AI: perils and promise

Vol 46: Issue 3 | September 2023

IN SHORT

• Climate change will affect populations in different ways, depending on their geography, demographics and health status.
• Insurers need to identify which impacts will have the biggest effects on their business.
• New approaches such as microinsurance, parametric risk transfer solutions and digital technology could help insurers to close the protection gap.

Less than a year since its launch, ChatGPT seems to be an omni-presence in our lives as it grabs headlines and redefines some tasks and workflows in offices, universities and homes.

In case you missed it, ChatGPT is a chatbot that was developed by OpenAI and released publicly on 30 November 2022. ‘GPT’ stands for generative pre-trained transformer.

Essentially, you ask the chatbot a question, and it uses artificial intelligence (AI) to generate a unique response in text, images or video — drawing on language patterns it accesses from the pre-2021 internet.

Human feedback — voting responses up or down — improves ChatGPT’s performance over time, so that future responses will be more accurate and relevant. 

Users can ask ChatGPT for factual information, such as ‘What is the average annual rainfall in Auckland?’, right through to asking it to write a 5,000-word university essay on platypus habitats, requesting a graph of a company’s 10-year growth or getting it to write a short, professional email to a client, seeking additional information to complete a project. 

The focus for industry leaders and legal and compliance teams, however, now needs to shift from incredulity about the power of OpenAI’s famous chatbot to some critical questions. 

What business risks and opportunities do ChatGPT and other generative AI tools — ones that create new content in response to a prompt — present? How should insurers respond strategically and legally to this technology? How can they protect the data and intellectual property of their customers?

Jono Soo, head of cyber specialty at Marsh New Zealand, says insurance industry leaders need to learn the nuances of the technology and be mindful of any associated risks. 
“The key is to understand how [ChatGPT] is intersecting with current areas of coverage,” he says. 

"In some respects, it’s just another advancement of existing AI technology, so there doesn’t need to be a knee-jerk reaction from markets at this stage.”

In its Global Risks Report 2023, Marsh states that advancements in AI, quantum computing and biotechnology will create risks and “enable the misuse of personal information through legitimate legal mechanisms, weakening individual digital sovereignty and the right to privacy, even in well-regulated, democratic regimes”.

 Soo says any concerns about generative AI technology are clearly not just an issue for the insurance industry. “Governments need to be across this as well, and societies in general.

We really need to find some frameworks to quickly regulate the development and use of this technology and start putting some guard rails around it.”

Risks on the radar

Clyde & Co digital law and data privacy lawyer Alec Christie says there is an understandable desire among many insurance leaders to test the capabilities of ChatGPT as they seek to automate mundane processes, respond to internal and external queries, and generate quick answers to myriad questions.

“On the simple tasks, where there’s generally available information, ChatGPT can put together a reasonably good answer,” he says.

“The problem comes if there’s incorrect information or a pervasive urban myth, because it picks that up and ingests it.”

Christie’s concern for insurers and others is the risk of an “almost incidental and accidental ability to infringe intellectual property and not comply with privacy provisions”.

“If you’re not addressing the potential risks from an IP and privacy point of view, you’re opening yourself up to a huge risk,” he warns.

Like Christie, other lawyers and leaders are trying to comprehend some of the risks — legal and reputational — that exist for insurers and other businesses from generative AI tools. 

Defamation lawsuits and other liability concerns

ChatGPT is smart, but it is not foolproof. Indeed, verifying information is one of its weaknesses because its responses are based on patterns, rather than up-to-date information or knowledge. 

This opens the way for possible defamation action, as shown in the case of an Australian regional mayor, Hepburn Shire’s Brian Hood, who is threatening to sue OpenAI if it does not correct ChatGPT’s false claims that he had served time in jail for bribery.

The case could be the first defamation lawsuit against ChatGPT and highlights a common problem: the possibility of chatbots providing incorrect, although superficially plausible, information.

To protect themselves, businesses should issue guidance that requires employees to review any output generated by chatbots for accuracy.

Sebastian Hartley, a general litigator at Holland Beckett in New Zealand, says the temptation to “delegate work” to chatbots creates a risk if their output is not properly checked.

“You really need that verification, because if information is being acted on, or it’s being packaged and then sent out to others, liability attaches to that information.”

Significantly, there is no specific regulatory framework in New Zealand or Australia to oversee the use and abuse of AI tools.

“People are yet to really wake up to the legal implications,” says Hartley. “We find that with most new technologies, there’s a lag before we really begin to see businesses getting lawyers involved. Typically, something has to go wrong first.”

Data privacy and confidentiality fears

A massive dataset of text — such as books, articles and websites — are used to pre-train ChatGPT and similar ‘large language models’. And this raises questions about the privacy and security of the data that is used to train these AI tools.

For example, when the chat history of ChatGPT is not disabled, any information that is entered into it could become part of its training dataset.

If an employee at an insurer inserts sensitive and personal customer data into an AI search, that data can then be used as part of responses to other chat requests.

“As an insurer, if you’ve got all this great data on a lot of people who may have been with you for many, many years, it’s something you’ve got to seriously protect,”

says Christie, who advises businesses to set up compliance frameworks for the handling of such data.

Phishing emails and cyber attacks

Hackers are already using ChatGPT to write advanced phishing emails and create malware. This is making it more difficult than ever to detect the difference between AI and human-crafted emails and texts. 

The technology has also lowered the entry bar for sophisticated attacks from a larger cohort of scammers, who are using ChatGPT for in-depth learning about their targets. As a result, businesses can expect a new wave of attacks, and insurers are likely to face a rising number of cyber claims.

Insurers, be on your guard

What does this all mean for insurers? Hartley believes generative AI can be a great tool for insurers to crunch data and analyse risk.

However, it is a case of user beware. “It’s garbage in, garbage out, as computer scientists have said for a long time,” says Hartley.

“These tools are only as smart as what you’re feeding them. I suspect that within about five years, we’ll be in a radically different position as AI models learn not only to make better sense checks of what they’re producing, but also get better at selecting and even recommending the input.”

Soo agrees that the key for insurers will be to establish policy guidelines around the use of generative AI and to ask existing and potential customers about their use of the technology.

“We find, with our background in the cyber space, insurers have been really good at asking the right questions through the underwriting process, including what risk-management controls are in place,” he says. “That will be important with AI, too.” 

Christie says the red flags with ChatGPT for insurers are clear: “Don’t just jump into the lake without checking how deep it is first.” That means paying attention to risk management, doing your due diligence and putting in place smart usage policies for generative AI tools.

“There’s a lot of potential upside for businesses and insurers,” adds Christie, “but you’ve got to manage it, because sticking your head in the sand is not going to cut it when these issues start to get the attention of regulators.” 

An ethical minefield

These are three of the biggest ethical challenges stemming from generative AI.

1. Biased, sexist or racist responses 

Programming bias is one of the major flaws of ChatGPT and its rivals. In many instances, AI may provide racist and sexist responses because of prejudiced data.

The case of another large language model, Galactica, is instructive. In November 2022, Facebook owner Meta unveiled Galactica — after training it on millions of examples of scientific articles, websites, textbooks and lecture notes — with a view to serving the science community.

Just three days later, Meta shut down the platform after a backlash, with users criticising it for spreading misinformation and spewing out racist diatribes.

These potential biases mean it will be crucial for insurers to carefully consider the quality and diversity of any data they use with AI platforms. 

“Unlawful discrimination is clearly one of the risks of using ChatGPT or AI for data analysis,” says Holland Beckett’s Sebastian Hartley. “If you’re not consciously working to check and correct those sorts of biases, it can have a systemic discrimination effect, which will obviously be pertinent for insurers.”

2. Plagiarism and cheating

People experimenting with ChatGPT have been quick to realise that they can use it to cheat. 

There have been reports of people taking advantage of the tool to win coding contests, while many school and university students are using it to write their essays and assignments.

However, plagiarism-detection software is quickly improving. Some universities, too, are allowing the use of AI in assignments, providing it is disclosed.

3. Intellectual property breaches

Expect lawyers and ethicists to be busy as they grapple with uncertainties around the ownership of AI-generated works and content. 

Clyde & Co’s Alec Christie says the threat of copyright or intellectual property (IP) breaches with generative AI must be taken seriously. 

Clyde & Co is helping to draft generative AI usage policies for clients. Such policies set “the rules of the road”, covering factors such as why a company needs to use the technology, who can use it and for what purposes it can be used. 

“And, most importantly, how can we double-check that any usage does not involve an IP infringement?” says Christie.

He believes pressure will mount on insurers to ensure that their customers have such a usage policy in place.

Read this article and all the other articles from the latest issue of the Journal e-magazine

Read issue 3 of the Journal