How does AI handle uncertainty?

Probabilistic reasoning

Probabilistic reasoning is about calculating the likelihood of different outcomes based on data This approach allows AI to make educated guesses when certainty is impossible. Think of it like rolling a dice. If you know the dice has six sides, you’d reason that each number (1 through 6) has an equal chance of being rolled: around 16.7%. If you started seeing that some numbers show up more often than others, you’d adjust your expectations accordingly. Probabilistic reasoning in AI works similarly, but instead of a dice, it’s dealing with data, lots and lots of it. This reasoning uses probability distributions, which map out all possible outcomes and the likelihood of each. Some common tools AI uses to handle uncertainty are:

Bayesian inference

A method where AI updates its predictions based on new information, similar to how we change our opinions as we get new facts.

Markov decision processes

A mathematical model that helps AI decide the next step in a sequence of actions based on the probability of different outcomes.

Hidden Markov models

Used when some parts of the situation are unknown or “hidden.” These models help AI make guesses about hidden variables by analyzing what it can see.

Imagine an AI system built to predict stock market trends. The stock market is influenced by countless factors, from global news events to investor psychology, making it inherently unpredictable. To handle this uncertainty, an AI model would use probabilistic reasoning to evaluate how likely certain trends are based on past data. When new data comes in the model updates its predictions, recalculating the probabilities of different market directions.

Bayesian inference helps the AI adjust, improving its ability to “hedge its bets” about which way the market might move next. The AI doesn’t give a 100% “yes” or “no” answer. Instead, it might say, “There’s a 70% chance the market will go up tomorrow based on recent patterns”.

Dealing with unknowns

One major hurdle in handling real-world complexities is the concept of hidden variables. In many situations, some important details remain unknown, creating blind spots for AI. Let’s say you’re trying to identify potential diseases based on symptoms. The model might be pretty accurate, but what if a patient has a rare underlying condition that isn’t in the dataset? This hidden variable could skew the decision-making, resulting in less accurate predictions.

The Hidden Markov models and other advanced probabilistic methods help AI cope by allowing it to account for possible “unknowns” or missing information In practice these models are not perfect. Handling hidden variables remains one of the toughest challenges in AI. When AI makes a prediction, it often attaches a confidence level to it. This confidence level helps humans understand the model’s level of certainty and assess how much to trust it. High confidence is great, but there’s a trade-off. If AI sticks only to high-certainty answers, it will miss out on a lot of potential predictions.

For instance, in a medical setting, a model might have low confidence about a rare diagnosis, but if it only reports high-confidence predictions, it could miss an important possibility. To find balance, some AI systems use a technique called confidence calibration. They adjust predictions to better reflect real-world accuracy. This way, the model knows when to provide more cautious suggestions and when to be more assertive, based on historical accuracy.

Real world examples

Self-driving cars

Autonomous vehicles use probabilistic reasoning. If a sensor picks up a blurry object, the car has to calculate whether it’s more likely a pedestrian or a shadow and respond accordingly.

Spam filters

Ever wondered how your email’s spam filter “decides” what to flag as spam? It uses probabilistic reasoning to calculate the likelihood that an email with certain words, phrases, or attachments is spam. It does not to be 100% certain, it just needs a high enough probability to separate spam from non-spam.

Voice assistants (like Siri or Alexa):

When you ask a voice assistant a question, it doesn’t always know exactly what you mean. But it’s able to handle uncertainty by estimating the probabilities of different interpretations. For instance, if you say, “Play the new song by [artist],” it might look for recent releases and guess what you mean, even if it’s not entirely certain.

The limits of probabilistic reasoning

Although probabilistic reasoning is powerful, it has its limitations. Sometimes, real-world situations are just too complex, chaotic, or novel for AI to handle with high accuracy. An unexpected natural disaster, for instance, might overwhelm an AI weather prediction model since it’s based on historical data that doesn’t account for all future possibilities. The quality of an AI’s predictions depends heavily on the data it’s trained on. If the training data doesn’t cover enough diverse situations, its probabilistic reasoning will be limited in scope, reducing its accuracy when encountering something entirely new.

Understanding how AI handles uncertainty can help us appreciate both its capabilities and limitations. When we know that AI relies on probabilistic reasoning, we can better grasp why it might give a “best guess” rather than a certain answer. And in situations where it’s important to get a 100% correct answer, we know to seek human expertise.

AI’s probabilistic reasoning is actually a bit like ours, it can handle a lot of uncertainty, make educated guesses, and learn over time. But it doesn’t always get it right. It needs high-quality data, diverse experiences, and continuous updates to keep making the best possible decisions.