As it turns out, yes, but just barely

AI is ubiquitous. One of the most common places for AI in the home these days is the smartphone, smart homes and appliances. Recently, as we sat down to dinner, we had a conversation with Alexa that went something like this:

Me: Alexa, play Cubs highlights. [This is a feature that’s touted on Alexa’s home screen. Ask Alexa to play highlights for your favorite team.]

Alexa: I found something on the web. [I know when Alexa starts like this, there’s a problem. It’s not going to go well. Alexa shows a list of a handful of videos. Granted, most of them are baseball videos with players making exceptional plays over the past 5 years. My fault. Try again.]

Me: Alexa, show me highlights for the latest Chicago Cubs baseball game. [I’m hoping it doesn’t detect the condescending tone as I accept the blame for it not being able to understand my request.]

Alexa: Highlights for the Chicago Cubs game will be available two hours after the game is over. [Progress. I had no idea that they may be playing at this very moment. Luck is on my side. I’m suddenly hopeful.]

Me: Alexa, fine show me yesterday’s effing highlights, then. [Yes, my frustration is beginning to show. I’m so close to breaking the code. I can almost taste it.]

Alexa: I’m sorry, I don’t know that one. [It says this far too often. Perhaps I wasn’t clear.]

Me: Are you kidding me? Play, video highlights for the Major League Baseball game between the Chicago Cubs and Pittsburgh Pirates for Monday July 25, 2022 at Wrigley Field. [This time I’m confident I’ve nailed it. I’ve spit out a specific, unambiguous request that is a skill I know Alexa has. It has done this before. ]

Alexa: [Silence. Nothing. No response. I’ve forgotten to say the magic wake up word, Alexa.]

The average IQ of an 18-year old is around 100. The average IQ of a human 6-year old is 55. Google AI IQ was evaluated to be 47. Siri’s IQ is estimated to be 24. Bing and Baidu are in the 30’s. I didn’t find an evaluation of Alexa’s IQ, but my experience was much like talking to a preschooler.

Some may say, it’s not fair to give a computer an IQ test. But, that’s perfectly the point. The promise of AI is to do what humans do, only better. So far, each head-to-head – or, shall we say, neural network to neural network –challenge has been very focused. Playing chess. Diagnosing disease. Milking cows. Driving cars. The robot usually wins. What I want to see is Watson milking a cow while driving a car and playing Jeopardy. Now, that would be the trifecta. Humans can’t even look for their cigarettes while they’re driving without getting into an accident.

AI’s IQ

Outwitted by a machine. I suspect I’m not alone. I got to thinking, if this is state of the art, how smart are these things? Can we compare a human’s intelligence to a machine?

Scientists are assessing systems’ abilities to learn and reason. So far, the synthetic humans haven’t done as well as the real thing. Researchers are using the shortcomings to identify the gaps so that we better understand where additional development and progress needs to be made.

Just so that you don’t miss the point and forget what the “I” in AI represents, marketers have now coined the term Smart AI.

Is AI Sentient?

Do robots have feelings? Can computers experience emotions? No. Let’s move on. If you do want to read about it, one (former) Google engine does claim the AI model that Google is working on is sentient. He had a creepy chat with a bot that convinced him that the computer has feelings. The computer fears for its life. I can’t even believe I wrote that sentence. Computers have no life to fear. Computers can’t think. Algorithms are not thought.

I would not be surprised, however, if a computer responds to a command in the very near future with: “I’m sorry, Dave, I can’t do that.”

Where Does AI Fail?

Or, more precisely, why do AI projects fail? They fail for the same reasons that IT projects have always failed. Projects fail due to mismanagement, or failure in managing time, scope or budget..:

• Unclear or undefined vision. Poor strategy. You may have heard management say, “We just need to check the box.” If the value proposition cannot be defined, the purpose is unclear.
• Unrealistic expectations. This may be due to misunderstandings, poor communication, or unrealistic scheduling. Unrealistic expectations may also stem from lack of comprehension of AI tools capabilities and methodology.
• Unacceptable requirements. The business requirements are not well defined. The metrics for success are unclear. Also in this category is the undervaluing of employees who understand the data.
• Unbudgeted and underestimated projects. Costs have not been fully and objectively estimated. Contingencies have not been planned for and anticipated. The time contribution of staff who are already too busy has been underestimated.
• Unforeseen circumstances. Yes, chance happens, but I think this falls under poor planning.

See, also, our previous post 12 Reasons for Failure In Analytics and Business Intelligence.

AI, today, is very powerful and can help companies achieve tremendous success. When AI initiatives fail, the failure can almost always be traced to one of the above.

Where Does AI Excel?

AI is good at repetitive, complex tasks. (To be fair, it can do simple, non-repetitive tasks, too. But, it would be cheaper to have your preschooler do it.) It is good at finding patterns and relationships, if they exist, in vast amounts of data.

• AI does well when looking for events which don’t match specific patterns.
  • Detecting credit card fraud is about finding transactions which don’t follow usage patterns. It tends to err on the side of caution. I have received calls from my credit card with an over-zealous algorithm when I filled up my rental car with gas in Dallas and then filled up my personal car in Chicago. It was legit, but unusual enough to get flagged.

“American Express processes $1 trillion in transactions and has 110 million AmEx cards in operation. They rely heavily on data analytics and machine learning algorithms to help detect fraud in near real time, therefore saving millions in losses”.

• Pharmaceutical fraud and abuse. Systems can find unusual patterns of behavior based on many programmed rules. For example, if a patient saw three different doctors around town on the same day with similar complaints of pain, additional investigation might be warranted to rule out abuse.
• AI in healthcare has had some excellent successes.
  • AI and deep learning was taught to compare X-rays to normal findings. It was able to augment a radiologists work by flagging abnormalities for a radiologist to check.
• AI works well with social and shopping. One reason why we see this so much is that there is low risk. The risk of AI being wrong and having severe consequences is low.
• • If you liked/bought this, we think you’ll like this. From Amazon to Netflix and YouTube, they all use some form of pattern recognition. Instagram AI considers your interactions to focus your feed. This tends to work best if the algorithm can put your preferences in a bucket or group of other users who have made similar choices, or if your interests are narrow.
  • AI has enjoyed some success with facial recognition. Facebook is able to identify a previously tagged person in a new photo. Some early security-related facial recognition systems were fooled by masks.
• AI has enjoyed successes in farming using machine learning, IoT sensors and connected systems.
  • AI assisted smart tractors plant and harvest fields to maximize yield, minimize fertilizer and improve food production costs.
  • With data points from 3-D maps, soil sensors, drones, weather patterns, supervised machine learning finds patterns in large data sets to predict the best time to plant crops and predict yields before they’re even planted.
  • Dairy farms use AI robots to have cows milk themselves, AI and machine learning also monitor the cow’s vital signs, activity, food and water intake to keep them healthy and contented.
  • With the help of AI, farmers who are less than 2% of the population feed 300 million in the rest of the USA.
  • Artificial Intelligence in Agriculture

There are also great stories of AI success in the service industries, retail, media and manufacturing. AI really is everywhere.

AI Strengths and Weaknesses Contrasted

A solid understanding of AI’s strengths and weaknesses may contribute to the success of your AI initiatives. Remember, too, that the capabilities currently in the right-hand column are opportunities. These are the areas in which vendors and bleeding edge adopters are currently making progress. We will look at the capabilities which currently challenge AI again in a year and document the left-shift. If you study the following chart carefully, I would not be surprised if there was some movement between the time I write this and the time it is published.

What’s AI’s Future?

If AI were smarter, it might predict what the future holds. It’s clear that there are many misconceptions about what AI can and cannot do. Many misconceptions and AI illiteracy are the result of tech marketing over-hyping existing capabilities. AI is impressive for what it can do today. I predict that many of the weaknesses in the right-hand column will shift to the left and become strengths in the next 2 or 3 years.

[After I finished this article, I presented the previous paragraph to OpenAI, an open AI platform language generator. You may have seen some of the art generated by its DALL-E. I wanted to know what it thought about the future of AI. Here’s what it had to say. ]

The future of AI is not about buying a few servers and installing an off-the-shelf software package. It’s about finding and hiring the right people, building the right team, and making the right investments in both hardware and software.

Some potential successes of AI over the next few years include:

• Increasing accuracy of predictions and recommendations
• Improving decision-making processes
• Speeding up research and development
• Helping to automate and optimize business processes

However, there are also some potential failures of AI that businesses should be aware of, such as:

• Over-reliance on AI leading to suboptimal decisions
• Lack of understanding of how AI works leading to misuse
• Bias in data used to train AI models leading to inaccurate results
• Security and privacy concerns around data used to train AI models

So, what does this mean for businesses investing in AI to supplement their traditional analytics? The short answer is, there are no short-cuts. 85% of AI initiatives fail. Interestingly, this is similar to oft-quoted statistics related to traditional IT and BI projects. The same hard work that has always been required before you can get value out of analytics must still be done. The vision must exist, be realistic and achievable. The dirty work is data preparation, data wrangling and data cleansing. This will always need to be done. In training AI, even more so. There are currently no shortcuts to human intervention. Humans are still required to define the algorithms. Humans are required to identify the “right” answer.

In summary, for AI to be successful, humans need to:

• Establish the infrastructure. This is essentially establishing the boundaries in which AI will work. It’s about whether the foundation can support unstructured data, blockchain, IoT, appropriate security.
• Aid in discovery. Find and determine the availability of data. Data to train AI must exist and be available.
• Curate the data. When presented with a large data set and, consequently, a large number of potential results, a domain expert may be required to evaluate the results. Curation will also include the validation of data context.

To borrow a phrase from the data scientists, for companies to be successful with AI, to be able to add value to existing analytics capabilities, they need to be able to separate the signal from the noise, the message from the hype.

Seven years ago, IBM’s Ginni Rometty said something like, Watson Health [AI] is our moonshot. In other words, AI – the equivalent of a lunar landing – is an inspirational, achievable, stretch goal. I don’t think we’ve landed on the moon. Yet. IBM, and many other companies continue to work toward the goal of transformative AI.

If AI is the moon, the moon is in sight and it’s closer than it has ever been.