On Saturday (19.9.2020), the Tamilnadu state government had released a policy document on Artificial Intelligence titled "Safe & Ethical Artificial Intelligence Policy 2020".
The office bearers of Free Software Foundation Tamil Nadu (FSFTN) had a discussion on the document on Monday afternoon (21.9.2020).
First, we would like to appreciate the initiative taken by the state government to come up with a policy to regulate the development, auditing and deployment of Artificial Intelligent systems in the public sphere.
The policy highlights the dangers of unethical and poorly designed AI models that doesn't consider various social indicators. The policy makers have also come up with a framework called TAM-DEF which emphasizes Transparency, Auditing, Accountability, protection against Misuse, Digital divide / data Deficit, Ethical usage, Fairness and equity of the algorithmic models.
Data and Algorithms are center to understand how a particular system arrives to a decision. Hence, both of them carry huge significance in Artificially Intelligent systems. The policy has rightly pointed out that any machine learning algorithm that is being developed and deployed cannot be a black-box.
At the same time, FSFTN demands that, not just the algorithms, the actual implementation of those algorithms which are to be deployed in e-governance services must be made available for public audition under one of the FOSS licenses.
This ensures that various FOSS communities, Citizen Science groups and civil society bodies can audit the source code, identify faults or bugs with the implementation and could also help improve the software to serve the public better.
In order to satisfy the TAM-DEF parameters, the policy document puts forth a 7 metric scorecard called DEEP-MAX, according to which all AI/ML system would undergo a phase of test run periodically and 7 parameters of DEEP-MAX will be given a binary score of either 0 when it fails the test and 1 when it passes the test. The scorecard will then be added to the Blockchain infrastructure setup by TNeGA so that they will be tamper-proof.
The policy document says all models will be fed with a standard set of test data to check against the 7 DEEP-MAX parameters, however there is no clarity on the correctness of the test data and the evaluation procedure itself. For example, what is the threshold criteria to determine, say, for example, the algorithm properly takes into account the social discrimination? Will the evaluation too will be left to algorithms? or will it be performed by humans?
Data Collection, Protection and Possibility of Mass Surveillance
Rolling out machine learning algorithms for public services means collecting more and more data about citizens. With a proper Data Protection law not in place yet in our country, what are the measures TN government would take legally to ensure data protection? There is no provision in the policy document that deals with mass surveillance with AI/ML systems. Will these systems be used by the State and Police to hunt down dissenters and protesters is not addressed.
Blockchain and Environmental Impact
The govt. has considered to use Blockchain to store the algorithm scorecard. It is a well known fact that Blockchain's proof-of-work requires high computing power and hence consuming more electricity. Have the policy makers considered the amount of Carbon footprint this will add to the environment? Have they considered any other tamper-proof mechanism to store the scorecard is not clear.
To sum up, FSFTN appreciates the initiative and supports only as long as the power of AI is harnessed for the public good while keeping it safe, ethically compatible with human values and environmental values.