Fixing Gen AI trust issues: three focus areas to keep on your radar

People might think Gen AI is a simple case of, “I type in a question, I get the answer.” But underneath the hood of this technology is a large and complex model that’s inherently opaque. A feature of the underlying Large Language Model is that – much like a black box, there are always uncertainties – both statistical and systemic.

The term “hallucination” describes those times when AI presents incorrect information as fact. For example, I’ve been working with my colleagues at Fujitsu Research Europe on genomics AI. We built a model that can analyze genetic material such as skin samples for any signs of diseases; we compare subject samples with known good samples and determine whether the differences require further attention. More importantly, we provide the underlying reasoning. Out of curiosity we gave this same task to a Gen AI system, and it came back with a plausible explanation, but the outcome was factually incorrect. I use this for illustrative purposes: in the short to medium term at least, organizations need to use their own professional reviews and experience to validate the results.

Gen AI’s capabilities enable us to question the status quo and swiftly make groundbreaking discoveries, but we still need human knowledge and input to work alongside it. Put them together, and we can quickly test hypotheses that we may not have considered without the technology. So, see it as a way to augment your decision-making capability.

“If AI is used appropriately, we will create a better world – I truly believe that. But people justifiably have their concerns about the technology, and are skeptical about trusting Gen AI systems.”

Bias in input training data or algorithms can dramatically skew the integrity of Gen AI outputs. But detecting bias in training data is complex. At Fujitsu, we have an entire research unit in Europe and Israel working on detecting bias in AI training data.

It's impossible to completely eliminate risks such as bias, but it’s important to actively manage hazards in business. And there are several ways to tackle bias.

First, pull together a diverse multidisciplinary team to help inform and train your Gen AI systems. This will ensure that multiple perspectives are guiding how the tech is implemented and governed. These teams can, for instance, contribute to refining the guardrails embedded into Gen AI systems that teach the tool to avoid ethical issues such as unfair bias in its outputs.

Also consider recruiting an AI ethics specialist and using your technology specialists to routinely “red team” your Gen AI systems. Red teaming exercises stress-test the accuracy and moral code of your tools and you can use the learnings to upgrade your system guardrails.

And look to the wider industry for support. Some providers are experimenting with creating bias-management tools and embedding ethical codes and moral constitutions for Gen AI to operate with.

To benefit from this transformative technology and use it responsibly, business leaders must focus on unraveling the range of risks and governance challenges associated with Gen AI. This means they need to:

For many organizations, rapid advances in Gen AI have the potential to transform almost every facet of life. To seize its opportunities, organizations will need to use Gen AI alongside their existing professional expertise – together with a healthy dose of skepticism.