#friction4responsibleai

Humans may think that because AI is pre-programmed to reduce errors, it is accurate and neutral. To counteract these biases we hold about AI and ensure more responsible use of AI, “good friction” can help - this involves introducing strategic obstacles in AI systems that encourage ethical practices, such as greater transparency and human oversight.

I've been a police officer with the Singapore Police Force for more than a decade, and human decision-making has always been an important part of the operational decision-making process, even if we augment our decisions with data. This is partly because data can also be unintentionally biased. For instance, operational data predicting crime rates often face the issue of being self-fulfilling prophecies, as sending officers to patrol an area more often leads to a greater likelihood of detecting crime and having that area listed as a hotspot for more policing. Data fed to AI systems may also not capture everything - data can only capture the measurable and quantifiable aspects of our world, but there are intangible aspects of the world that is difficult to capture in a structured format, e.g. data may not capture the poor security measures in a store that an officer observed last shift and had to advise the store owner on, leading them to question whether that area may become more susceptible to opportunistic crimes like theft.

One of the ways to strategically use friction to achieve more responsible AI systems is, as Gosline mentions, to practice acts of inconvenience for areas where we question: “Should AI be doing this?”. As such, you do not see officers relying on hotspot crime data alone to plan their patrol routes. Here, the friction added is more time required for decision-making for officers for patrol route planning before their shift, but this is a “good friction” that the officer can afford, and it allows the officer to think through their various recent experiences, recent events and data that they are provided on crime in their area. The sum of their human experience, informed by crime data provided by the system undoubtedly provides a better informed decision. Within an organization, it is important to conduct “friction audits” to identify the areas where introducing friction could encourage human-first thinking that benefits the customer and community.

As much as I love Tom Cruise, Minority Report is not happening anytime soon (left). Instead, the best "data" officers can gather about the ground is really just from good old fashioned community policing, and getting to know their neighborhood (right).

Another way we can try to introduce “good friction” is to be on the lookout for “dark patterns”. In my time with the Singapore Police, we’ve identified several lobster traps when it comes to scams. Very often, victims of scams will give away their bank account details, and money gets transferred out of their bank accounts very quickly - but it becomes extremely difficult to get the money back once that transfer is made. In an effort to combat such dark patterns, good friction was introduced in the banking systems, where monetary limits are set on both local and overseas bank transfers, and delays are introduced if a new payee is added to a bank account. There are also frictions that you can impose when your rational System 2 brain is in charge, such a "money lock" feature to secure your money for a set period of time - to prevent any losses when your System 1 brain succumbs to any scams or dark patterns. Such frictions undoubtedly make it more inconvenient for the customer, but they have been accepted so far, as they have been useful in reducing the amount lost to scams and the incidences of scams in Singapore.