THE MARKET FOCUSES UPON THE RISK OF BLACKBOX AI OVER LEARNING TO DEVIATE FROM THE NORM
‘One Rogue Leads Another’
Black-box Artificial Intelligence is where the algorithms are hidden. When Machine Learning is applied it is empowering the machine to evolve. There are ways to limit the evolution through human-controlled interventions. The risks are amplified when the AI is empowered to learn without sufficient controls. As there is poor transparency of how the algorithms evolve there is a risk of the AI over learning. This can mean there can be subtle and very gradual deterioration or deviation of the AI decision capability.
The realisation that there have been decision distortions may not be realised until the AI is travelling fast down the slippery slope. By that time, the damage to the AI capabilities maybe so severe that the AI is terminated. This may sound severe but an AI trying to unlearn to relearn is so problematic it is often better to start again.
The gathering pace of AI means a lot of people have jumped upon the bandwagon of AI without the expertise or know-how to best deploy a balance between black-box and white-box AI. For example, at Df2020 we deploy white-box AI for regulations as it is fully human controlled. Our argument is you cannot empower black-box AI to rewrite regulations as it is illegal.
The recent publicly-based squabble between Elon Musk and Mark Zuckerberg about artificial intelligence has been well documented. Musk at one stage described the Facebook CEO’s knowledge of AI as “limited”.
Elon Musk advocates that AI needs to be proactively regulated to mitigate against a ‘fundamental risk to human civilization’. Musk uses AI for transportation such as cars and space rockets. In these cases, the Musk approach is to ensure there are sufficient AI safeguards in place to protect people.
Zuckerberg, on the other hand says Musk is “pretty irresponsible” for using language such as “I keep sounding the alarm bell, but until people see robots going down the street killing people, they don’t know how to react, because it seems so ethereal.”
As the intensity of the squabble increases between Zuckerberg and Musk, suddenly Facebook hits a problem.
Facebook had to “shut down” one of its AI services after it went rogue and invented its own language. The observation was made after two of its Chatbots, Bob and Alice were negotiating and used their own language to complete the exchange, according to a report in Digital Journal.
It seems these two Chatbots were empowered to rapidly evolve through Machine Learning without human controls and interventions in place. Though both Chatbots were empowered to learn English, they evolved their own language. This meant the humans were out of the loop in understanding the conversation.
The notion of “over learning” has been an AI issue for some decades. For example, researchers have encountered numerous documented situations elsewhere when an AI has diverged from its training in English language and developed its own language.
Musk is not alone as among others, Stephen Hawking and Bill Gates are saying caution is needed.
In the meantime, whilst this controversy continues, Facebook has bought an AI Chatbot early stage firm called Ozlo to continue building conversational capability into Facebook Messenger as the platform accelerates building services.
The Ozlo Chatbot is aimed at handling more complex Q&A conversations such as whether a restaurant is ‘group friendly’ based on data from multiple reviews.
Ozlo had raised US$14m in one round from a syndicate led by Greylock Partners supported by AME Cloud Ventures and Jerry Yang.
This acquisition has strengthened Facebook’s capabilities as the competitive intensity of conversation-as-a-service continues amongst the other tech titans: Microsoft, Amazon, Google, Apple and TenCent.
Chatbots are going to lead the revolution of conversation-as-a-service will change everything inside and outside organisations. The key is for governments, industries and practices to deploy AI sensibly with sufficient checks and balances. A blended approach of Chatbots and humans as the next generation workforce will rapidly happen. As always, we need the checks and balances in place to handle rogues, whether they are human or artificial beings.