Generative AI presents very big opportunities but also carries many risks. We review the current risk mitigation strategies and explore where we could go next as an industry where all players must play their part to achieve the best potential of this technology.
Benefits of Generative AI
Since 2022, there has been a 1000% increase in the use of Generative AI content on the internet. Some say the AI could be the biggest game changer in our fight against hunger and poverty while others claim that it could widen the gap between rich and poor. Time will tell and with AI, time moves faster than other technologies we have experienced before - in just a few months, many industries have already embraced Generative AI at different depths to introduce enhanced products. For example:
1. Using publicly available pre-trained AI Models: Grammarly, SciSummary and Legalyze are some of the companies using Generative AI to enhance their products with better interfaces and newer capabilities.
2. Self-trained AI Models: BloombergGPT , Meta’s SeamlessM4T, Tabnine are some of the examples where companies have spent a lot more resources to train generative AI models on their own custom corpus to build tools which function better in their domains.
3. Generative AI for Audio, Video and Imagery: Meta’s Audiocraft, Eleven Lab’s Dubbing Studio, and Video generators from Synthesia and Deepbrain are some examples.
Risks
Clearly, Generative AI has numerous benefits. So, what is the conundrum here? While Generative AI can solve many hard problems, it also presents a clear and present danger to our society and way of life: Generative AI has been become the killer app for spammers, fraudsters, and misinformation spreaders. It has become a winning tool for politicians, especially where winning is important but extremely difficult or expensive due to regulatory environment. Several recent examples have emerged:
1. An AI generated voice recording on Facebook was used to disrupted elections in Slovakia.
2. Politicians in Argentina used AI tools like Midjourney to generate fake images of rivals to influence voters. Politicians also claimed to be victimized by attackers that used Generative AI when in fact the information was as real and credible as it could be.
3. Robocalls using an AI voice resembling President Joe Biden urged democratic primary voters to skip.
Current Solutions
Social media platforms have resorted to a mix of solutions ranging from automated detection tools to crowdsourced content moderation and fact checkers. For example:
1. Meta uses fact checking organizations like Demagog and StopFake
2. X uses CommunityNotes, an approved set of volunteers.
However, research shows these solutions have much lower efficacy in the political domain. First, the speed of response at the required scale does not match the viral speed at which the attack spreads. Second, seeking out consensus across divergent perspectives in a polarized political environment can be a daunting challenge for fact checkers too. The links I provided in the above examples in fact indicate many such limitations.
How do we address the gaps in 2024?
What can we do to address the possibility that there could be much more magnified versions of these attacks that may undermine political events in 2024?
The approach to solving this must be multi-pronged.
Step 1: Improved legislative stance against tech and media companies to increase ethical accountability.
The general technology for Generative AI is mostly open sourced and well known in the research community. However, tech and media companies are the dominant enablers of such technology to the masses. By making it harder for the tech companies to ignore their ethical responsibilities, we can build a strong foundation for a comprehensive solution. The legislative actions taken in Europe, US, Canada and many other countries on AI Governance, Risk Framework and Compliance are very good examples of how this is being achieved.
Step 2: Use technology to ensure fast and accurate detection of Generative AI in digital content.
Generative AI content is mostly produced and consumed inside applications and devices that have capabilities to mark and verify that content. Examples:
1. Applications that generate media like photoshop, Audiocraft and Synthesia.
2. Social networks that host media like Facebook, Instagram and X.
3. Applications that consume media like Web Browsers, Youtube and other mobile apps
The next step would be to improve how these applications can secure digital content and alert recipients as quickly and reliably as possible to the presence of generative AI. Here we can use at least three techniques:
1. Watermarking: Watermarking involves hiding a label, sound or text inside the media and obscuring it so it is hard to detect. Such techniques are much more effective with images and video than text and can be used in applications that generate or modify media. There is absolutely no harm in generating watermarks within these applications. However, since generative AI is mostly open source, hackers and spammers can circumvent watermarking by using alternate services/training their own models.
2. Pattern Detection: Tools like SynthID and GPTZero use pattern detection to identify generative AI content like text and images. Such tools can be integrated in Social Networks and Browsers. Care must be taken to constantly improve these tools and account for errors as Generative AI will continue to evolve and become better and evading detection.
3. Coalition for Content Provenance and Authenticity (C2PA): C2PA is the most promising option as it presents a fast, reliable, and yet accurate way compared to other options. C2PA can be used to add cryptographic information to detect the tampering of media files and streams. C2PA is an open standard that focuses on providing the history and context for any digital media. This technology is very similar to X.509 certificates used by websites and payment services. Just like a certificate owned by a website allows a browser to verify the authenticity of the website and its owner, C2PA helps consumers identify who created a digital file and its complete modification history. Examples of applications that already implement C2PA include Adobe Photoshop and a Chrome extension released by Digimarc that can help consumers check images. C2PA can also be used along with blockchain technologies to certify the authenticity of images for news organizations.
C2PA has built-in safeguards (including PKI-based digital signatures) to ensure that the authorship or origin information is accurate and can’t be faked or falsified. C2PA has been available as early as 2022 and but its adoption is still very weak. Government intervention may be an option to accelerate adoption across the industry.
In conclusion, while Generative AI presents extraordinary benefits, it is imperative to address its associated risks through comprehensive strategies involving technology and policy in addition to just public awareness.
