Artificial intelligence (AI) is often talked about as a potential incredible productivity boost for workers or a job replacement. In the current headlines, you will see many Hollywood writers and actors going on strike over a number of issues, including the use of artificial intelligence technology to replace human writers and, in the future, actors. On the other hand, you might see posts and videos on social media sites where creators are advocating that other people start using artificial intelligence tools to make money using "get rich quick" style schemes. Often, these articles or social media posts are sensationalizing the impact of these technologies on relative beginners in a field. Typically, the tools discussed in these articles and posts are large general-purpose models that are advertised as general-purpose and capable of handling several different tasks. What is being talked about in most of the hype is not what most companies using AI models are actually basing their systems on. With all this information, should you be afraid of artificial intelligence becoming a problem? Replace artist? Take your job?
To understand these issues, it is important to understand what AI use cases are already being used in your daily life. Most actual "AI" that companies have built and continue to build is narrower in scope, making the models better than those larger models at the specific tasks they were designed to handle. These more focused models are built using a technique called machine learning (ML) and are often used for tasks that most people take for granted.
ML at its simplest uses curated data to tune statistical models. To get an idea of how widespread this use is, I'll give a few examples of where these models are used.
- Credit card companies use large datasets of customer transactions and a small subset of records that show fraud to train ML models to determine whether new transactions for a given account are fraudulent or legitimate.
- Email providers use large email data sets that classify emails into primary, promotional, and social categories to train ML models to determine which category new emails fit into.
- E-commerce websites use large datasets of items that users have purchased to train ML models to recommend items that users might be interested in viewing or purchasing next.
- Rideshare company uses large dataset of trips, drivers, and users to train ML model to determine how much a ride should cost between 2 places
Most of these examples happen all around you, providing value to consumers without taking a person's job or causing problems. Having said that, applying ML without appropriate ethical considerations, supervision and care can cause or at least exacerbate problems.
Some examples of ML accidents are:
- Models trained on historical crime data end up perpetuating bias — AI is sending people to prison — and getting it wrong | MIT Technology Review
- Model trained on historical real estate prices costs Zillow $88,188.1 million in one year — Zillow’s shuttered house flipping business lost $<>.<> billion in 2021 — Wall Street Journal
These two problems with ML highlight 2 potential problems with AI and ML that have existed and will continue to exist in the future.
The first problem is that AI and ML require training data (examples) and a lot of training data. This usually means that models are trained with existing biases that are then reflected in the results they output, or more simply: garbage in means garbage out.
The second problem is that the results of AI and machine learning models are often viewed as unquestionable by “data-driven” leaders because they have been trained on the data they must understand better than humans. This is often not the case and supervision is required, especially when edge cases occur that the model was not trained to handle. More simply, a model trained to predict the market using only rising markets will never predict falling markets, although if left unchecked it may predict a lot most of the time when market conditions slow down or change , which can have dire consequences.
So, to summarize the question “Should I be afraid of AI?”, the answer depends on how AI is implemented and what the controls are around it. For many narrow use cases, you may already actively interact with AI in your daily life. To the extent that the AI is trained on high-quality data, under the supervision of a team that carefully considers the impact and effectiveness of its models, you shouldn't worry too much about AI. Bigger problems can arise when teams accidentally train models and implement them without proper oversight or ethical considerations, which can be cause for concern in artificial intelligence.