As we delve deeper into the realm of AI, a growing complexity of challenges and risks are becoming more lucid. In a recent Guardian’s report, AI could even eavesdrop typing sound and identify which keys people pressed for passwords! How can we ensure responsible use of AI when it is becoming more powerful and ubiquitous? In this article, we are trying to reimagine AI governance, and propose AI can play a more prominent role and become our co-pilot to pave the way for trustworthy AI.
The pressing need for ethical and trustworthy AI
There are manifold risks associated with AI or generative AI to be specific. To list a few most commonly addressed, generative AI can produce inaccurate or even fabricated responses. It’s also struggling with algorithmic bias that reflects and reinforces the existing bias in society. For example, we may all heard about some AI recruitment tools that can disadvantage female candidates, or certain facial detection that works more accurately on Caucasian faces. In addition, AI-generated works are faced with the challenge of violating copyrights and data privacy. And more seriously, generative AI’s widespread availability substantially escalates risks of it being used maliciously for cyber attacks and fraud.
To mitigate and prevent risks, governments and AI industry leaders around the world are rolling out AI governance initiatives and policies. In June, EU made the first move and carried out AI Act, which is deemed as the world’s first comprehensive AI law. In July, seven leading AI companies, including Amazon, OpenAI and Meta, followed suits and introduced voluntary safeguards at the White House. And in the upcoming fall, we are expecting the new international standard ISO/IEC 42001 on AI Management Systems to be published.
Here in Singapore, where WIZ.AI is headquartered, the government takes a balanced approach to facilitate innovation and safeguard consumer interests. The government regularly facilitates dialogues and collaborations among the public sector, industry and academia. In a recent discussion where WIZ.AI participated, we also exemplified our commitment to responsible use of AI. Our company is ISO 27001 certified, which is a globally recognised standard for information security management systems (ISMS). This certification demonstrates our commitment to the highest levels of data security and privacy. In addition, we also certified by SOC2 Type II Report, an independent audit report that provides assurance on the controls of a service organization. It evaluates the design and operating effectiveness of the organization’s controls related to security, availability, processing integrity, confidentiality, and privacy.
Responsible use of AI: current practices
Besides broader initiatives mentioned, currently there are diverse AI governance practices from companies developing and deploying AI as well.
- Incorporate guardrails:
- Develop compliance processes for risk assessment and review
- Retain proprietary data within secure or localized infrastructures, under direct monitoring of the organization
- Grant system users with different roles and data access
- Leverage privacy enhancing technologies (PETs) to derive insights from consumer datasets, safeguarding personal data confidentiality
- Intelligence collection about ongoing breaches and threats
- Eliminate toxic data that contains hate, abuse, profanity content, as well as those with private information and license constraints, during the data filtering stage
- Testing systems for AIGC tools’ safety, explainability as well as their robustness against security attacks
- Uplift transparency:
- Document the data pile used for training, as well as how the model is trained and validated, increasing explainability in AI-decision making
- Label content that’s generated by AI, or clearly inform users that they are interacting with AI systems
- Human validation:
- Loop humans to validate AI outputs before public distribution
- Create feedback loop for users to identify issues with AI system’s functionality and behaviors
- Ongoing monitoring by humans to detect unexpected behaviors, flaws or vulnerabilities after model deployment
AI as co-pilot in AI governance
After examining current initiatives and practices for responsible use of AI, we are wondering if we are being a bit too human-centric sometimes, and overlook the role AI can play in its own governance?
For starters, Generative AI is already assisting with risk mitigation. Companies have been utilizing generative AI to “red team” AI solutions and generating edge cases to test system’s robustness. There are real-life examples with generative AI goes beyond laboratory testings too. China’s Ant Group introduced an LLMs-enabled AI assistant for anti-fraud education, resulted in a 10% reduction in reported fraud cases. Waabi, a Toronto-based start-up uses generative AI to automate vehicle safety. This helps the company operate at only 5% of the previous cost. WIZ.AI leverages LLMs to detect sensitive content for better compliance and help maintain sensitive content errors at <5% level.
Generative AI already takes up a role as an assistant mitigating risks at a lower cost, what if it takes a more proactive role and becomes a co-pilot to achieve responsible AI more efficiently?
At WIZ.AI, we decided to test the above hypothesis. We trained one LLM to monitor the other LLM for bias, violence, and profanity content in workplace chats. The monitoring LLM successfully flagged many issues based on the given prompts, without further human involvement. However, we still haven’t fully realized the co-pilot status because there are still loopholes that need human to check the root causes and tweak the prompts accordingly, and there are even cases where human cannot yet identify the root cause that leading to AI’s obvious false judgements.
With the further exploration and development of our LLM technology though, we can envision a future where human-AI co-pilot in AI governance can be achieved, when humans can understand generative AI’s decision making process on a granular level. By then, the well-trained AI models will “tutor” other models on workplace’s “responsible code of conduct” independently with zero mistake. Humans will only need to validate the learning results without too much involvement in tedious work, such as fixing errors, digging the roots causes, and shift attention to more important work.
Forward-looking of a universal responsible code of conduct
Take an even more forward-looking perspective, a universal “responsible code of conduct” module can be developed and by-default plugged into every newly launched LLM or their much more intelligent “offsprings”. There can even be specific versions by region, by country or by individual organizations building on top of the foundational module. This process ensures AI systems are “aware” of their responsibilities and ethical guidelines upfront and ultimately liberates human labor. Naturally, humans will still be in the loop, judging AIs’ “performance”after their onboarding, reviewing and iterating the “responsible code of conduct” module on an ongoing basis.
Looking ahead, we see a world where human-AI collaboration permeates every possible aspect, ushering in an era of ideal co-existence between humanity and technology.
Reference
2023 World Artificial Intelligence Conference, Trustworthy AI Forum. Retrieved from: https://36kr.com/live/2332158061807241
IBM Research. (2023). Generative AI for business. Retrieved from: https://www.youtube.com/watch?v=FrDnPTPgEmk
Speech by Minister of Communications and Information, Mrs Josephine Teo, at the Opening of the Personal Data Protection Week on 18 July 2023. Retrieved from: https://www.mci.gov.sg/pressroom/news-and-stories/pressroom/2023/7/speech-by-minister-josephine-teo-at-the-opening-of-the-personal-data-protection-week-on-18-july-2023
Brand Smith, Microsoft. (2023). Advancing AI governance in Europe and internationally. Retrieved from: https://blogs.microsoft.com/eupolicy/2023/06/29/advancing-ai-governance-europe-brad-smith/
Guardian. (2023). AI can identify passwords by sound of keys being pressed, study suggests. Retrieved from: https://www.theguardian.com/technology/2023/aug/08/ai-could-identify-passwords-by-sound-of-keys-being-pressed-study-suggests?CMP=fb_a-technology_b-gdntech
Saritha Rai. Bloomberg. (2023). Startup Sees AI as a Safer Way to Train AI in Cars. Retrieved from: https://www.bloomberg.com/news/newsletters/2023-08-08/raquel-urtasun-startup-waabi-trains-autonomous-vehicles-with-ai
