We’ll be the first to say that publishing and discussing AI news can lead to tremendous knowledge sharing and developments.
But, at what point does propagating a topic turn harmful?
In this month’s digest, we uncover the danger of model sentience conversations, how to implement AI ethics in your company, and a new framework built to measure the ethical performance of AI.
Thanks for joining the conversation!
Cal Al-Dhubaib, CEO
What Do We Mean by Trusted AI?
Trusted AI is the discipline of designing AI-powered solutions that maximize the value of humans while being more fair, transparent, and privacy-preserving.
What it's about: Blake Lemoine, an AI researcher at Google, was recently placed on leave after insisting the company’s conversational language model LaMDA is sentient. This article goes on to cover a more alarming issue: Focusing on sentience diverts creators’ and companies’ attention away from more pressing AI ethics concerns.
Why it matters: Conversations surrounding language models’ sentience are nothing new. The topic has been discussed as early as the 1960s with the primitive chatbot ELIZA, and has more recently been theorized by leaders like Elon Musk and Sam Altman.
Former Google Ethical AI team co-lead Timnit Gebru puts it best—Blake Lemoine is a “victim of an insatiable hype cycle,” filled with claims of humanlike cognition in machines. But the real danger in propagandizing model sentience is that we’re ignoring real risks of AI, like AI colonialism and false arrests.
While sentience might make for entertaining dinner conversation, leaders and data scientists must instead focus on identifying and mitigating realistic indicators of model bias—ones that may end in serious consequences if we aren’t careful.
Why it matters: Designing, piloting, and iterating AI models is an investment that requires support from a diverse team of stakeholders—not just one or two technologists. And as a company leader, mitigating the risk of negative bias and discrimination in your AI must be a top priority.
While issues like fairness, transparency, and privacy should be core considerations of any AI, your organization will also want to identify and evaluate risk tied to your specific AI use case. For example, creators of self-driving cars must also assess the risk of pedestrian safety, while creators of chatbots must evaluate model outputs for inappropriate responses.
Addressing the ethical implications of AI hasn’t always been at the forefront of business agendas. But, as we see more regulations put in place and organizations held accountable for the unintended consequences of AI, it’s evident that designing trustworthy AI is imperative.
Department of Defense Launches Responsible AI Strategy
What it's about: The United States Department of Defense recently released a 47-page document detailing its plans for the use of responsible AI. The document is broken into six foundational tenets of proposed action: Responsible AI Governance, Warfighter Trust, AI Product and Acquisition Lifecycle, Requirements Validation, Responsible AI Ecosystem, and AI Workforce.
Why it matters: Explainability is a key guiding principle in the design of trustworthy AI solutions. Although the United States still lacks federal AI regulation, a strategy offering this level of transparency into intended action is significant.
The Department of Defense’s responsible AI document underscores the importance of a strategy-first, design-second mindset. This approach can prevent unwanted consequences of artificial intelligence—and when a nation’s federal defense systems are involved, avoiding negative outcomes is critical to the safety of its citizens.
The AI strategy outlined in this document also sets performance goals for each of the six tenets of action. These goals give us clear insight into how the government will measure the success of their artificial intelligence solutions. And when citizens are able to understand the use cases and performance metrics of AI solutions like those laid out in this strategy, we stand a better chance of holding our government accountable to protecting our privacy and security.
Singapore Launches A.I. Verify Framework and Toolkit
What it's about: Singapore’s Minister of Communications and Information launched one of the world’s most robust AI governance frameworks, called A.I. Verify. Organizations can use this framework and its associated tools to measure the ethical performance of their AI against standardized solutions. The reports generated by the toolkit are designed to promote transparency and accountability among businesses and their stakeholders.
Why it matters: From keeping humans in your feedback loop to monitoring input data, there are a number of ways that data scientists and leaders can measure AI for bias.
Developments like the A.I. Verify framework and toolkit make it that much easier for organizations to prioritize fairness, transparency, and privacy in AI. While only available as a minimum viable product for now, it’s a promising step towards more accessible tools for minimizing potential negative consequences of AI.
While most AI-ready leaders today understand the importance of designing trustworthy AI, many require support when it comes to evaluating a solution and communicating the changes that need to be made. The comprehensive reports generated by A.I. Verify (or any trusted third-party) can ultimately lead to improved transparency and communication among key stakeholders.
What is the difference between "real" and "synthetic" data, and is one more trustworthy than another?
When we begin the discovery process, one of our first steps is identifying a dataset that we can use to train and prototype our models. Ideally, this data should be:
Representative of the eventual production data.
Rich enough to provide features to model with.
Sufficiently large to enable training of ML models.
Ideally, we have "real" data that matches these criteria. However this often isn't the case; either the real data is too small to train with, or perhaps the use case is so unique that no appropriate datasets exist. In this case, we turn to our backup plan: synthetic data.
Synthetic data looks just like real data, only it's been artificially generated. These datasets are created to have the same characteristics we expect to see in production, and allow us to train and evaluate models long before production data begins rolling in.
High quality synthetic data can then be almost indistinguishable from real data, while enabling early analysis that might otherwise be impossible. Meanwhile, "real" data still has pitfalls; if user patterns change before a model is deployed, even real data can fall flat.
At the end of the day, there's good and bad "real" data, just as there is good and bad "synthetic" data. So whether you have it, find it, or make it, as long as it's representative and reasonably large, your models should train happily on either type.
