In these days, the media headlines frequently focus on Artificial Intelligence, its uses, potentialities and dangers. In more than one forum we wonder, -and we answer to ourselves- about the use that the tax administration can give to these technologies to better fulfill its mission, both to facilitate compliance and to effectively exercise control. We are not referring here to generative Artificial Intelligence, the one that the papers are talking about today, but to the more “normal” one, the one that seeks to solve classification and regression problems, based on things like machine learning and neural networks. This is, to illustrate the point, like the one that within a photo library  can find those photos in which cats appear and discard those in which they do not appear although from time to time it fails and presents us with a hare. In the administration there is a very large universe of situations that can be classified as “of interest”; forgetful taxpayers, those who forget to declare some things; actor taxpayers, those who simulate operations or make use of documents about simulated operations; erratic taxpayers, those who report things that are actually unintentional errors (and non so non-intentional ones); or in most cases, reliable taxpayers, those who want to fulfill their obligations and whom we have to help so that their actions to comply are not too expensive or their waiting times are not too long when they wait for a refund or a service. Of course, in these cases, failing in the classification is not as innocuous as identifying a cat in a photo when in reality, it was none other than a disguised dog. In the case of the administration, a mistake in the classification can really result, as the saying goes, that they pass us a cat for a hare. There are consequences for the administration and for taxpayers.

Many of the advanced analytics and Artificial Intelligence applications are supported by open tools (Python, Scikit-Learn or TensorFlow); in open collaboration environments on GitHub where projects can be consulted, downloaded or cloned; and by free access of open data sources for algorithm training, such as image sets catalogued as Digi-Face M1; to platforms to manage the work of individuals who collaborate by tagging the data training sets, such as Label Studio; to the algorithms themselves, which allow a new song to be heard on the voice of someone else, and it looks good, like the music of Breezer in the voice of Gallagher. The Electronic Invoice Anomaly Detector it is an open tool built within the framework of the CIAT Advanced Analytics Center with the support of several of our member countries, in whose process we work these two servers, and where Microsoft put the engineering effort.

But, as another old saying goes, no pain, no gain and if you want it light blue, it will cost you. Unlike what happens with the algorithms to classify cats, the algorithms for the specific processes of the tax administration are not all available to download from some cloud. Investments are required and we are not just talking about the costs involved in providing computing capacity in our own data centers or in the cloud. Human resources are required for its development, training, testing, validation and for its constant evolution and adjustment. All this must be done without compromising the satisfaction of the increasingly demanding requirements of traditional applications. Even the largest administrations have difficulties to develop everything in-house  and even if they could, they should not do it: in addition to Artificial Intelligence specialists, data scientists, auditors, economists or lawyers will be needed and share Intelligence. The smaller administrations will have it even more difficult. Undoubtedly, external support is required and the hiring of services for these developments may be inevitable. And then a question arises, how do we hire them? how is it executed? and how is it paid for? Is the problem the same as that of traditional computing or is it a different one?

Let us imagine that it is about developing an algorithm that classifies a situation as a breach or not. We might think that anyone who could distinguish real cats or real credit card transactions from fake ones could do it. Of course, as a tax administration we do not want to share the data with third parties (unlike cat photos), and we would like whoever develops would do so without accessing any of the administration’s data. The researcher’s problem is to do better than tossing a coin in the air or consulting our experts. If they work with synthetic data, invented and out of touch with reality, it may happen that the algorithms, no matter how effective they are with synthetic data, could be useless when applied to real situations and not those reserved for Fantasy Island. If we anonymize the data, the question immediately arises as to whether the anonymized data is really anonymous or whether a third party could, from them, identify actors, or some of them.

Then comes the issue of quality, as in cancer testing. Should we ask for 60% of hits to serve us or should we demand more? What do we call success? It is the same an algorithm that has 1% false negatives (a fraudulent operator requesting a refund is classified as dependable) that 1% false positives (a valid operator requesting a refund is classified as fraudulent). Which is more serious if someone is diagnosed with cancer with the price of a scare or an unnecessary operation or if it is not detected and let go without treatment? How much does it cost in human and machine resources to do each test?

But even if privacy is ensured and if the algorithms seem good with the test cases (few and from past years) we always have the doubt of what will actually happen because it may happen that reality has changed. Artificial Intelligence systems today learn by “playing against themselves,” for example, chess or Go, but tax Artificial Intelligences learn from inspections of taxpayers, which cost taxpayers resources and uneasiness.

A taxpayer has the right to decide whether the person answering is an Artificial Intelligence or a person. He has the right to know the system that assists him, for example, he must know if the Artificial Intelligence that medicates him is specialized or if, to save money, one for veterinary use is being applied and he must have the right to know if he is inspected because there are enough well-founded reasons or if it is because the Administration has this Artificial Intelligence in the “college of Artificial Intelligence” learning and that they classify him as cats.

Artificial Intelligences that select taxpayers should not have biases. The case of the Netherlands is well known, where its use was criticized for the unintentional use of certain variables in the selection of taxpayers, which led to them being, in proportion, selected more from certain neighborhoods and thus from certain communities than from others. How are we going to train our Artificial Intelligence? With data and criteria from the past and let whoever falls fall (strategy that generated the mentioned problem in the Netherlands) or with new criteria? If the criteria are new, who pays for the experiment? The taxpayers that the machine selects by playing against itself.

Solving this is difficult. Do the tax administrations have resources to do Research and Development in this area, or is it better to wait for them to mature?

For our part, talking to each other, we thought that the best thing to do for now, is to involve experts from outside the administration as part of development teams led by someone in the administration. External resources can be replaced by others when they do not “measure up”. It must be ensured that the knowledge stays within the agency and that the use of human, financial and processing resources are reasonable and significantly less than the benefits obtained from their application.

In other words, in order to tame the beast of Artificial Intelligence for its tax use, the best thing that can be done is to use the beasts from outside (there are many and cheap) and the tamer from inside. And let the beast be of our size.

Greetings and good luck.

3,258 total views, 2 views today