Tone-deaf AI advocates need a transparency algorithm
Some in the facial recognition community complain that opponents will never be satisfied there is enough transparency in the development and use of biometric technology.
That argument might be the market’s first truism, but it does nothing to quell demands for openness given the intimate yet largely invisible nature of advanced facial recognition systems.
One system, in the European Union, is designed to read people’s thoughts via so-called micro-expressions that play across a person’s face. IBorderCtrl is the subject of a lawsuit filed by a member of the EU parliament seeking specific information about how the AI algorithm was developed and tested.
The problem is a lack of visibility into the project’s processes that might otherwise salve the fears of civil rights advocates and AI skeptics.
Also, there is nothing approaching consensus on whether transient expressions indicate thoughts or intentions. And it is unknown how AI would interpret ingrained reactions by people of different cultures or the expressions of people with muscle tics, for example.
The EU, according to The Guardian, invested millions of euros into iBorderCtrl to extrapolate thoughts and emotions from even fleeting facial expressions. The three-year project apparently ended in August 2019, after limited border tests. Its site, iborderctrl.eu, is a tombstone.
Maybe iBorderCtrl researchers were on to something with their automatic deception detection system. It is possible that the system could be convincingly accurate, effective in spotting trouble and sensitive to civil rights.
But according to The Guardian article, the information requests by the German member of the European Parliament have been dismissed out of hand, citing the sanctity of trade secrets. Undeterred, the parliamentarian (who is not alone in worrying about the development) has taken the matter to the European court of justice in a lawsuit expected to be heard in 2021.
Online publication The Intercept last year found iBorderCtrl lacking. It noted that at least one of the supporting tests done by the consortium of EU agencies and universities was so small as to be anecdotal — 32 people — and achieved 75 percent accuracy.
Some information from the EU is available.
An EU-funded idea nursery, Horizon 2020, gave €4.5 million to iBorderCtrl in 2016 to develop what proponents bill as a lie detector that watches for the smallest signs of duplicity in expressions.
In materials aimed at generating support for the program, it is marketed as a non-invasive way to speed up border crossings while increasing security.
Part of the system involves subjecting individuals to an initial risk assessment conducted by a putty-faced avatar.