This short opening shows three real examples of what AI can do and how it can go wrong. It takes about 15 minutes. At each question, stop and think before you move on.
Use the Next button or the right arrow key to move through the slides.
Dr. Ali Al Bataineh, Higher Colleges of Technology, aalbataineh@hct.ac.ae
The exchange below was written to show a common kind of AI mistake. Read the answer first and decide whether you would trust it.
What is the compressive strength of standard ASTM A36 structural steel?
ASTM A36 structural steel has a compressive strength of 36,000 to 58,000 psi, depending on the shape and length of the member. Its yield strength is 36 ksi, its tensile strength is 58 to 80 ksi, and its modulus of elasticity is 29,000 ksi.
1 ksi = 1,000 psi (pounds per square inch).
Most of the numbers are correct. A36 steel has a minimum yield strength of 36 ksi, a tensile strength of 58 to 80 ksi and a modulus of elasticity of about 29,000 ksi.
The mistake is the first sentence. 58,000 psi is the lowest tensile strength of A36, the stress at which it breaks when it is pulled, and it does not belong in a range for compression. Engineers use the same yield strength for steel whether it is pulled or squeezed. A short, thick column can carry a load close to that yield point, while a long, slender column fails at a lower load by bending sideways, which is called buckling. In design, the shape and length of a column can lower the stress it can carry, but they never raise it to the tensile strength.
Steel values: AISC/NSBA Steel Bridge Design Handbook, Chapter 1 (2022), and AISC 360-16, Specification for Structural Steel Buildings.
In 2018, Joy Buolamwini of the MIT Media Lab and Timnit Gebru of Microsoft Research tested three commercial face-analysis systems, from IBM, Microsoft and Face++. Each system looked at a photo of a face and guessed whether the person was male or female. The researchers tested them on photos of 1,270 members of parliament from three African and three European countries.
So IBM's system got about one in three of the darker-skinned women wrong. Across the three systems, the error rate for darker-skinned women reached 34.7%, while the highest error rate for lighter-skinned men was 0.8%.
The companies did not say what data their systems were trained on. The researchers did find that two widely used collections of face photos were mostly lighter-skinned people (79.6% and 86.2%). When some groups are rare in the data used to build and test a system, its mistakes on those groups can go unnoticed until someone checks. Nobody has to intend a bias for it to appear.
These were commercial services that developers paid to use, not lab experiments.
Source: J. Buolamwini and T. Gebru, "Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification," Proceedings of Machine Learning Research 81 (2018), 77–91.
With generative design software, an engineer describes what a part must do, such as the loads it must carry and where it must attach, and the software generates many possible shapes that meet those goals.
In 2018, General Motors and Autodesk used generative design to rework a seat-belt bracket. The software produced more than 150 valid designs. The one GM chose replaced eight separate pieces with a single 3D-printed part. It was a test part, not one used in production cars.
The software optimizes only for the goals and limits the engineer gives it. If the loads or the constraints are wrong, it will find a very good answer to the wrong problem, so the engineer's judgment about what the part really has to do still matters.
Sources: Autodesk News, "How GM and Autodesk are using generative design for vehicles of the future" (May 3, 2018); General Motors news release, "Advanced Software Design Technology Leads GM to Lighter, More Efficient Vehicles" (May 22, 2018). The drawings are schematics, not the real parts.
AI can be confident and wrong, it can work better for some people than for others, and it can help find designs that people would not have tried. What do you need to know to use it well?
Next, try the labs. In the object-recognition lab you will see confident mistakes for yourself, and in Grid World you will train a small AI from rewards and penalties.
Hands-On AI for High School · Dr. Ali Al Bataineh, Higher Colleges of Technology