Introduction
Every fiscal quarter, automated writing algorithms produce thousands of corporate earning articles using little more than structured data. Companies such as Automated Insights and Narrative Science can write articles on almost any domain with clean and well-structured data. The machine-written articles have variability, tone, and style, and at times are difficult to differentiate them from human-written articles. The automated writing algorithms offer scale, speed, and labor-saving cost. However, the effects of the compromise on media organizations are nuance and accuracy. Thousands of machine-written and published articles had to be corrected for some kind of error (Diakopoulos). The errors range from relatively harmless like the location of a company to more critical ones like the wrong choice of words.
Problems Associated with the Errors
Algorithmic curation is behind the Facebook newsfeed that is one of the most influential news sources. A recent research has shown that Facebook is the news source for approximately 61% of millennials concerning government and politics. The amount of hard news promoted in an individuals’ Facebook news feed has the power to affect the voter turnout in an election. A recent study together with this information has shown that biased Facebook search results can shift the voter preference of undecided voters. This raises questions concerning the extent to which algorithmic curation and ranking systems could impact the democratic voting process in society.
The impact of automated decision making is being felt in virtually all sectors of industry including dynamic product pricing that affect the economy. Uber, the ride-sharing app uses a surge-pricing algorithm that appeals to basic economic theories. The pricing algorithm encourages more drivers to go online so as to try and meet the demand. However, investigations show that instead of encouraging new drivers to get on the road, the pricing algorithm redistributes drivers who are already on the road. Drivers move to neighborhoods with higher surge prices leaving other neighborhoods with inadequate car supply resulting in longer waiting times. The investigations raised concerns regarding which neighborhoods get better services because Uber cars are rival goods. Higher prices and good services for one neighborhood could mean bad services for another neighborhood.
Proposed Solutions to the Problem
Diakopoulos in his article “Accountability in algorithmic decision making” suggests that the solution to the errors is government and private sector accountability in the use of algorithms. Transparency can facilitate accountability and should be demanded from the government and encouraged in industry. However, the mandate for accountability of algorithms is different for the two sectors. In modern democratic states, the government elected by citizens provides social services and exercises power and control through moderated regulations and norms. However, algorithms are largely unregulated and exercise power over citizens without any accountability. The citizens in a democratic society should not accept scenarios where there is no algorithm transparency given the important decision policy they influence.
Although corporations do not have the same mandate as for public accountability as government organizations, they may be forced to act through social pressure. The most convincing argument is that higher data quality will result to a better inference that consequently leads to more satisfied customers. The best way to achieve this is to develop processes that will enhance correction of false negatives by the end user. Allowing customers access to inspect data, dispute, and make corrections on inaccurate labels will improve the data quality in for algorithmic processes.
The reason why both public and private corporations abstract their transparency is because they fear losing their competitive advantage through exposure of trade secrets or revealing their manipulation and game strategies. Complete source code transparency of algorithms could be suicide to corporations in most cases and should not be encouraged. Instead, only certain key pieces of information like aggregate results and benchmarks should be revealed. This will be enough in communicating the algorithmic performance of corporations to the public.
The freedom of Information Act (FOIA) in the United States forces the government to disclose their records when requested unless when the government is in business with a third party system that is protected by trade secrets. There has been at least one successful attempt to compel the government to release its source code. As such, FOIA can be of some importance in dealing with the government use of algorithms. However, FOIA does not require organizations to produce documents that do not already exist. This means that as long algorithm decision was never directly stored in a document, then FOIA cannot compel the government to disclose its source code. To deal with this loophole in the Act, audit trails that record stepwise correlations and inferences could be used. Thus, guidelines should be created to trigger an audit trail whenever an algorithm is used in decision making. To achieve this, Diakopoulos proposes what he calls Freedom of Information Processing Act (FOIPA) that will be an amendment to the existent FOIA (Diakopoulos). FOIPA would mitigate all issues associated with releasing of source codes and allow the public to submit benchmark datasets that would be processed by government algorithms and the results released publicly. This would allow all interested parties to scrutinize government algorithm and benchmarks and discover any discrimination and censorship cases.
Algorithmic Transparency Standards
One question that still lingers is what should be disclosed concerning algorithms. A group of 50 people from the media and academic sector convened to discuss ideas that could support the development of a robust policy for news and information stewardship using algorithms. The group discussed cases of algorithmically enhanced curation and brainstormed on different aspects of various algorithms used that might be publicly disclosed. Based on the generated ideas, five categories of information that might be disclosed were agreed upon (Diakopoulos).
The first was human involvement that included transparency in the purpose and intent of the algorithm, and any human editorial processes, goals and social content that the algorithm was cast. The second was data that creates the opportunity to communicate the quality of data that drives an algorithm. This includes such things as the accuracy, uncertainty, assumptions, limitations and timeliness of data. The third was the model used as input in the algorithms. The model process and tools used in modeling are also made transparent because some software modeling tools have limitations. The fourth is inference that deals with answers questions such as the accuracy and margin of error in made by an algorithm in performing classifications or predictions. The final one is algorithmic presence that dictates the disclosure of algorithm use if and when in use and also alerting the public of any information in a curated experience that have been filtered away (Diakopoulos).
I do believe that the Freedom of Information Processing Act (FOIPA) should be enacted. However, the Act should not only cover the public corporations but also private corporations as well. This is to ensure equity between the private and public sectors. I do support the five categories of information to be disclosed to the public to be requested. As this ensures only the relevant information in regards to the public is disclosed. This will prevent rival organizations from stealing each other trade secrets and game strategies.
Works cited
Diakopoulos, Nicholas. “Accountability in Algorithmic Decision Making.” Communications of the ACM 59.2 (2016): 56–62. Web.