Led by OpenAI poster child, ChatGPT, in a matter of months, hundreds of millions of individuals across the globe have been activated to the life-changing possibilities of machine learning. As hundreds of new AI tools are released each week, work productivity is on the rise[1], and global consultancy behemoth, PwC, has predicted that AI will contribute US$15.7 trillion to the global economy by 2030.[2] Yet with the power to create exponential, socio-anthropological transformation, comes great responsibility. The first part of our series on emerging AI platforms like ChatGPT highlighted some serious legal challenges raised by these disruptive technologies. In this latest instalment, we look at some more issues that encluster this increasingly out-of-control space.
Below we look at what to do about biased outputs, the transparency paradox, and mounting challenges for regulators.
AI programs such as ChatGPT are trained on billions of pieces of internet data, including web pages, articles, and social media posts. Many of us have quickly become accustomed to its vast intelligence, and seemingly nonchalant ability to provide detailed, insightful and often creative responses, based on a range of natural language processing tasks, at processing speeds one thousand times faster than most humans.[3] However, it is a simple truism that the models are only as smart – and principled – as the data they are trained on. If the underlying data it relies on is biased, and the AI model is unable to identify that bias, then the tool can serve to perpetuate the bias, which may lead to socially undesirable forms of discrimination.
Large language models are trained to perform the task of predicting the next word in a sequence of words, by optimising for human preferences. OpenAI collects a large, high-quality dataset of human-to-summary comparisons, trains the model to predict the human-preferred summary, and uses that model as a reward function to fine-tune the summarisation policy using reinforcement learning. These techniques are designed to avoid outputs that are false, toxic, biased or just not useful to the user.
‘Bias’ is a phenomenon that skews the result of an algorithm in favour or against an idea. Bias is considered a systematic error that occurs in the machine learning model itself due to incorrect assumptions. Technically, it is the error between average model prediction and the ground truth.[4]
Discrimination occurs under principles of Australian and general international law when a person, or a group of people, is treated less favourably than another person or group because of their background or certain personal characteristics, such as race, colour, sex, language, religion, political or other opinion, national or social origin. When machine learning algorithms are trained on biased data, they can perpetuate discrimination.
Instances of bias have been widely reported in ChatGPT outputs, with racist and sexist content produced by the AI going viral on social media. Some users managed to jailbreak the chatbot to create ChatGPT’s politically incorrect alter ego, DAN.[5] Though developers have continued to upgrade ChatGPT’s safeguards against these materials, ‘black hat’ type users continue to search for the next virtual chink in the AI’s digital armour.
In theory, increasing the complexity of the model to count for bias and variance, will decrease the overall bias while increasing the variance to an acceptable level. Increasing the training data set can also help to balance this trade-off, to some extent. A large data set offers more data points for the algorithm to generalize data easily. However, underfitting or low bias models are not that sensitive to the training data set.
Managing this issue is challenging for the law given the open nature of the web, and the plethora of the biased training material on the internet. This requires some creative solutions, such as:
Removing bias from information is a large and difficult problem that has long challenged philosophers, thinkers and lawmakers. Even identifying implicit and unconscious biases can be difficult. In fairness, programs like ChatGPT have been making valiant attempts to remove bias and discrimination from their outputs. However, completely laying this long-standing human challenge at the feet of developers and owners of these powerful and increasingly influential programs may be asking too much. Industry self-regulation may need to be supported by greater legal and regulatory oversight.
As we spoke about in the first part of our series, it is unclear who to hold legally accountable and liable for illegal, false, toxic or offensive outputs generated by AI programs.
The secret sauce of platforms like ChatGPT and Dall-e 2 has so far been well protected, with many wondering how exactly they ‘do what they do.’ However, an absence of transparency can decrease consumer trust and confidence, and make AI susceptible to allegations of unfairness and discrimination. It is a common complaint by users that ChatGPT does not breadcrumb its conclusions.
AI lack of transparency presents a paradox: on one hand, it is much harder for hackers to infiltrate the software of an opaque AI, meaning that the systems are more secure for regular users if the ‘impenetrable black box’ approach is maintained. On the other hand, it is difficult for consumers to trust a technology that does not explain how it processes data or maintains cybersecurity.
Ultimately, this raises a question about how much information developers and owners, such as OpenAI, should be required to disclose to users about their tech. As people the world over quickly come to depend on these intelligent platforms, there are legal and moral questions at play here. This challenge is exacerbated by the lack of consistency in AI operations across international borders, with different countries having differing approaches to AI regulation.
Unsurprisingly, the rapid growth and deployment of AI models has outpaced the development of regulatory frameworks to ensure their safe and ethical use. At present, there is a void of regulation governing AI, which is further complicated by profound challenges in determining how to regulate this rapidly evolving technology. This comes down to issues like whether a scientific, technical or human-based legal vocabulary should be used and establishing definitions to account for such complicated systems. Scientific language may be harder for regulators and users to understand, but on the other hand legal language may not be suitable to effectively regulate AI.
The lack of regulation has led to concerns about the potential misuse of AI models. Balancing the need for innovation with the constant striving for progress has always been a difficult task for the law. While there is a risk that hastily overregulating AI will stifle innovation and investment in developing technologies, the law needs to ensure that models are safe and ethical.
Further, since AI operates over the internet, the regulation of AI is both a global and a local challenge. So far there has been a notable ack of consistency among nation states as to how to regulate this tech. The European Union is currently progressing a draft version of the AI Act through their legislative process, which, if passed, would be the first significant piece of legislation regulating AI by a major power.[8] Key elements that are sought to be regulated include:
In a similar way to what the EU’s GDPR has done for privacy regulation, the AI Act could potentially serve as the template and high watermark for other jurisdictions seeking to regulate machine-learning models. For Australia’s part, the Department of Industry has published AI Ethics Principles, but so far there are no specific laws governing AI. Other countries, including Canada, are proposing their own statutes.
Given the global nature of AI, a globally co-ordinated approach would be ideal, though at this stage appears more idealistic, and potentially even unrealistic.
Delicately balancing the briar patch of competing legal, commercial and ethical considerations surrounding the rise of AI arguably presents a generational challenge for lawmakers. Can these global platforms be entirely bleached of bias and discriminatory tendencies, or is that an unrealistic ideal in the complex, plural world in which we live. Do these platforms need to be more transparent to engender consumer trust, or do they need at least some level of opacity to maintain security. Is a globally co-ordinated approach to regulation of AI possible, or does the typical desire for States to manage laws within their own boundaries make this unrealistic.
These are difficult and thorny questions, that for some time at least will continue to both entangle and enchant people across the world.
E+Co are experts in digital industries and the protection and commercialisation of IP assets in an increasingly online, onchain and web 3.0 world. For advice on developing laws surrounding AI and the digital landscape generally, please contact us below.
[1] Researchers from Massachusetts Institute of Technology and Stanford University analysed the impact generative AI tools like ChatGPT had on productivity at an unnamed Fortune 500 software firm. Using data from 5,179 customer support agents, the research team found that workers who had access to an AI-based conversational assistant were almost 14% more productive than those who did not. At: https://fortune.com/2023/04/24/chatgpt-ai-worker-productivity-fortune-500-study/amp/
[2] ‘Sizing the prize: PwC’s Global Artificial Intelligence Study: Exploiting the AI Revolution’, PwC (Article, March 2023) <https://www.pwc.com/gx/en/issues/data-and-analytics/publications/artificial-intelligence-study.html>.
[3] Chris Havergal, ‘Robot scientist ‘works 1,000 times faster’ than human researchers’ Times Higher Education (Article, 8 July 2020) <https://www.timeshighereducation.com/news/robot-scientist-works-1000-times-faster-human-researchers>.
[4] See: https://www.bmc.com/blogs/bias-variance-machine-learning/
[5] Josh Taylor, ‘ChatGPT’s alter ego, Dan: users jailbreak AI program to get around ethical safeguards’ (Article, 8 March 2023) <https://www.theguardian.com/technology/2023/mar/08/chatgpt-alter-ego-dan-users-jailbreak-ai-program-to-get-around-ethical-safeguards>.
[6] Poornima Apte, ‘5 Ways to Prevent AI Bias’ ITPro Today (Article, 23 September 2022) <https://www.itprotoday.com/artificial-intelligence/5-ways-prevent-ai-bias>.
[7] ‘Artificial Intelligence Act: Council calls for promoting safe AI that respects fundamental rights’ Council of the EU (Press Release, 6 December 2022) <https://www.consilium.europa.eu/en/press/press-releases/2022/12/06/artificial-intelligence-act-council-calls-for-promoting-safe-ai-that-respects-fundamental-rights/>.
[8] ‘Developments’ The AI Act (Web Page, 22 March 2023) <https://artificialintelligenceact.eu/developments/>.
