AI is used in employment decisions ranging from investigation outcomes, performance assessments, pay-reviews, candidate assessments, and disciplinary outcomes.
As employment and privacy lawyers, we love to tell clients to keep a ‘human in the loop’ to review AI outputs before making decisions.
All well and good, but what does it actually mean? You’d be forgiven for seeing it as an opaque cop-out.
The daunting challenges facing a human reviewer who’s trying to understand and explain AI outputs:
- AI generated explanations are guess-work: if you ask an AI tool to explain its decision, the explanation is still predicted text – the model’s best guess at what good reasoning for that task looks like. It is not a window into the complex computation that actually produced the AI’s output.
- AI does not repeat itself: run the same input twice and you get different output, due to built-in randomness. This makes consistency across employees – and after-the-fact auditing – genuinely hard.
- Confident fabrication: AI states things fluently and authoritatively even when wrong – for example, missed four deadlines when there were none. It can confidently express qualitative inferences that are not grounded in fact. It can subtly misquote or conflate two entries in data. Put this together and a reviewer skimming for errors is poorly placed to catch a confident, specific falsehood.
Practical solutions
So, what are the practical answers? How does an employer comply with the good faith duty in section 4 of the Employment Relations Act 2000 and IPP8 (Accuracy) in the Privacy Act 2020?
The answers below do not solve the problem entirely. But there are practical tips that go a long way to implementing more reliable, safer use of AI and compliance with the employer duty of good faith.
Four key tips:
- Prompts: perhaps most importantly – build a rule-based prompt that limits review to specific documents, demands citations, nudges the tool to admit and identify uncertainty, separate fact from inference, argue both sides, require ‘reasoning first, conclusion last’, and acknowledge missing information. We have developed and refined a rule-based prompt instruction that you can implement. It is geared specifically to the duty of good faith. Get in touch with us for details..
- Records: capture the prompt, the output, the model and version, the date, and who reviewed it. That information must be carefully secured – but you will need it to justify decisions and meet information requests under the Privacy Act 2020.
- Review: a human must cross check the citations and sources for accuracy before making preliminary decisions.
- Feedback: be transparent. Provide the prompt and output to the employee. Ask for their feedback. The employee may well have additional context to share with you – for example, illness, sick leave, annual leave, or a mix-up in the output. This is critical to the duty of good faith and ultimately beneficial to help employers make better-informed decisions.
Our employment, privacy and AI law specialists are well-placed to give further practical tips and provide you with our in-house-built prompt models, AI Use Policies, Privacy Impact Assessments and Data Breach Response Plans.
Our view remains that the current technological minefield presents a unique opportunity for business success. A company with a legally and technologically robust AI strategy is streets ahead of the competition and can move forward confidently and responsibly with deployment and use of AI tools.
