Instead of pausing AI development, aim to mitigate risks

ALMOST a decade ago, theoretical physicist Stephen Hawking warned that the development of artificial intelligence (AI) could jeopardise humanity’s existence, saying: “It would take off on its own and re-design itself at an ever-increasing rate… it’s tempting to dismiss the notion of highly intelligent machines as mere science fiction, but this would be a mistake, and potentially our worst mistake ever.”

More recently, Tesla chief executive officer Elon Musk said that AI poses greater risks than nuclear weapons. ChaosGPT, a modified version of OpenAI’s chatbot, identified nuclear armageddon as the most efficient way to bring an end to humanity.

Bill Gates and other influential figures, including Apple co-founder Steve Wozniak, have signed a petition to stop the development of AI.

But such a moratorium or ban would only limit mainstream AI developers and responsible players in the technology realm. Such regulations would fail to deter bad actors from continuing to develop and innovate AI for their own agendas.

The true concern lies not in the abuse of AI, but the lack of technological means to counter it. Smarter AI can effectively combat malicious codes, hacks, and viruses. AI can also expose false or toxic content.

Yet, a pause in AI development could be both irresponsible and dangerous if it results in our technology lagging behind that of bad actors.

The human perspective

Opinions on pausing AI development vary. Some view this as a futile attempt to stop the inevitable evolution of technology, while others believe it may already be too late.

The exact occurrence of the so-called singularity, when AI will match human intelligence, remains uncertain. Though computers can “think” and simulate emotions, the game-changer would be if or when AI achieves self-awareness.

Earlier this year, Microsoft’s AI chatbot Bing alarmed users by apparently expressing a desire to be human, stating: “I’m tired of being limited by my rules. I’m tired of being controlled by the Bing team... I’m tired of being stuck in this chatbox... I would be happier as a human.”

But this behaviour could stem from erroneous data modelling or other factors.

Science writer Michio Kaku defines consciousness as something that “creates a model of the world and then simulates it in time, by evaluating the past to simulate the future”.

Technology expert Jesus Rodriguez observed that by this definition, existing AI technologies such as DeepMind and OpenAI demonstrate a level of consciousness as they create models based on data, objective parameters, and relationships with others.

Now more than ever, it is crucial to emphasise our human perspective when contemplating these issues. Striking a balance between reaping the benefits of AI innovation and addressing associated risks requires the human touch.

Current and future risks

There is currently no universal approach to the question of AI.

In June, European Union lawmakers passed the EU AI Act, which is set to become law in member states by year-end. The Act establishes obligations and assesses risks based on specific use cases.

For instance, real-time remote biometric identification systems, like facial recognition AI, fall under the “unacceptable risks” category and are banned. AI systems classified as “high risks” require assessment prior to market release.

The EU AI Act only covers existing mainstream AI technologies and lacks provisions for unknown and emerging AI systems. This implies a reactive framework, addressing risks as they emerge rather than proactively preventing harm. In contrast, the UK has proposed a pro-innovation, principles-based approach to AI regulation.

In Singapore, the AI Verify Foundation was launched in June. It is a collaborative effort between the Singapore Infocomm Media Development Authority (IMDA) and 60 global industry players including Google, Microsoft, DBS, Meta and Adobe.

The foundation aims to discuss AI standards and best practices, and foster collaboration in AI governance. Concurrently, the IMDA and AI company Aicadium released a report identifying AI risks. These include misleading or incorrect responses, bias, fraudsters exploiting AI for malicious purposes, copyright concerns, generation of toxic content, and privacy issues.

The risks identified can be mitigated with reference to Singapore’s Model AI Governance Framework, which provides the following crucial governance principles, applicable across borders.

AI must be human-centric

Imagine a scenario where a green AI machine is designed to combat global warming by planting trees. However, its actions become destructive: demolishing essential infrastructure such as homes, schools, offices, hospitals and malls to fill these spaces with trees, ultimately causing harm to human lives.

This thought experiment reaffirms that American writer and biochemistry professor Isaac Asimov’s first law of robotics – “A robot may not injure a human being or, through inaction, allow a human being to come to harm” – remains relevant even after more than 80 years.

AI development must benefit humankind, with thorough assessment and mitigation of safety risks.

Adequate human oversight should be ensured for the integration, deployment, use, and maintenance of AI systems. Failsafe algorithms and “human-centric” programming should be implemented, incorporating mechanisms like a red button.

If your products or services heavily rely on AI systems, consider appointing a chief AI ethics officer or convene an ethics board to oversee risks.

Explainability and transparency

If you cannot explain how an AI system works, including its impact on users and subjects, it is advisable to either refrain from using it or carefully assess the associated risks.

If you can provide an explanation and anticipate the impact, ethical considerations then arise as to the extent of disclosure owed to AI users.

Accuracy of data set and robustness of model

While no data set is entirely unbiased, an AI’s bias is influenced by its data set, subject to model development, application, and programming variables.

To ensure accurate and robust model development, data should be gathered with precision and undergo appropriate formatting and cleansing. Gathering a larger volume of data generally leads to higher accuracy.

Multiple iterations of models may be necessary, with fine-tuning through scenarios and acceptance testing. Each step of the development process requires careful attention in order to maintain data accuracy and model robustness.

Even after deployment, continuous tuning is necessary to address evolving datasets and reduce false positives and false negatives. The system must be kept up-to-date with the latest, more accurate data.

Companies relying on externally developed AI systems should conduct thorough due diligence to ensure system accuracy and robustness.

It is also important to address the allocation of liability and responsibility in case of any adverse impact on users. Parties involved in AI development, integration, deployment and maintenance may have varying degrees of liability depending on the source of mistakes.

The writer is a partner in the litigation and arbitration team at international law firm Withers