Featuring:
Brandon Purcell, Vice President and Principal Analyst
Show Notes:
The term “AI misalignment” might not be familiar to many business leaders yet, but Vice President and Principal Analyst Brandon Purcell says it may be the solution to an “existential threat” that many businesses face in the AI era. In this episode, Purcell discusses the risks of misaligned AI and outlines a solution called “align by design.”
The conversation begins with Purcell defining the concept of AI misalignment and outlining its potential causes and risks. He says misalignment stems from the fact that AI’s effectiveness is contingent on the quality of the data its models are trained on. And in most cases, the data used to train the models comes from the internet, which may not accurately reflect reality. That misalignment can lead AI astray, with potential for serious business consequences.
Are businesses aware of this threat? Purcell says that most often, he is the one raising the issue of AI misalignment with business leaders who are more accustomed to addressing topics such as AI governance and responsible AI. While AI alignment includes risk mitigation and governance, it is ultimately focused on achieving intended business objectives (more carrot than stick, as Purcell puts it).
From there, Purcell reviews three types of AI misalignment: outer misalignment, where the AI system’s objective doesn’t exist as a discrete variable in the data; inner misalignment, where AI learns an objective but over time learns an additional unintended objective; and user misalignment, where users pull AI out of its intended objectives. Purcell provides real-world examples of all three types of AI misalignment to help illustrate the risks and addresses how AI misalignment can impact agentic AI specifically, given how quickly that type of AI has grown recently.
What’s the answer? To minimize these risks, Purcell details the “align by design” approach, which is a proactive approach to designing AI to meet your intended objectives while adhering to desired guidelines and standards. This approach leverages Forrester’s seven levers of trust to build trust and ensure that AI systems do what you want them to do, not what you don’t want them to do. After describing how the accountability and empathy trust levers are used within the “align by design” approach, Purcell drills down into the transparency lever, describing how AI transparency can vary between types of AI (agentic vs. generative, for example) and will be a key differentiator for firms leveraging AI in the future.
The episode closes with Purcell providing two predictions about AI misalignment. First, he predicts that three to five very large examples of AI misalignment will make headlines, including some that include generative and agentic AI. “I also think that companies that take an ‘align by design’ approach … will actually see a quicker ROI from their AI investments,” he says.
