where the two sources conflict, Google uses the structured data, according to its documentation.

Shopping made the AI field mandatory. Merchant Center required the trainedAlgorithmicMedia tag on AI-generated product images from February 19, 2024, across main, additional and lifestyle image links, and prohibited removing it. Text needed a separate route, since a feed title cannot carry embedded metadata: the specification added structured title and structured description attributes in April 2024.

Social platforms turned the field into visible labels. Nick Clegg, then Meta's president of global affairs, said on February 6, 2024 that Meta would label images from Google, OpenAI, Microsoft, Adobe, Midjourney and Shutterstock by reading "AI generated" information in the Coalition for Content Provenance and Authenticity (C2PA) and IPTC standards. Pinterest began showing an "AI modified" label on April 30, 2025, combining metadata with its own classifiers, according to TechCrunch. Google's About this image panel, extended to Circle to Search and Lens on July 31, 2024, shows creator details where metadata exists.

Industry frameworks build on the same layer. The Interactive Advertising Bureau's (IAB) January 2026 framework paired consumer labels with machine-readable metadata, and version 2, issued on August 18, 2026, cited a New York University (NYU) Stern finding that an AI label cut click-through by 31.5%.

Where it falls short

The fields are unsigned. Anyone can add, alter or remove them, a weakness Google notes in its About this image material. A missing tag proves nothing.

Stripping is routine. When IPTC members tested 16 sharing sites and file services in spring 2019, Twitter removed metadata from saved files, Instagram offered no download at all, and Facebook discarded XMP while keeping a few rights fields in IIM and Exif. That alone erases Digital Source Type, which lives only in XMP. Clegg conceded as much in 2024, warning that those who "actively want to deceive people with AI-generated content will look for ways around safeguards".

Over-labelling has proved as contentious as under-labelling. Meta's "Made with AI" label, introduced in May 2024, attached itself to real photographs retouched with tools such as Adobe's Generative Fill, and Meta renamed it "AI info" on July 1, 2024 after photographers complained. A single-value field records a category of origin, not the extent of a change.

The layers also collide. When the IPTC added the four 2025.1 properties to its own sample image, the edit invalidated that file's C2PA signature, because any change to a signed file breaks its hash. Data Mining depends on crawlers choosing to read it, the same honour system that governs robots.txt.

Regulators have set a higher bar. Under Article 50(2) of the European Union's (EU) AI Act, providers must mark synthetic output in a machine-readable format, and the Code of Practice published on June 10, 2026 asks for digitally signed, time-stamped metadata plus an imperceptible watermark. Brendan Quinn, the IPTC's managing director, wrote that "the only technology that meets these criteria is C2PA", while conceding that the code names no standard. On that reading, plain IPTC fields do not qualify. An IAB Austria guide adds that hidden metadata cannot replace the visible label a qualifying advert needs.

Google's own pages diverge. As of October 2026, its Search documentation recognises four AI values, while the Merchant Center image page lists three and omits the inpainting term.

Not the same as

Exif. Written by cameras and phones, Exif records exposure, lens, timestamp and sometimes location. IPTC fields describe content and rights; the standard deliberately excludes technical camera data.

XMP. XMP is a container, not a vocabulary; IPTC fields share it with Adobe and Dublin Core properties.

C2PA. The C2PA standard signs a manifest cryptographically. It reuses IPTC Digital Source Type values inside its creation actions and can wrap IPTC fields in a signed assertion, so the two are layers, not rivals. SynthID is different again: a watermark in the pixels that survives stripping but carries little information.

Structured data and platform attestations. Schema.org markup describes an image on one page and must be repeated wherever the image appears; IPTC fields travel with the file. Google's synthetic content attestation field in the Display & Video 360 and Campaign Manager 360 application programming interfaces (APIs) is a declaration stored by the platform, not data inside the asset.

Recent developments

Google's July 9, 2026 AI label setting added a "Manage AI label" control to Merchant Center, a workflow described at the time as largely superseding the 2024 metadata tag. Yet a Search Central page Google edited on October 1, 2026 still describes only the IPTC route for AI-generated product images. Merchant Center will also reject images below 500 by 500 pixels from January 31, 2027, and a picture enlarged with generative tools to clear that bar falls under the same metadata rule.

At its Photo Metadata Conference on September 3, 2026, which drew more than 250 registrants from over 100 organisations, the IPTC launched Metawatch, a public service tracking how news sites use IPTC, C2PA and AI opt-out metadata. Proposed changes to the photo standard go to its Standards Committee at an online meeting from October 19 to 21, 2026.

Timeline

  • 1965: News organisations found the International Press Telecommunications Council, based in London
  • 1990: The IPTC develops the Information Interchange Model
  • September 2001: Adobe introduces the Extensible Metadata Platform
  • October 8, 2004: IPTC Core 1.0, defined with Adobe for XMP, is approved
  • July 2, 2008: The IPTC Extension schema 1.0 is approved and the Photo Metadata Standard is published as one document
  • 2012: XMP becomes ISO standard 16684-1
  • June 2017: Version 2017.1 adds Web Statement of Rights
  • 2018: Google begins supporting IPTC image rights metadata
  • Spring 2019: IPTC members test metadata retention on 16 sharing sites and file services
  • December 2019: Version 2019.1 introduces image regions
  • August 2020: Google Images launches the Licensable badge
  • November 2021: Version 2021.1 adds accessibility alt text
  • April 20, 2022: The IPTC releases a draft Digital Source Type vocabulary for synthetic media
  • May 9, 2023: The IPTC recommends trainedAlgorithmicMedia for AI-generated media
  • May 10, 2023: Google commits to IPTC metadata on its AI-generated images
  • July 27, 2023: The compositeWithTrainedAlgorithmicMedia term is added
  • October 2023: Version 2023.1 adds the Data Mining property
  • February 6, 2024: Meta announces labels based on C2PA and IPTC signals
  • February 19, 2024: Merchant Center requires trainedAlgorithmicMedia on AI-generated product images
  • April 2024: Merchant Center adds structured title and structured description attributes
  • May 2024: Meta begins applying "Made with AI" labels
  • July 1, 2024: Meta renames the label "AI info"
  • October 2024: The IPTC relabels its main AI terms "Created using Generative AI" and "Edited with Generative AI"
  • April 30, 2025: Pinterest begins showing "AI modified" labels
  • October 16, 2025: The IPTC Standards Committee approves Extension schema 1.9
  • November 2025: Version 2025.1 is published with four AI properties
  • January 2026: The IAB publishes its AI Transparency and Disclosure Framework
  • June 10, 2026: The EU Code of Practice on Transparency of AI-Generated Content is published
  • July 9, 2026: Google introduces its AI label setting across five products, including Merchant Center
  • August 2, 2026: Article 50 transparency obligations apply in the EU
  • August 18, 2026: The IAB publishes version 2 of its disclosure framework
  • September 3, 2026: The IPTC launches Metawatch at its Photo Metadata Conference
  • October 1, 2026: Google edits Search guidance that still cites the IPTC route for AI product images
  • October 19-21, 2026: The IPTC autumn meeting, including a Standards Committee session
  • January 31, 2027: Merchant Center's 500 by 500 pixel image minimum takes effect

Summary

Who. The International Press Telecommunications Council, a London-based standards body of more than 60 news and media organisations, maintains the standard with the PLUS Coalition and Adobe's XMP technology. News agencies, stock libraries, photographers, editing software and AI generators write it; Google Images, Merchant Center, Meta and Pinterest read it.

What. A set of standardised fields, 25 in IPTC Core and 41 in IPTC Extension as of version 2025.1, embedded in image files in IIM or XMP format. They record caption, creator, credit, copyright, licence links, training permissions and, through Digital Source Type, whether AI made or edited the image.

When. The IIM dates from 1990, IPTC Core for XMP from October 2004 and the Extension schema from July 2008. AI terms were recommended in May 2023, made mandatory in Merchant Center from February 19, 2024 and extended with four AI properties in November 2025.

Where. Inside JPEG, TIFF, PNG and other image files, mostly in the XMP packet, and in the systems that ingest them: search engines, shopping feeds, social platforms and asset libraries.

Why. Credit and rights information that travels with a picture survives the trip between agencies, publishers and platforms better than a separate database. Generative AI gave the same fields a disclosure role, but because they are unsigned and easily stripped, regulators and platforms increasingly pair or replace them with C2PA signatures, watermarks and platform-side declarations.