What becomes possible when you can trust how AI behaves?

+ Microsoft's humanist AI, our CISI beta launch, and World Summit AI

There is certainly no shortage of AI news to discuss at the moment. But for this first note, we want to focus on a question that has come up repeatedly in our recent conversations: what becomes possible when you can trust how AI behaves?

Responsible AI is often discussed primarily in terms of managing risk. We think that misses a significant part of the opportunity. 

Every couple of months, we’ll share some of what we are seeing and learning through our work, alongside developments in AI that have important implications for businesses.

01 — THE LEAD (FROM OUR CEO, MAURY SHENK)

What becomes possible when you can trust how AI behaves?

Consider an organization that wants to use AI to give advice to its customers.

A mass-market AI model may be perfectly capable of understanding a question and producing a credible answer. But that does not mean the organization is ready to put it in front of customers.

It may need the AI to interpret questions according to a particular professional framework, apply the organization's own policies and standards, recognize situations in which it should qualify or decline an answer, and behave appropriately when it encounters something its designers did not anticipate.

And, crucially, the organization needs some way of knowing whether it is actually doing those things.

Establishing that confidence changes the question. Instead of asking only whether AI can perform a task, organizations can begin to ask what new applications or services become possible when they can trust how it will perform it.

Achieving that confidence is not straightforward. Large language models do not arrive as neutral building blocks. They already exhibit behavioral tendencies shaped by their training and by decisions made during their development. Those defaults will not necessarily correspond to the standards of the organizations deploying them.

Prompts and instructions can shape behavior, but they do not by themselves provide assurance. AI systems are non-deterministic. A satisfactory response to one test does not guarantee an appropriate response when the wording, context or circumstances change.

For an organization to have meaningful confidence in an AI application, it therefore needs to be clear about what appropriate behavior actually means, have ways to control that behavior, and be able to evaluate what the AI does against those requirements.

At Ordinary Wisdom, we think about this in three broad stages: define, control and evaluate.

01    Define 

The behavior required for the particular organization and application, drawing on its policies, values, professional standards and regulatory obligations.

02    Control

How the AI behaves against those requirements.

03    Evaluate

Its actual outputs against them, not simply once before launch, but on an ongoing basis.

We deploy a patent-pending pipeline of multiple machine learning and data science techniques across the three stages.

There is an obvious risk-management benefit to doing this. But the more interesting consequence is what it can make possible.

An organization that would not be prepared to deploy a general-purpose AI system in a judgment-heavy environment may be able to deploy one whose behavior has been deliberately shaped around its requirements and can be evaluated accordingly.

That could mean introducing AI into a workflow that previously depended on human judgment. Or it could mean creating an entirely new service that would not otherwise have been practical to offer.

For organizations in regulated industries especially, the ability to establish that confidence can determine which AI applications are viable. The opportunity from responsible AI is not simply to make applications they are already considering safer. Greater control and confidence can expand the range of AI applications they are able to build and deploy in the first place.

So perhaps the more useful question is not simply, how do we reduce the risks of using AI? It is:

What could your organization do with AI if you had greater confidence in how it would behave?

If you have an application in mind, we’d be interested to hear about it.

Maury

02 — WORTH READING

Microsoft's 'Humanist AI' and Values

Microsoft’s new “humanist AI” approach puts human values at the center of its vision for AI. Maury looks at what this means in practice, and why the question of whose values AI should reflect matters.

Read more →

The Economist on AI Model Values

A recent briefing from The Economist explores how the values expressed by AI models can differ significantly from those of people, reinforcing the importance of values in how AI systems are assessed and governed.

Read more →

03 — FROM ORDINARY WISDOM

04 — PLATFORM UPDATE

Our first major platform update since launch

On Monday, we’re releasing the first major update to the Ordinary Wisdom platform since our launch in June.

The release makes it easier for organizations to evaluate, compare and manage how their AI systems behave. It includes significant improvements to LLM output evaluation, including a new capability to compare two configurations side by side, alongside a simpler and more intuitive user experience.

We’ve also improved the management of document portfolios, constitutions, tests and other data elements, as well as making updates to our API and fully revising the platform documentation.

It’s an important step forward as we continue to develop the platform through beta, shaped by what we’re learning from real-world use cases.

Explore Ordinary Wisdom →

Heading to Amsterdam for

World Summit AI?

Our CEO Maury will be at World Summit AI in Amsterdam from October 6-8, and is particularly interested in speaking with organizations that see an opportunity to use AI but aren’t yet confident they can deploy it with the control they need.

If you have a specific use case in mind, let’s discuss what would need to be true to make it possible.

Not attending? Let's discuss your use case →