We need to build better bias in AI

You can check out all the on-demand sessions here.

Artificial intelligence systems help us to realize our goals. They undermine them at their best. We have heard of high-profile instances of artificial intelligence bias, such as Amazon's machine learning recruitment engine that discriminated against women or the racist results from Goggles. The cases work against their creators' intentions. As a result of these examples, the public perception of artificial intelligence was shaped into something that was bad and we need to eliminate it.

It's unrealistic to take all bias out of an artificial intelligence system. The new wave of models are being designed with some level of subjectivity in mind. Translating results is one of the most sophisticated systems of today. Organizations shouldn't try to eliminate bias entirely.

When it comes to the goals of the systems, organizations require them to be subjective in a way that meshes with the project's intent.

It's clear that this is in the field of artificial intelligence. Speech-to-text systems can now be used for calls and videos. The emerging wave of solutions report speech but also summarize it. Rather than a simple transcript, these systems work with humans to extend how they already work by creating a list of actions after a meeting.

The intelligent security summit is on demand.

Understand the role of artificial intelligence and machine learning in case studies. You can watch sessions on demand.

The system needs to understand context and interpret what is important in these examples. Artificial intelligence is being built to act like humans and subjectivity is an important part of the package.

The business of bias

The technological leap that took us from speech-to-text to conversational intelligence in just a few years is small compared to the future potential of this branch of artificial intelligence.

Professor Albert Mehrabian stated in his seminal work, Silent Messages, that meaning in conversation is most often conveyed through non-verbal signals. The words themselves make up less than 10% of the total. The majority of conversation intelligence solutions rely on interpreting text.

We might call what these intelligence systems are interpreting the Metadata of Human Conversation. That is, tone, pauses, context, facial expressions and so on, bias is not only a requirement but also a value proposition.

Conversation intelligence is an example of a machine learning field. Some of the most interesting and potentially profitable applications of artificial intelligence are interpreting what already exists.

The first wave of artificial intelligence systems were meant to be neutral and fast, so bias was seen as bad. We need subjectivity because the systems can mimic what humans do. Over the course of a single generation, we need to update our expectations of artificial intelligence.

Rooting out bad bias

The issue of accountability becomes important as the use of artificial intelligence increases.

It is easy to point the finger at the data or the program. It is easy to see how dependent projects are on easily accessible upstream libraries, protocols and datasets.

Problematic data sources are not the only vulnerabilities. The way we test and measure models can be affected by undesirable biases. Humans are the ones who build the models. How we go on to use the results is one of the decisions we make. Skewed results can be mitigated by having diverse teams and a collaborative work culture in which team members are free to share their ideas and input.

Accountability in AI

Building more diverse teams will help build bias. Change has been slow despite research showing that more diverse teams lead to increased performance. This is also true in the field of artificial intelligence.

We should keep pushing for culture change, but it's just one part of the bias debate. There are regulations governing the bias of the artificial intelligence.

Companies should be more careful with their artificial intelligence. The aim of the Algorithmic Fairness Act is to protect the interests of citizens from harm caused by unfair artificial intelligence. The EU's proposed regulation will heavily regulate the use of artificial intelligence in high-risk situations. New York City will require companies to perform artificial intelligence audits to evaluate race and gender biases.

Building AI systems we can trust

To build a new model, for example, organizations need to think carefully about the data they are feeding into the system, as well as how the system is being used. At the point where employees are trained in using it, as well as test and measurement, and results interpretation, they must go further to make sure that the consequences don't get in the way.

As the field of artificial intelligence gets more regulated, companies need to be more transparent in how they use their technology. They need a framework that understands and governs implicit and explicit biases.

They are not likely to achieve their goals without culture change. The conversation about bias needs to expand to keep up with the new generation of artificial intelligence systems. Governments, organizations and citizens alike will need to be able to measure all the biases to which our systems are subject as artificial intelligence machines are increasingly built to enhance what we are capable of.

Symbl.ai is a company founded by the CEO and co- founder.

The VentureBeat community welcomes you.

Data decision makers can share data related insights and innovation.

Join us at DataDecisionMakers to read about cutting-edge ideas and up-to-date information.

You could possibly contribute an article of your own.

Data decision makers have more to say.