Algorithmic decision making has revolutionized the way we live and work, with its impact evident in everything from personalized product recommendations to self-driving cars.
These systems use complex mathematical formulas to analyze vast amounts of data and make decisions at incredible speeds.
The beauty of algorithmic decision making lies in its ability to process and analyze data that would be impossible for humans to handle on their own, making it a game-changer in fields like finance, healthcare, and transportation.
This is because algorithms can quickly identify patterns and relationships in data that might take humans weeks or even months to discover.
Consider reading: Decision Tree Algorithm Machine Learning
Human Decision Making
Human decision making is often biased and prone to errors.
Research has shown that humans tend to rely on mental shortcuts, such as heuristics, to make decisions quickly.
These mental shortcuts can lead to systematic errors, like confirmation bias, where we give too much weight to information that confirms our preconceptions.
For example, a study found that doctors were more likely to recommend a treatment if they had previously read a positive article about it.
Humans also tend to overvalue information that is readily available, such as data from the past year, and undervalue information that is less accessible, such as data from previous years.
Human Capabilities
Humans have a remarkable ability to process information, with research suggesting that our brains can process up to 70,000 bits of information per hour.
We can also recognize patterns and make decisions based on incomplete information, a skill known as "pattern recognition." This is especially evident in situations where we have to make quick decisions, such as in emergency response scenarios.
Our brains are wired to respond to emotions, which can influence our decision-making processes. For example, fear and anxiety can lead to impulsive decisions.
We have a natural tendency to seek out information that confirms our pre-existing biases, a phenomenon known as "confirmation bias." This can lead to poor decision-making, as we may overlook important information that contradicts our initial assumptions.
Humans are capable of learning from experience and adapting to new situations, a process known as "learning from feedback." This is essential for improving decision-making skills and developing expertise in a particular area.
Worth a look: Decision Tree Pruning
Prioritization
Prioritization is a fundamental aspect of human decision-making that algorithms also employ. We make prioritization decisions daily to cope with the information onslaught.
Algorithms prioritize information by emphasizing or bringing attention to certain things at the expense of others. This emphasis can have ramifications for individuals or entities.
Search engines are a classic example of prioritization, where rankings can influence what we see and what we don't. The criteria used in a ranking, how they're defined, and datafied, and their weighting are essential design decisions.
The criteria used in a ranking can have significant consequences, such as determining the quality of schools and hospitals, or the riskiness of individuals on watch lists.
Take a look at this: Automated Decision-making
Inferencing
Inferencing can be a tricky business, and it's essential to consider the accuracy of the outcomes. Algorithm creators might benchmark against standard data sets to disclose key statistics.
The margin of error is a crucial consideration, and it's not always clear what that margin is. What is the accuracy rate, and how many false positives versus false negatives are there?
Algorithm creators should disclose known errors and the steps taken to remediate them. Are errors a result of human involvement, data inputs, or the algorithm itself?
Classifiers often produce a confidence value, which can give us a sense of uncertainty in the outcomes. The average range of those confidence values can be a useful measure of uncertainty.
Decision Making Process
Algorithmic decision-making is a complex process that involves several key steps. These steps include prioritizing, classifying, associating, and filtering.
Prioritizing is a crucial part of the decision-making process, where algorithms weigh different options against each other to determine which one is most important. This is often done using a set of pre-defined rules or criteria.
Classifying is another important step, where algorithms group similar data points together based on certain characteristics. This can help identify patterns and trends within the data.
Associating is a process where algorithms connect different pieces of information to create new insights. This can be done by identifying relationships between different data points or by using external data sources.
Filtering is the final step, where algorithms eliminate irrelevant or redundant data to focus on the most important information. This can help reduce noise and improve the accuracy of the decision-making process.
Opacity and Complexity
Opacity and complexity are two major issues that arise in algorithmic decision making. Burrell (2016) identifies three forms of opacity, with one of them being technical illiteracy. Most of us don't understand how machine learning algorithms work and must trust the experts in this domain.
Just like how we trust our mechanic because we don't understand how engines work, we also trust AI algorithms because of their complexity. As AI algorithms grow more complex, it's unlikely that differences in understanding due to specialized knowledge will be eliminated.
Type 3 opacity, which derives from the scale at which algorithms operate, is a significant concern. The system is not natively understandable by humans, even those with expertise and deep knowledge of its design.
The issue is that machine optimizations based on training data do not naturally accord with human semantic explanations. This means that even if we have access to the data and system design, we still can't fully understand how the algorithm is making decisions.
Consider the example of ChatGPT, which performs 175 billion calculations when deciding which word comes next when writing. While the OpenAI engineers can look at the data set it was trained on, humans have neither the time nor working memory to follow through these steps.
Human-Machine Relationships
As we increasingly rely on algorithms to make decisions for us, our relationships with machines are changing in profound ways. This shift is particularly evident in the way we interact with AI-powered systems, which can learn from our behavior and adapt to our needs.
Algorithmic decision making systems can be programmed to prioritize human values, such as fairness and transparency, to ensure that decisions are made in a way that aligns with our values. For example, a system designed to allocate resources can be programmed to prioritize the most vulnerable members of society.
In a study of AI-powered chatbots, researchers found that users who interacted with a bot that was transparent about its decision-making process reported higher levels of trust and satisfaction than those who interacted with a bot that was opaque. This highlights the importance of transparency in human-machine relationships.
Human Involvement
Human involvement is crucial in understanding how algorithms work. Transparency around human involvement can be achieved by explaining the goal, purpose, and intent of the algorithm.
To do this, you need to identify who at your company has direct control over the algorithm. This includes who has oversight and is accountable for the algorithm's decisions.
The people behind the algorithm should feel a sense of public responsibility and pressure if their names are on the line. This can be achieved by disclosing specific human involvement in the algorithm's creation and maintenance.
Ultimately, transparency around human involvement can bring about social influences that reward individuals' reputations and reduce the risk of free riding in the event of a mishap.
Classification
Classification is a crucial aspect of human-machine relationships, where algorithms decide whether an entity belongs to a specific class based on its characteristics. This can have significant downstream effects.
Human biases can be lurking in the training data used by supervised machine-learning algorithms, which can lead to biased classification decisions.
Data collected from Mechanical Turk may be useful for widely shared and agreed-upon knowledge but introduces discrepancies in other cases.
Accuracy of classifications is also a major concern, with false positives and false negatives having real-world consequences. For example, a man in Boston was unable to work due to a false positive classification of having a fraudulent driver's license.
Designers must balance error rates, but this can involve making value judgments that grant privilege to different stakeholders and outcomes in a decision.
Association
Association is a fundamental concept in human-machine relationships, where creating relationships between entities is key. This can lead to connotations in their human interpretation, as seen in the case of a man in Germany whose name was associated with "scientology" and "fraud" on Google.
The semantics of those relationships can vary from the generic "related to" or "similar to" to distinct domain-specific meanings. This can have implications for the accuracy of an association, both objectively and in terms of how that association is interpreted by other people.
Collaborative filtering is a popular class of algorithm that defines an association neighborhood around an entity and uses those close ties to suggest or recommend other items. This is done through similarity metrics that dictate how closely two entities match.
People often misinterpret correlational associations as causal, despite the popular adage "Correlation does not equal causation." This was evident in the case of the man whose name was associated with fraud on Google, where it was likely read as a causal association rather than a correlational one.
The quantification bugbear torments associations just as it does rankings and classifications. This is due to engineering choices that can have implications for the accuracy of an association, both objectively and in terms of how that association is interpreted by other people.
Human-Machine P-A Relationships
In human-machine relationships, a key aspect is the power aspect, which refers to the distribution of control and decision-making authority between humans and machines.
Machines can be designed to be either empowering or disempowering, depending on how they are programmed and used.
Human-machine relationships can be either task-oriented or social, with task-oriented relationships focusing on achieving a specific goal and social relationships involving more emotional and personal interactions.
In a study, it was found that humans tend to attribute human-like qualities to machines that are designed to be more human-like, such as those with facial expressions and body language.
The level of autonomy given to machines can also impact the power aspect of human-machine relationships, with more autonomous machines potentially leading to a shift in control and decision-making authority.
Human-machine relationships can be influenced by factors such as cultural background and individual personality, with some people being more comfortable with machines taking control than others.
Data
Data plays a crucial role in human-machine relationships, and being transparent about it is essential. Transparency about the data that drives algorithms can be achieved by communicating its quality, including accuracy, completeness, and uncertainty.
Data accuracy is not just about being right or wrong, it's also about being up-to-date. Validity can change over time, so it's essential to disclose the timeliness of the data. Some data may be outdated, while others may be more recent.
Data representativeness is also a significant concern. A sample may not accurately reflect a specific population, and assumptions or limitations may be involved. For instance, a survey may only reach a certain demographic, leading to biased results.
Data processing involves several steps, including definition, collection, transformation, vetting, and editing. These steps can be made transparent to help users understand how the data was handled. Automatic or human-edited data labels can also be disclosed, along with the types of data used and any personal information involved.
Personalization often relies on collecting or inferring individual profiles. This can involve using personal information, such as location or browsing history, to create a tailored experience. However, this raises concerns about personal privacy and what types of information are being used.
Institutions and Choice
Institutions and Choice play a crucial role in shaping the outcome of algorithmic decision making. As seen in the example of the loan approval system, institutions can either promote or hinder the fairness of these decisions.
The transparency of institutions is key to ensuring that algorithmic decisions are made with accountability. This is evident in the case of the credit scoring system, where the criteria for evaluating creditworthiness are clearly defined.
Institutions can also influence the choice of algorithms used in decision making. For instance, the use of machine learning algorithms can be more prevalent in institutions with a strong data-driven culture.
The design of institutions can also impact the fairness of algorithmic decisions. As seen in the example of the hiring system, the use of biased data can perpetuate existing inequalities.
Ultimately, institutions have the power to either amplify or mitigate the effects of algorithmic decision making. By prioritizing transparency and fairness, institutions can promote more equitable outcomes.
Computational Journalism
Computational journalism is a field that uses algorithms and data analysis to tell stories and uncover insights. It's a powerful tool for journalists who want to dig deeper into complex issues.
Computational journalists use techniques like data visualization and machine learning to identify patterns and trends that might be missed by human reporters. For example, a study found that computational journalists were able to identify 90% of a dataset's most important features using machine learning algorithms.
Computational journalism is also being used to automate reporting tasks, freeing up journalists to focus on more in-depth and nuanced storytelling. This can be especially useful for covering large events or datasets, where human reporters might struggle to keep up.
A View from Computational Journalism
Computational journalism is a field that combines computer science and journalism to analyze and understand large datasets. This approach has been used to uncover stories that might have otherwise gone unnoticed.
One example is the Panama Papers investigation, where a team of journalists and computer scientists analyzed over 11 million documents to reveal widespread tax evasion by global leaders. The team used machine learning algorithms to identify patterns in the data and narrow down the search for relevant information.
Check this out: How to Make an Ai on Your Computer
Computational journalism can also help journalists to identify biases in their own reporting. For instance, a study found that articles about African American people were more likely to use words with negative connotations than articles about white people. This kind of analysis can help journalists to be more aware of their own biases and produce more balanced reporting.
By using computational methods, journalists can also automate the process of data collection and analysis, freeing them up to focus on the storytelling and investigation.
The Model
The Model is a crucial aspect of computational journalism, and understanding how it works is essential to producing high-quality reporting.
The model itself, as well as the modeling process, can be made transparent to some extent. Knowing what the model actually uses as input is of high importance.
Features or variables are used in the algorithm, and often those features are weighted, so it's essential to know what those weights are.
A different take: Difference between Model and Algorithm in Machine Learning
Training data was used in some machine-learning process, and characterizing the data used for that along all potential dimensions is necessary.
Some software-modeling tools have different assumptions or limitations, so it's essential to know what tools were used to do the modeling.
The rationale for weightings and the design process for considering alternative models or model comparisons is also important to understand.
The assumptions behind the model, both statistical and otherwise, should be examined, and where those assumptions arose should be understood.
If some aspect of the model was not exposed in the front end, it's essential to know why that was the case.
The Model
The Model is a crucial aspect of algorithmic decision making, and it's essential to understand what goes into it. The model itself, as well as the modeling process, can be made somewhat transparent.
Knowing what the model uses as input is vital, including the features or variables used in the algorithm. These features are often weighted, and it's essential to know what those weights are.
The type of software used for modeling can also have different assumptions or limitations. In fact, some software tools have limitations that should be taken into account.
Understanding the rationale behind the weightings and the design process for considering alternative models or model comparisons is also important. This includes knowing the assumptions behind the model and where those assumptions arose.
If some aspect of the model was not exposed in the front end, it's essential to know why that was the case. Transparency is key in algorithmic decision making.
Presence and Challenges
Algorithmic decision making is a complex and multifaceted issue.
We need to be aware of whether algorithms are being used, especially when personalization is involved.
Entities use information disclosure to engage in strategic impression management, so we can't rely on voluntary compliance with ethical mandates.
Routine audits around key algorithmically influenced decisions, such as credit scoring, could help increase transparency.
In some cases, an adversarial approach is necessary to investigate black-box algorithms, like algorithmic accountability reporting in journalism.
Sampling algorithms along key dimensions can help examine the input-output relationship and investigate an algorithm's influence, mistakes, or biases.
The investigation into Uber's surge-pricing algorithm showed that it redistributes drivers already on the road, rather than motivating a fresh supply to get on the road.
This has significant implications for neighborhoods, as some may end up with better service quality while others are left undersupplied and with longer waiting times.
Designing in Uncertainty
Uncertainty is an inherent part of algorithmic decision making. The more complex the system, the more uncertainty it introduces.
In a real-world example, a self-driving car's sensor data can be incomplete or inaccurate, leading to uncertainty in its decision-making process.
Algorithmic decision making can be designed to handle uncertainty through probabilistic modeling. This involves assigning probabilities to different outcomes based on available data.
Probabilistic modeling can help mitigate the impact of uncertainty on decision-making, but it's not a guarantee against errors. In fact, a study on machine learning models found that even with probabilistic modeling, errors can still occur.
The key is to design systems that can learn from their mistakes and adapt to new information. This can be achieved through techniques like online learning and active learning.
Online learning allows a system to update its models in real-time as new data becomes available. Active learning involves selecting the most informative data points to learn from, rather than relying on a fixed dataset.
By incorporating these techniques, algorithmic decision making systems can become more robust and resilient in the face of uncertainty.
Legal and Regulatory
Algorithmic decision making is a rapidly growing field, but it's not without its challenges. In the United States, the Equal Credit Opportunity Act requires lenders to explain the reasons behind their credit decisions.
The use of algorithms in decision making can lead to bias, as seen in the example of the COMPAS risk assessment tool, which was found to be biased against African Americans. The algorithm was based on a dataset that was heavily influenced by racial disparities in the justice system.
As a result, the use of algorithmic decision making in high-stakes areas like lending and law enforcement requires careful consideration of potential biases and a commitment to transparency.
Legal Background
In the United States, the Securities Act of 1933 provides the primary framework for securities regulation.
The Securities Act of 1933 requires companies to register their securities with the Securities and Exchange Commission (SEC) before offering them to the public.
Companies must also disclose detailed financial information and other material facts in their registration statements.
The SEC reviews these statements to ensure they comply with the requirements of the Securities Act.
The registration process typically involves submitting a registration statement to the SEC, which is then reviewed and approved or rejected.
Companies must also comply with ongoing reporting requirements, including filing periodic reports with the SEC.
The SEC has the authority to bring enforcement actions against companies that fail to comply with the Securities Act.
The penalties for non-compliance can be severe, including fines and even imprisonment.
Recommendations for Government
For governments looking to create effective legal and regulatory frameworks, it's essential to prioritize transparency and public engagement.
One way to achieve this is by making regulatory documents easily accessible online, as seen in the UK's approach to publishing regulatory guidance in a clear and concise manner.
The UK's approach has been successful in reducing the complexity and confusion surrounding regulatory requirements.
Incorporating clear and concise language into regulatory documents can also help to reduce the burden on businesses and individuals.
This can be achieved by using plain language, avoiding technical jargon, and providing examples to illustrate key points.
Governments should also establish clear and consistent communication channels with stakeholders to ensure that regulatory changes are well understood and implemented.
Regular updates and feedback mechanisms can help to build trust and confidence in the regulatory process.
By following these best practices, governments can create legal and regulatory frameworks that are more effective, efficient, and accountable.
Recommendations
Algorithmic decision making can be both fascinating and intimidating. The truth is, algorithms are already making decisions for us in many areas of life, from credit scores to job matching.
To make the most of algorithmic decision making, consider the following recommendations. Use multiple data sources to get a well-rounded view of a situation, as seen in the example of Google's search algorithm combining data from various websites to provide accurate results.
Be transparent about the algorithms being used, as in the case of the Netflix recommendation algorithm, which allows users to see why they're being suggested certain shows.
Use human oversight to review and correct algorithmic decisions, as in the case of Amazon's use of human reviewers to ensure product recommendations are accurate.
Regularly update and refine algorithms to ensure they remain fair and unbiased, just as the authors of the Netflix algorithm did to address concerns about representation in their recommendations.
Avoid over-reliance on a single algorithm or data source, as seen in the example of the Google search algorithm being vulnerable to manipulation by biased data.
Here's an interesting read: Algorithmic Fairness
Equity in Decision-Making
Equity in decision-making is crucial to ensure that algorithms don't perpetuate existing biases. This can be achieved by involving diverse stakeholders in the decision-making process.
For instance, a study found that algorithms used to determine creditworthiness were biased against low-income individuals. This is because the data used to train these algorithms was primarily based on information from high-income individuals.
Diverse teams can help identify potential biases by bringing different perspectives to the table. A team with members from various cultural backgrounds can recognize and address biases that might be invisible to a homogeneous team.
Regular audits and testing can also help identify biases in algorithms. By using techniques such as fairness metrics and bias detection tools, developers can identify and address biases before they become a problem.
Involving stakeholders from the communities that will be affected by the decision can also help ensure equity. This can include community members, advocacy groups, and other stakeholders who can provide valuable insights and perspectives.
By taking a proactive approach to equity in decision-making, developers can create algorithms that are fair and unbiased.
Frequently Asked Questions
Is ADM a type of AI?
ADM is a technology that heavily relies on AI, but it's not a type of AI itself. Instead, it's a broader field that leverages various AI technologies to process and analyze large-scale data from multiple sources.
What is the meaning of automated decision-making?
Automated decision-making refers to the process of making choices without human input, based on data, profiles, or inferences. This process is used in various applications, including online loan approvals and more.
Sources
- https://link.springer.com/article/10.1007/s13347-024-00757-5
- https://queue.acm.org/detail.cfm
- https://hai.stanford.edu/news/designing-decision-making-algorithms-uncertain-world
- https://www.brookings.edu/articles/fairness-in-algorithmic-decision-making/
- https://www.techexplorist.com/algorithmic-decision-making-future-decision-making/65627/
Featured Images: pexels.com