We’re increasingly relying on artificial intelligence in our daily lives. There’s potential for vast improvements in the ways we live and work, but there are some important caveats that are crucial to understand, particularly if you’re charged with bringing these tools into your organization:

  1. Machine learning makes AI so powerful.
  2. AI is only as good as the data.
  3. AI is still application specific.
  4. AI has a pretty steep “learning curve.”

Machine Learning Makes AI So Powerful

Machine learning is a subset of AI algorithms that teach a computer to learn to solve a problem on its own. The system doesn’t need to “see” every example. For example, you don’t have to configure a robot vacuum cleaner with a map of every possible space. You can drop it in any space and it knows how to figure out the map itself. This is quite powerful, but it does have two very important constraints. First, these systems learn on data examples of “right and wrong” answers, which makes them highly data dependent. Secondly, they require a specific set of instructions on how to learn, which makes them quite literal. They have no nuance (i.e., human judgement).

Takeaway for business leaders: Machine learning is a powerful shortcut that allows computers to solve complex problems without having to “see” every possible iteration, but there are important caveats.

AI is Only as Good as the Data

The reason why AI applications are exploding right now is because of the vast amounts of data available to “train” them. Chat GPT can review hundreds of thousands of fairy tales before creating a completely new tale about your nephew with astounding accuracy. However, AI algorithms can’t “see” anything that isn’t in the data, and they have no prior context to identify incorrect or biased data. This is why, for example, hiring algorithms can easily be biased if they’re not explicitly instructed not to be. Even if you remove explicit gender identifiers, the algorithms find a way to replicate the “successful” outcomes of the past. They will find women’s advocacy groups or women’s colleges deep in a resume, identifying that as a reason why similar applicants were rejected in the past, and why this applicant should be rejected.

Takeaway for business leaders: You have an important role in developing applications of AI. You know the data and the business processes that create it. You can help the developers account for potential biases in the data.

AI is Still Application Specific

For those that have a healthy paranoia about Skynet (Terminator) or Synthients (Matrix), the good news is that AI is pretty far from replicating, let alone surpassing, human intelligence. To date, there’s not one single “AI,” but rather there are many different algorithms coded individually for each problem. Practically, this means that Chat GBT will be great at writing a newspaper article but can’t drive a car.

Importantly, humans code each AI application. If the humans miss something in the instructions, the AI lacks the nuanced judgement to account for it. The best example I’ve seen of this was the use of a popular navigation app to flee the California wildfires. Users were directed into the path of the wildfires because the app was coded to only navigate around traffic hazards that created gridlock. The roads consumed by fire, obviously unpassable, were clear of traffic.

Takeaway for business leaders: Human judgement is still needed to guide AI, identifying these unintended consequences and their implications. You’re the best positioned to do this for their business.

AI Has a Pretty Big ‘Learning Curve’

Creators of Chat GPT have released the open-source code for free specifically to test for unintended consequences. They have programmed safeguards against the more nefarious uses of the code and want to see how people find ways around them. They’re crowdsourcing these “unintended consequences” before the application could be used for something with more serious consequences.

Coding AI involves training the algorithm on a set of available data with known outcomes, and then repeatedly showing it “new” sets of data to see how it reacts. Exceptions and nuances are identified and accounted for in continuous calibration of the algorithm. Human feedback is essential. As with Chat GPT, only humans can explain why the output wasn’t right. This means it may take time for the AI to be fully operational.

Takeaway for business leaders: Learning time needs to be planned into any implementation. You should define the minimum acceptable accuracy, and make sure you understand the cost to improve, both in terms of time and resources.

AI will change the way we do business, but there are some important caveats. Business leaders have an important role to play in implementing and operationalizing AI, as they’re the ones that can best provide the business context.