The Trust Consortium
on Generative AI
Establish trust in Artificial Intelligence by building a practical mechanism to identify Generative AI content.
Our Call
Create a hash-sharing consortium in order to identify and establish trust in AI-generated content.
Develop a system built on tried and tested methodologies previously used to identify and stop online abuse and digital copyright infringement
Industry collaboration must underpin this consortium. Systemic approach will only work if all industry-players work together to build common standards.
In order to prevent generative-AI misuse by harmful actors: criminal theft, terrorist and violent extremists, state-actors pursuing disinformation.
Introduction
The transformative power of Generative Artificial Intelligence and Large Language Models is exciting and terrifying in equal measure.
When it comes to misuse, the threat is palpable. Be that by criminals infringing copyright, terrorists quickly accelerating the spread of propaganda or nation-states disrupting democratic processes such as the US elections.
While it is now almost impossible to stop the juggernaut, steps can be taken to establish firm guardrails in these, the early days of the generative AI revolution.
While there are already calls for regulation in this space, we propose a practical and industry-led initiative that focuses on a practical solution based on known, tried and trusted technologies: hashing and similarity-matching.
We specifically focus on content-generation for synthetic media: be that images, video or sound.
We propose a mechanism that is robust, economically sustainable and industry-wide to effectively challenge generative AI misuse.
WHAT WE PROPOSE
A Database, powered by an industry-led consortium, to identify Generative AI content
We call for generative AI providers to create unique content identifiers or "hashes" for the content they generate and share these with an independent database.
The database would be hosted and run by an independent and international body and would act as a ‘clearing house’ for AI generated content. It would then become a valuable tool, fostering an ecosystem resilient against disinformation.
When it comes to images – a specific area of concern – the database would adopt advanced methodologies of "similarity matching": technologies that identify similar images, even if they have been slightly altered.
Users of the database would include social media companies who would scan uploaded content against a comprehensive database to identify matches, flagging potential terrorist and violent extremist or state-crafted disinformation.
The initiative would be economically sustainable by also offering a paid API for hash and lookup services. Use-cases could include tagging and flagging AI-generated content on respective platforms. This could, for example, give advertisers confidence that their products are not featured alongside misused AI-generated content.
Taking lessons from the enforcement of digital copyright and hash-sharing databases, such a system and initiative would need to leverage the collective wisdom and experience of the industry.
Crucially, this mechanism would be underpinned by a consortium consisting of major players such as OpenAI, Microsoft, Google, Meta, and other Generative AI providers. This collective effort could pave the way for innovative start-ups to contribute fresh perspectives and agile solutions to the broader initiative.
€30m or 6% turnover
Proposed EU fines for noncompliance with the prohibition of artificial intelligence practices
$1.8 trillion
Getty Images suing Stability AI in legal damages for copyright violation
10%
Of Fortune 500 enterprises will generate content with AI tools in 2023
Why we need this mechanism now
PREVENT MISUSE AND DISINFORMATION
While at its infancy, the power of Generative AI, both AI companies themselves and policymakers around the world have raised red flags about the misuse of this new technology.
SECURE PROACTIVE INDUSTRY LEADERSHIP
By establishing an independent database and mechanism to prevent misuse, Generative- AI impage producers and platforms must own the problem. Ahead of punitive measures by regulators, for example through the European Union, tech companies must be proactive.
UPHOLD TRUST IN CONTENT MODERATION
Content moderation is a key pillar for tech platforms as they seek to establish trust in the services they provide. The rapid-rise of Generative-AI risks upending this carefully developed process.
Taking Lessons Learned
The mechanism we propose takes important lessons from the enforcement of digital copyright and the deployment of hash-sharing databases designed to protect children from exploitation.
Digital copyright
Digital copyright exploitation has taught us invaluable lessons. What is needed now is a robust, economically sustainable, industry-wide initiative that borrows from these lessons to pre-emptively tackle the challenges of Generative AI misuse.
Watermarking, a well-established technique in digital asset protection, could be implemented for images produced by generative AI solutions. By embedding a unique identifier into the image's metadata or the image itself, AI-generated content can be readily differentiated from human-crafted imagery.
In addition, content ID systems, akin to those used by platforms like YouTube for copyright protection, could be repurposed to combat the misuse of Generative AI.
Harnessing the experience of these systems would speed up the detection and removal of potential terrorist and violent extremist content or state-sponsored disinformation.
Databases to alert and verify
The creation of hash-databases for AI-generated images and text represents a crucial countermeasure. Yet, classical hashing techniques can fall short, vulnerable to minor manipulations in images.
We therefore suggest the adoption of more advanced methodologies of "similarity matching", used in Microsoft's PhotoDNA (PDNA) and Facebook's PDQ algorithm. These technologies can identify similar images, even if they have been slightly altered, equipping us with more potent tools in the fight against disinformation.
An independent body could offer a paid API for hash and similarity lookup services. This would not only serve as a valuable tool in the fight against disinformation but also ensure the system's self-sustainability. This approach takes a leaf from the book of existing hash-sharing initiatives, like the one run by the National Center for Missing & Exploited Children (NCMEC) for Child Sexual Abuse Material (CSAM) detection.
We would also draw lessons from Tech Against Terrorism’s Terrorist Content Analytics Platform: a system which is becoming the world’s largest database of terrorist and violent extremist content and is fast becoming an essential tool for counterterrorism.
PROPOSED BY
The Online Harms Foundation provides tech companies the tools and resources they need to effectively counter online harms on their platforms.
We implement the initiative Tech Against Terrorism, a UN-supported public-private partnership to disrupt terrorist and violent extremist exploitation of the internet.
The Online Harms Foundation works with democratic states and regulators to guide and advise on relevant legislation ensuring that tackling online harms respects human rights and fundamental freedoms.