Three Reasons to Create Supply Chain Transparency for Generative AI Now
Inspirations From Germany’s Supply Chain Act
Transparency in Physical Supply Chains
At the beginning of this year, a new law went into effect in Germany: Lieferkettensorgfaltenpflichtgesetz. What reads like a word for the next hangman game with your kids can be easily translated into English as Supply Chain Act. Its basic premise is for businesses which have more than 3,000 employees based in Germany to ensure that risks of human rights violations and environmental impact within their supply chain are minimized or eliminated. It is a good example of what is becoming more and more relevant in AI as well: Transparency across the AI supply chain. There are three key areas to highlight and what AI leaders exploring generative AI scenarios need to know…
» Watch the latest episodes of “What’s the BUZZ?” on YouTube or listen to it wherever you get your podcasts. «
1) Raw Material: Data
To produce any physical good or product, a business uses raw materials. For AI-driven systems, this raw material is data. Generative AI tools such as ChatGPT and MidJourney have been soaring in popularity, in part due to their ease of use, creating text or images based on plain-language instructions. The underlying AI models have been trained on publicly available data that’s been scraped off the internet (e.g., details in this article). But recent events such as the Italian data protection authority filing a probe into OpenAI ChatGPT, discussions about fake images such as the one of Pope Francis wearing a puffer jacket, a high-profile lawsuit against text-to-image AI tools, and popular Internet forum Reddit charging for programmatic access to their content are catapulting one topic into the limelight: Data.
What data has been used to train the model?
Where does it come from?
What biases, inaccuracies, and misinformation is included in the training data?
How does this data influence the language of the generated output (falsehoods, hate speech, etc.)?
Does the use of this data for training violate intellectual property rights?
What data is being collected about the users?
Have people been informed upfront that their data will be used?
Have they given their consent?
Why it matters for AI leaders: In a business environment, your company is responsible for the information it produces and publicly shares. While generative AI is still an emerging field, the reputational, legal, and financial risks from regurgitating inaccurate or biased information are real — for example a generative AI based chatbot on a company’s website responding with racial slurs.
AI leaders exploring which generative AI scenarios to pursue, should prioritize those that have a low risk to the business. These are currently scenarios in which a human reviews the AI-generated output, before accepting or using it — for example, asking AI to create a draft of a sales e-mail or press release.
2) Ethics: Work Conditions
Another critical dimension is ethics, including the work conditions along the supply chain. In the case of generative AI, TIME reported in January the mental health impact that content moderation for large language model (LLM) training has on contractors in Kenya. Although human reviews of flagged data and content are an additional safety mechanism in building LLMs, the human cost of achieving this level of safety needs to be taken into account more broadly. Industries such as consumer products and fashion retail have longe established certifications to highlight supply chain transparency (e.g., fair trade, child-labor free production). Similar certifications for work conditions of creating generative AI models responsibly could emerge in the future.
Why it matters to AI leaders: To date, AI leaders have largely been building AI models based on their own business’ data or third-party data. However, incorporating LLMs and other foundation models into products adds additional responsibility for AI leaders and the business they work for. Both, in aspects of generating highly accurate output as well as of having transparency how the technology components they include in their products have been built. Although this is currently not yet top of mind for many, the growing, global scale of generative AI and the plethora of use cases and products it is going to be being used in, will likely bump this topic up on the corporate agenda.
3) Environmental impact: Energy consumption
Training and hosting LLMs and other foundation models consumes a significant amount of electricity for running the cloud infrastructure that these workloads run on. At a time when environmental and sustainable goals are on the agenda for many businesses, using or even training LLMs is in stark contrast to environmentally friendly practices. For example, Stanford University have shared in their latest annual AI Index Report that training an LLM produces 25x the emissions of a single person’s flight from New York to San Francisco. While not every business that uses LLMs also automatically has a large carbon footprint, incorporating LLMs does come with increased resource requirements, including power consumption. Consequentially, greater AI supply chain transparency will also include a calculation and publishing of produced carbon emissions.
Why it matters to AI leaders: For years, companies have been offsetting carbon emissions of their business travelers by supporting reforestation and clean-air projects. Given the environmental impact for generative AI, AI leaders need to play a key role in reducing overall emissions in the sourcing and delivery of AI that they include in their products. Working with smaller, less resource-intensive models (that have fewer parameters) could help further reduce the environmental impact in the future, for example.
Summary
Similar to supply chain transparency for physical goods and products, building transparency into the generative AI supply chain will become a growing demand. Whether it is the data itself, the working conditions under which it is being curated, or the environmental impact of training large models, AI leaders are best positioned to make the choices today that will create greater supply chain transparency and that allow them to also use it as a distinction tomorrow.
What other dimensions would you add?
PS: » The last 4 weeks have shown us: We have learned a lot as an industry. The next 4 weeks will teach us even more. « These were the opening lines when I started the draft for this post at the end of January, shortly after the TIME article had been published. Four months into 2023, we have not been disappointed — and the call for AI supply chain transparency is getting more imminent.
» Watch the latest episodes of “What’s the BUZZ?” on YouTube or listen to it wherever you get your podcasts. «
What’s next?
Join us for the upcoming episodes of “What’s the BUZZ?”:
April 25 - Ramsay Brown, Founder & CEO Mission Control, will be on the show when we talk about How Businesses Can Trust Generative AI in times of rapid innovation.
May 9 - Brian Evergreen, Founder & CEO The Profitable Good Company & Author, will discuss how manufacturing businesses can Create A Human Future With AI.
June 8 - Ravit Dotan, Director The Collaborative AI Responsibility Lab at University of Pittsburgh, will join when we cover how responsible AI practices evolve in times of generative AI.
Follow me on LinkedIn for daily posts about how you can set up & scale your AI program in the enterprise. Activate notifications (🔔) and never miss an update.
Together, let’s turn hype into outcome. 👍🏻
—Andreas