This week, the Victorian government announced it will act to protect workers from “inappropriate and unfair uses” of artificial intelligence (AI) in the workplace. This will include introducing new laws to ensure AI hiring decisions are “free from bias or discrimination”.
The state government also plans to target “intrusive” workplace surveillance, which will include giving workers a right to know if they are being monitored, and requiring a human to have “final say” in any “significant automated decision made using surveillance data”.
Reactions to this announcement have been mixed. Business groups and some employment lawyers describe the proposal as “unnecessary overreach”, but it is seen as a “smart move” by unions.
As someone who researches how AI systems can discriminate, for me, this is a welcome focus. Eliminating all bias and discrimination from AI systems may not be achievable. But we can do more to protect job candidates from this new form of harm.
Bias and discrimination in AI hiring
There is strong evidence to support the state government’s assertion that AI tools can “systematically disadvantage women and people from diverse backgrounds”.
An analysis in the United States of one of the largest datasets to date – 3 million job applicants who submitted 4 million applications that were screened by algorithms built by the same software vendor – found clear racial discrimination against Black and Asian applicants.
If such a system is used by multiple employers, whole groups may be locked out of employment.
My own research has highlighted the ways AI hiring systems may exclude some job seekers with disabilities, older workers, women or those who speak or write English as a second language – all categories protected by discrimination legislation.
Discrimination may be embedded in the AI system via the data or the algorithmic model, or it might result from the way the system is used by an organisation.
Do we need new laws in Victoria?
The Victorian Equal Opportunity Act was drafted before the development of AI technologies. While this law clearly applies to discrimination during the hiring process, there are gaps.
For example, this law does not currently place any obligation on employers to make reasonable adjustments for people with disability in hiring processes.
Adding third-party AI hiring systems into this mix makes things even more complicated. Is the AI developer or the employer responsible if the software makes a biased decision, and will a hiring manager notice?
If no one can see inside the “black box”, it is extremely difficult for an individual to gather evidence to make a complaint. Reversing the burden of proof is one way this could be addressed.
We also need a greater focus on preventing AI harms before they occur. The federal government’s abandonment of its proposal to enact mandatory guardrails for high-risk AI systems has left a vacuum the Victorian government now seeks to fill.
These guardrails would have required AI developers and companies using these systems to take steps prior to rolling it out to reduce AI harms through testing, transparency and accountability.
Victoria isn’t the first state to make this push. In a similar vein, New South Wales recently passed its own laws to regulate AI in the workplace.
As it is desirable that this regulation occur at a federal level to avoid a patchwork of state and territory laws, the federal Labor government is this week considering at its national conference a potential AI governance role for the Fair Work Commission.
Safeguards against bias
The Victorian Equal Opportunity Act already imposes a positive duty on employers to take “reasonable and proportionate measures” to eliminate all forms of discrimination, including AI-based discrimination.
The insertion of a “human in [the hiring] process to ensure that biases and unintended consequences don’t take place”, as planned by the Victorian government, is one such measure.
Another measure is to conduct regular independent audits checking for bias. What employers using AI are currently lacking is clear regulatory guidance on the steps this duty requires they take to prevent such discrimination.
Additionally, the use of AI to analyse emotions has been scientifically discredited. There is evidence to suggest emotion recognition technology may discriminate on the basis of race, gender and disability.
It should be prohibited in the recruitment process and in workplaces more broadly, as the Victorian government has indicated it plans to do. The European Union’s AI Act bans AI systems used to infer emotions of a person in the workplace except for “medical” or “safety” reasons.
Vitaly Gariev/Unsplash
Bias-free AI?
Seeking to make AI hiring systems “free from bias and discrimination” is a difficult task. As a product of humans and based on human data, this may never be achievable. Even at a technical level, there is no generally accepted or standardised methodology for identifying discrimination by an AI hiring system.
But we can do better. The Victorian Equal Opportunity Act should be reviewed to identify and close gaps in protection. We need to focus on measures to prevent discriminatory harm before it occurs.
If an employer is unsure whether an AI hiring system encodes discrimination, it should not be used. We must also empower and resource the Victorian equality regulator to enforce these laws.
With a recent survey showing that less than a third of Australians view current regulations and laws as sufficient to protect against AI harms, there is a clear mandate for doing so.
Natalie Sheard receives postdoctoral research funding from the University of Melbourne.
