Discerning Authenticity and Protecting IP are Crucial Challenges in the AI Era

In 2022, an AI-generated artwork won the Colorado State Fair’s art competition. The winning artist, Jason Allen, used an AI image generator called Midjourney that turns lines of text into hyper-realistic graphics. Unsurprisingly, their use of AI to win the art competition triggered a backlash online.

A more recent example of AI-powered art involved a German artist who submitted a “photograph” generated by AI to win a prestigious international photography competition. However, the artist rejected the award in this case, arguing that “AI images and photography should not compete with each other.” The artist said he hoped the outcome would prompt more discussion on the issue of artificially generated images.

The impact of generative AI has spread beyond the bounds of making digital art. In April 2023, an anonymous TikToker dropped a “fake Drake” track that went viral on TikTok and YouTube. Even if virality is a short-lived phenomenon in the attention economy, the success of the fake track only amplified concerns over the inability to distinguish artificial from authentic content, stoked fears over the use of copyrighted works without owners’ consent, and raised questions about how to judge and value art created using generative AI tools like Midjourney.

Underlying these concerns is a central issue at stake for the Media & Entertainment industry with the advent of generative AI: ensuring that investment in Intellectual Property (IP), arguably the backbone of the industry, continues to be protected and rewarded. It is a thorny problem that deserves as much attention from everyone in the industry as has been given to the promise of generative AI to supercharge creative innovation and empower and scale humans. In this article, we take a deeper dive into how generative AI can be used to both discern and cloud artistic authenticity and discuss how we might mitigate risks to IP investment from this technology.

Using AI to Discern Authenticity

Due diligence plays an essential function in the traditional art market and comprises four core pillars: provenance, condition, clear title and fair market value, and authentication. While some might conceive of AI as a threat to the art world’s due diligence and forensic specialists, AI can significantly help in this area.

A Swiss company, Art Recognition, which made headlines in 2022 and 2023, offers an interesting example in this context. The company used their AI-powered technology to determine that a previously unknown painting was by Italian artist Raphael, that the contested Portrait de femme (Gabrielle) had a high probability (80.58%) of being painted by Renoir, and that a Titian work held in Switzerland’s Kunsthaus Zurich museum was most likely a fake.

Art Recognition’s advantage as an AI technology firm lies in its ability to leverage its dataset of images to identify the hallmarks of various artists by examining brushstrokes, edges, shapes, color variations, and composition elements at a speed and scale that far exceeds what is humanly possible. However, while AI is a useful tool to supplement forensic techniques used by experts in the field, in my view AI alone should not entirely displace a human expert and does not serve as a substitute for human judgement in the art authentication process.

Brushstrokes Analysis

The ability of AI models to determine whether an artwork is authentic or fake by analyzing brushstrokes is an especially powerful component of AI’s promise in this field. This approach has been used successfully in several high-profile cases, exposing fake artworks that had tricked even the most seasoned experts.

In 2017, for example, researchers at Rutgers University developed an AI system that was able to correctly identify the artist for a given painting in 80% of cases by studying unique patterns in brushstrokes, outperforming art historians.

A similar project led by scientists at Case Western Reserve University used machine learning to identify a distinctive ‘fingerprint’ in each artist’s specific way of applying paint. The research shows how machine learning analysis of small sections of topographical scans of paintings can match artworks to the correct artist with up to 96% accuracy. Kenneth Singer, a physics professor at Case Western Reserve who led the research, thinks that the next application for AI could be on media with less surface texture than paintings, like watercolors or drawings.

X-Ray Image Analysis

AI can also be used to authenticate art by analyzing infrared and X-ray images of artworks. These images can have hidden layers that can provide clues about the artwork’s authenticity. By examining X-ray images of art, AI algorithms can identify features and clues that may suggest a forgery. For instance, inconsistencies in the artist’s signature or use of materials that were not available when the artwork was thought to be created can be quite instructive.

I am curious to see if AI can go further and help experts analyze chemical compounds and pigments used in paintings to understand more about an artist’s creative process. For instance, in a very recent case of rare chemical compounds discovered in the “Mona Lisa,” the researchers used X-ray and infrared microanalyses. AI could potentially help us uncover similar findings faster and on a greater scale.

While the application of AI to assist in due diligence for art looks promising, the technology also has its limitations. AI models are only as good as the paintings, images, and data they are trained on. If those paintings and images are fake, AI-generated, or contain areas that have been touched up, the training data could lead AI models to mischaracterize artists’ traits and misidentify some artworks as authentic or fake. A similar problem relates to AI-powered valuation models for artworks based on data from past art auctions. The data sets used to train such models may not be comprehensive enough for the models to make accurate predictions or may be biased toward certain artistic styles and movements, causing the models to undervalue underrepresented artists or styles. Hence, the quality, scope, and nature of the training data are vital to how AI models learn about the content they are being trained in and generate their outputs, whether they be new digital artworks or music tracks, spotting fakes from real art, valuing artworks, and so on.

The Risks of AI to IP Investment and Ownership

While AI image generators like Midjourney and OpenAI’s DALL-E have impressed users with their ability to create dazzling images from natural language prompts, these systems raise tough questions about the ownership and authenticity of AI-generated content, the consent and disclosure around their use of copyrighted work, and the risks they pose to IP investment.

The training data behind generative AI models may include the work of artists who have yet to license their work or may use copyrighted work without consent from the copyright owner. A lack of awareness or consent from the IP owners around the use of their works by these systems is a big problem, both legally and ethically. For this reason, in January 2023, three artists filed a lawsuit against three generative AI platforms: Stability AI, Midjourney, and DeviantArt. The artists claimed that the companies were using copyrighted images to train their AI algorithms without their consent. In the very same month, Getty Images also filed a lawsuit against Stability AI over alleged copyright infringement.

Ownership of AI-generated artwork is also in question: If someone uses an AI tool to create their own art, to what extent can they claim full ownership of that artwork, and if they cannot, how authentic is that artwork? And can or should it be copyrightable?

Currently, there is no copyright protection for works generated solely by a machine under US Copyright law. Copyright may be granted, however, in cases where the creator can prove considerable human involvement; ultimately, copyrighting an AI-generated work remains a grey area subject to human interpretation and judgement.

Another important concern is the large-scale dissemination of AI-authored content, exacerbating the already significant problem of digital misinformation. AI tools offer scammers, con artists, and criminals a powerful and effective way to create artificial content or false information — articles, voices, images, photos, videos, songs, artworks, etc. — in the likeness or the style of the original creators that can be difficult to detect as fake or false. Besides the deliberate misuse of AI tools for nefarious purposes by such actors, authenticity gets crowded out as AI-authored content can be produced much faster than purely human-authored content. For example, according to Everypixel Journal, more images have been produced using AI in the last year than photographers have taken over the previous 150 years. This crowding-out problem dilutes creative authenticity and risks reducing the value of existing IP and undermining future investment in IP.

Mitigating the Risks and Possible Solutions

With the US Copyright Office (USCO) ruling that humans can establish copyright on AI-generated works — conditional on a certain level of human authorship, of course — it has become possible for artists to seek copyright protection for their AI artworks. For example, in September 2022, the USCO granted a first-of-its-kind registration for the comic book ‘Zarya of the Dawn’ generated with the help of Midjourney. However, after a subsequent review of that decision, USCO limited the copyright protection to cover only the text and arrangement of images in the book, not the individual images within it. The case offers a way forward, but it also highlights both the complexities of seeking copyright protection when AI is used in the creative process and how AI is disrupting the field of copyright protection.

Some of the leading generative AI tool providers are also working on solutions to address authenticity concerns and how to clearly identify generative content. Adobe’s Content Authenticity Initiative (CAI) is one such solution. The CAI combines cryptographic methods and a trust list to provide a transparent, tamper-evident history of digital content. Users can capture authorship and provenance information, either embedded into the image file or stored in the Cloud. When a user views a piece of CAI-enabled content, they can access this information to understand its provenance better.

The CAI has limitations, such as reliance on supported tools, placing the onus on creators to tag content, and the possibility of stripping the CAI data from a file, limiting its transparency and provenance. However, with a well-defined regulatory framework, it could offer a feasible and workable solution.

Since generative AI models depend on training data that often includes copyrighted content, it is also critical for creators to be compensated for using their content to protect and preserve the value of their IP. At a macro level, this is likely to involve partnership deals between content creators, owners and publishers, and large AI platform providers like Open AI (a case in point, the Financial Times reported earlier this year that media groups were trying to negotiate terms with AI platforms to get paid for their content used to train AI models). At a micro level, it means devising a model that can account for attribution back to the IP owners. A generative AI tool recently launched by Getty Images uses a creator compensation model based on two factors: (1) What proportion of the training data set does a content owner represent, and (2) How well has that owner’s content performed historically in their licensing world? It will be interesting to see how other AI tool providers tackle this problem, whether there will be convergence towards a common approach, and how such compensation models might evolve over time.

Concluding Thoughts

There is an odd paradox in the way generative AI tools can enhance our ability to discern authenticity (as in the context of authenticating artworks) while, at the same time, by disrupting the creative process generate more uncertainty over what is truly authentic and deserving of copyright protection. There is a need to ensure that generative AI systems offer a transparent and reliable way to capture and preserve provenance information so that generative content can be easily identified and IP owners can be compensated for using their work in the training and outputs of these systems. The stakes are high: Without adequate solutions to these problems, we risk lurching into a world of AI where content can be created at unprecedented speed and scale, but investment in IP is no longer valued and rewarded.

Author: Doron Fagelson,
Vice President of Media and Entertainment Practice at DataArt
Originally published on https://www.dataart.com/blog/