Sony announced last week that its Sony AI division developed new technology for determining which artists’ music is used in generative AI music outputs. While this announcement is getting some attention in the industry because of the Sony name, this joins a growing list of such technologies. It’s part of an emerging ecosystem of copyright management tools for generative AI that echoes the analogous set of tools that arose in the wake of Napster and the file-sharing era 25 years ago.
Among the most important of these technologies are AI detection and attribution. Sony’s new technology is an attribution engine: It makes an educated guess at which items of training data – songs, in this case – were most influential in generating an AI’s output.
For example, the Sony technology can state which musical artists the output resembles and at what percent influence.
Tracing Sources Of AI Output
PROMOTED
Sony’s attribution technology is at the R&D stage; the company hasn’t announced commercial availability or a brand name. Yet there are several attribution engines that are commercially available now – not from household-name companies like Sony, but from startups.
For music, there’s Sureel, which detects which artists’ “styles,” not just which copyrighted elements, are used in AI outputs. Sureel recently announced a partnership with STIM, the Swedish equivalent to ASCAP or BMI, for tracking royalties from music generated by AI platforms that take licenses from STIM. Other players in the music attribution space include Musical AI of Canada and Neutune of South Korea.
MORE FOR YOU
Inside Moltbook: The Social Network Where AI Agents Talk And Humans Just Watch
An Agent Revolt: Moltbook Is Not A Good Idea
Trump Approval Rating: Below 40% In 2 Latest Surveys Trump Alleges Are ‘Fake’
Attribution engines also exist for other types of content.
One of the more successful so far is ProRata, a spinout of the famed tech incubator Idealab. ProRata focuses on text content, though it has announced plans to expand to music. It has an arrangement through News/Media Alliance, the trade association for the news and magazine industries, to administer a voluntary blanket license for content from publishers – which means that publishers can have their content registered with ProRata and get paid royalties according to AI platforms’ use of it in their outputs in any licensing deals that ProRata and N/MA make with them.
ProRata also administers royalties from individual licensing deals that publishers have made with AI platforms.
Vermillio focuses on entertainment content; its TraceID system monitors the Internet for AI content that reflects its clients’ content as well as artists’ name, image, and likenesses (NIL); it can be used to issue takedown notices or administer license terms for matching AI content. Vermillio’s customers include the major talent agency WME.
Yet another attribution engine comes from the Israeli startup Bria. Bria focuses on images and video; it analyzes training data for various features (objects, composition, texture, etc.), and then compares an AI’s output to those sets of features. Like Musical AI, Bria’s attribution engine arose out of the company’s own generative AI platform.
The video player is currently playing an ad.
Forbes Daily: Join over 1 million Forbes Daily subscribers and get our best stories, exclusive reporting and essential analysis of the day’s news in your inbox every weekday.
Email Address
Sign Up
By signing up, you agree to receive this newsletter, other updates about Forbes and its affiliates’ offerings, our Terms of Service (including resolving disputes on an individual basis via arbitration), and you acknowledge our Privacy Statement. Forbes is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.
A Bridge Between Rightsholders And AI Platforms
Media companies find attribution engines attractive conceptually; they represent a path to getting paid for their content as generative AI output grows in sheer volume. But attribution engines have two challenges in getting accepted in the market: cooperation from AI platforms and acceptance by rightsholders.
Attribution engines require coordination with AI platforms: At a minimum, they need AI platforms’ training data to be able to map it to their outputs – although in some cases it may be possible to use a set of common content of a certain type in lieu of a specific AI’s training set.
Some generative AI platforms are likely to be more cooperative than others. At one end of the spectrum are AI companies that have no interest in copyright; at the other end are so-called ethical AI platforms that will only train on content for which they are licensed. Musical AI and Bria are two examples of this.
The likelihood that AI platforms will cooperate with attribution engines will depend on the outcomes of the many copyright lawsuits against them (80 at this writing): the more the courts favor rightsholders in these litigations, the more likely AI platforms will be to cooperate with attribution engines in relevant situations. It may well be a few years before this becomes clear.
Attribution engines will only produce adequate results from the outputs of noncooperative AI generators if they can obtain training data that is reasonably similar to the AI platform’s. (For example, a talent agency like WME can provide Vermillio’s TraceID with lots of assets that represent its clients, such as photos and video clips, so that it can detect deepfakes of those clients.)
Ethical AI companies have no need for attribution engines, as they should be able to track which elements of training data influenced each output themselves. That leaves a set of AI companies in the middle with which attribution engines might work.
The other challenge for attribution engines is acceptance by rightsholders. Attribution models produce estimates – educated guesses – of training data’s influence on outputs. Their methodologies for calculating those educated guesses are their secret sauces. This means that attribution engine companies must get rightsholders to trust their models.
Yet rightsholders have pushed for more transparent rights and royalty processes in recent years; they are not likely to trust royalty allocation models they don’t understand and might disagree with.
Overcoming this obstacle could involve some sort of neutral third-party model that a critical mass of rightsholders agrees to trust. This could happen through industry consensus borne of real-world experience, or it could happen by some sort of legal or regulatory action; neither is likely to be a short-term prospect.
AI Detection Flags AI-Generated Content
The other type of technology that is attracting a growing amount of activity is AI detection – analyzing content to determine the likelihood that it was AI-generated. This type of technology first appeared for text content, as an outgrowth of plagiarism detection services for schools and colleges such as TurnItIn and CopyLeaks.
Since the launch of ChatGPT in 2022, several startups – GPTZero, Winston AI, Originality.ai, Sapling and many others – have entered the market; and writing-support app Grammarly branched out to AI text detection too.
The primary application of this technology is still in schools, to detect student papers that were produced by ChatGPT or other AI text generators. Results have been mixed; some universities, including NYU, Vanderbilt and UCLA, have disabled them or discouraged their use due to false positives – students being accused of violating academic integrity policies by using ChatGPT to write papers when they actually wrote them themselves.
More recently, AI detection technologies have appeared for music. The Paris-based music service Deezer has developed its own technology for detecting music generated entirely by AI music platforms – as opposed to music created by humans with AI assistance. Deezer uses this to tag AI-generated music and to exclude it from algorithmic playlists. Deezer recently began marketing its AI detection technology for use by others; its first customer is SACEM, France’s analog to ASCAP and BMI.
IRCAM Amplify, a division of France’s renowned IRCAM music and acoustics research institute, has an AI Music Detector that claims 99% accuracy in detecting music produced by several AI platforms. Other providers of AI music detection include Pex, a leading provider of music recognition technology, and BeatDapp, a streaming fraud detection and prevention service.
AI detection is becoming a growing issue for music services and independent distributors due to the rapidly growing volume of AI content they receive daily and its connection with streaming fraud; but various other use cases are emerging: for example, a streaming music service may decide not to pay royalties on pure AI tracks, while a digital distributor may not accept pure AI tracks from its users for submission to streaming services.
Still more companies claim to detect AI content of other types. Hive and Reality Defender detect AI-generated content and deepfakes across multiple media types, including images and video. There are many others; the list grows on a regular basis.
AI detection and attribution are separate technologies, but they can work hand in hand. As the most obvious example, a music service might use AI detection to identify pure AI-generated music tracks uploaded to it and then feed them into an attribution engine to determine how to apportion royalties.
An Arms Race
AI detection is ultimately an arms race. Generative AI platforms are constantly tweaking their models – and ingesting more training data – to make their outputs increasingly difficult to distinguish from human content. And there are already third parties that claim to tweak existing AI output to help evade detection.
Still, these technologies don’t necessarily have to be completely accurate to be worthwhile. The criteria depend on the use case. For text AI detection in schools, for example, it’s much more important to eliminate false positives than false negatives; while in the music case, false positives may be less crucial when a human musician can appeal (a sign of human agency). Though we may not love it, we accept a world in which email spam detection is not 100% accurate.
And attribution technologies are always going to be estimates unless – perhaps – they are tightly integrated into each AI platform’s training and output algorithms. Yet it would be a mistake to assume that these arms races are lost causes as AI advances.
Echoes Of The Napster Era
We know this to be true because that’s exactly what happened 20-30 years ago. The Internet disrupted all content industries after it appeared in the 1990s. Technologies such as file-sharing (a la Napster) engendered a wave of development of copyright management technologies such as digital rights management (DRM), for encrypting content files to restrict their usage to individual users or devices, and fingerprinting (automated content recognition or ACR), for identifying content when a user uploads it to an online service.
These and similar technologies had their growing pains. Napster proffered ACR-based technology as a way to become copyright-compliant while defending the lawsuit that record companies brought against it; the judge in the case rejected the solution because it wouldn’t identify music with 100% accuracy. Years later, ACR schemes such as YouTube’s Content ID appeared; no one believes that they are completely accurate, but the industry has come to accept them.
Similarly, early DRM schemes were rejected for being cumbersome, overly restrictive, and glitch-prone. Yet today’s content services such as Spotify, Netflix, and Amazon Kindle use DRM to support access to content that hundreds of millions of users find convenient.
And the technologies developed during that era were subject to their own arms races. Developers of DRM technology were in arms races against hackers; this was so broadly anticipated that Congress passed a law in 1998 making it a violation of copyright law to hack DRMs. Today’s DRM technologies vary in their robustness against hacks, with those used for Hollywood video content – the most expensive to produce – generally considered to be among the strongest.
In all, today’s content services use a panoply of technologies for rights management that were designed to tame the initial chaos of the Internet and forged – for the most part – in the marketplace. Technologies that were considered bleeding-edge in the 2000s have adapted and are now part of the “plumbing” of today’s online services.
The same process is taking place now with the rise of generative AI. It’s going to be an exciting, dynamic world for content creators as well as distributors and technologists for the foreseeable future.