Is it time for a policy on artificial intelligence?

Is now the right time for Bangladesh to start addressing artificial intelligence (AI)? Are there more important issues deserving of the attention of policy makers and regulators at this time? 

Governments can choose to adopt network technologies, especially those dependent on the state's decisions to issue or not issue spectrum in particular bands or to assist or not with rights of way. The government of India, on the advice of Satyen Pitroda, decided in the late 1980s to pass on mobile telephony to focus on rural fixed telephony.   

Can today's governments choose to pass on artificial intelligence (AI) or connected devices (also known as Internet of Things or IoT)?

Why policy?

Today, a Bangladeshi with hearing disability (or even those who are not so disabled) can use Google's Live Transcribe, recently introduced in over 70 languages including Bengali. This allows the person who cannot hear to read on her or his smartphone what the other person is saying in real time. Live Transcribe is based on AI.

The person using Live Transcribe would have to be literate, possess a smartphone, have some kind of data connectivity and possess awareness of, and ability to, download the free app. According to the AfterAccess nationally representative survey conducted in Bangladesh in 2017, 24 percent of the 15-65 population of the country possessed smartphones, a number that should have increased by now. Persons with disabilities tend to use smartphones less, because of poverty and also because they are unaware how useful apps such as Live Transcribe can be.

A government does not have to authorise a person with hearing disability to use Live Transcribe. Short of blocking Google or the Play Store, there is nothing a government can do to prevent such use. That also means that there is nothing to stop data from that person's conversations from being used to "train" the AI that powers the app which helps a disabled person understand what is being said. The more the Bengali app is used, the better trained the underlying AI will be. Over time, real-time transcription will improve. 

The point is that AI is seeping into our lives whether we like it or not; whether a strategy is in place or not.

Why does a country (or a company) need an AI strategy? It needs a strategy to proactively adopt AI in its business processes, to position itself as a supplier of AI technology to other users, or both. Even now, Bangladeshi firms must be considering how to use AI to improve business processes and enhance their bottom lines. But there is little incentive for most businesses and the consultants who advise them to talk about the preconditions for effective absorption of AI in organisations. 

This has historical precedent. We first talked about adopting computers back in the 1980s; we addressed the supply side later. But with a large youth population (41.6 million in the 15-29 age group in 2015; now it should be higher), many with ICT credentials, can Bangladesh afford to wait?

Demand side

What is AI? It is machines that show some behaviours that mimic human intelligence. These days, what we have are narrow AI in specific domains. It is based on deep learning wherein the software is trained on massive amounts of domain-specific data. AI can make decisions or advise those who are making decisions on creditworthiness; assist judges on bail and sentencing, based on correlations that suggest the likelihood of recidivism; or diagnose medical conditions faster and with fewer errors. AI has been used to generate Tang Dynasty poetry by a Sri Lankan data scientist. The possibilities are endless.

The rules by which the software reaches its conclusions are opaque, so the results have to be verified against ground truth. Issues of bias or error have to be addressed. What works in California or with Chinese males may not necessarily work in Bangladesh. AI engineers are needed to customise software for specific conditions. 

So it's qualitatively different from when we started using computers. One could just buy spreadsheet software and use it with a little training. But AI requires skilled engineers to customise for best results, or least to minimise errors. In some cases, additional training data may be needed. Where are these engineers? How will they keep up with the rapidly changing state of the art in AI? What can be done about ossified curricula in computer-science programmes? Where are the data? Are they in "datafied" form amenable to analysis? When the data are about people, one would also need data-protection safeguards. 

All these elements could be addressed through a well-formulated national strategy. It is high time government officials, private sector leaders and civil society initiate a national conversation on the subject.

From big data to AI

Before AI, big data and data analytics were the buzzwords. It was possible to ask a consultant to, for example, run a company's customer records through a "black box" proprietary software to identify the most valuable customers, predict the ones most likely to defect and so on. Today, these things are likely to be marketed as AI. Data cleaning would most likely be necessary. 

Because analysis would be done in-house with internal data, data-protection issues are unlikely to crop up. It would be good to have a data scientist on staff, but not essential. Though the underlying software would most likely be open source, the consultants would have little incentive to open the black box unless the company or the in-house data scientist insists.

Things would get more complicated if the company starts working with external data sets, such as when it seeks to gain insights for marketing. Here, there are issues of representivity (does the data accurately depict the target population?) and also limitations, if any, on how the data may be used (for example, is the data pseudonymised or anonymised? If it is the former, patterns can be identified, even if the individual cannot be). There would be a greater necessity for domain knowledge and possibly also for in-house expertise in analytics.

For AI, training data is critically important. Many experts believe that China will lead in AI because of the greater availability of training data (China's digitalisation is highly advanced, for example in payments and facial recognition). Europe is likely to lag behind it because of excessive restrictions on data use and consent requirements.

Supply side

Barriers to entry are relatively low in most IT domains, including in AI. Given the need for training data and skilled AI engineers even on the demand side, it makes sense to also explore the opportunities of becoming suppliers of AI solutions and AI-infused products. Now, as many governments are considering data-protection legislation, the time is opportune to adopt national AI strategies.  Otherwise, Bangladesh may find the many opportunities of AI foreclosed by short-term considerations associated with doing business with Europe.

Developing and implementing policy requires resources and skills. When both are scarce, prioritisation is even more important. When is the right time to develop AI policy? Unless the conversation is started now, it may be too late for Bangladesh's youth and businesses.

Professor Rohan Samarajiva is Chair of the ICT Agency, the apex body for ICT within the government of Sri Lanka, and founding Chair of LIRNEasia, a think tank active across emerging economies in South and South East Asia. He served as Director General of the Telecommunications Regulatory Commission of Sri Lanka.