Enhancing AI Accuracy

We all know Generative AI has emerged as a dynamic force in HR. Finally developing that tailored learning content you never got to in hours rather than weeks, helping line managers 'tone down' their employee communication to an acceptable level fit for publishing at tribunal (we've all been there!). HR departments are increasingly using generative AI to refine, develop and enhance communications to their audiences, and supercharge their efficiency. While generative AI offers promising capabilities, it also introduces significant risks.

I think I’m losing it!

A primary concern is that generative AI can produce 'hallucinations' or inaccurate outputs that appear perfectly plausible to the untrained (or uneducated) eye. This issue arises because AI models generate responses based on existing data, and if it doesn't know, it sometimes fills the gaps with incorrect information. If users rely on these outputs without verification, they risk propagating inaccuracies that could impact decisions and undermine employee trust. Granted, the powers that be are aware of its bejewelled imagination, and they are making great strides to reduce these, but responses still aren't fool proof.

I’m an expert . . .

This has also created an increased risk of 'uneducated experts'. Those that think they know the right answer but have very little knowledge of the subject area. Everyone can prompt GenAI and copy and paste an answer, but issues arise when you don’t ask the right questions. Think about the conversations you have with your line managers and employees when they’re asking for advice. Very rarely will you give it without asking questions about the context of the scenario, the length of service, the absence history or make reference to a policy or procedure. Critical information is missed or not passed on in prompt, which inevitably places limitations on the responses. And then there is the fact checking. That evidence base that HR relies on. Do you think if they’re circumventing HR to ask for advice from GenAI they’ll take the time to check its response? Imagine the carnage that would ensue if your line managers were asking ChatGPT for HR advice without knowing your policies, regional employment law or even understand what custom and practice means.

Rewrite the past

As mentioned above, these model generate responses based on the past. What has gone before, whether that’s good, bad or indifferent, AI doesn’t know the difference. It’s for this reason that Bias is another critical risk. Because these systems learn from historical data, any biases embedded within that historical data can influence AI outputs, perpetuating discrimination and inequity in its responses. If you keep doing what you’ve always done, you’ll keep getting the same result. Don’t expect GenAI to change that. It’s not about reframing the question, but asking the right questions.

The real change comes in the ability to spot the risks and flaws, and to challenge this. To the experienced and trained HR eye, bias can be easy to spot, and this is a skill that needs to be developed and shared with others using GenAI. Teaching AI a new pathway also gives it a new history to rely on that is more credible than its previous past.

Context, Prompt, Challenge – rinse and repeat

In generative AI, the quality and specificity of prompts directly affect the relevance and accuracy of outputs. Prompt engineering is a skill that is quickly becoming essential in HR as AI technology advances. HR professionals need to become proficient at crafting prompts that clearly convey their objectives, ensuring the AI generates useful responses.

As HR professionals adopt prompt engineering skills, they will be better positioned to navigate AI’s complexities. HR teams should invest time in learning how to use generative AI tools, particularly in crafting high-quality, context-rich prompts that can unlock AI’s full potential while minimising risks.

Use diverse and high-quality datasets, implement continuous human oversight, and ensure the presence of ethical guardrails around AI-generated content. It’s not a one size fits all so it’s crucial for HR professionals to understand the limitations of generative AI and be prepared to intervene if the AI’s advice strays from organisational policies or values.

Ultimately, human judgement remains crucial in identifying when AI is producing misleading content, and continuous improvement is necessary to ensure AI models align with organisational standards.

For more on prompt engineering, check out our recent podcast with Kieran Gilmurray: AI: HR's Ally or Enemy