Does AI Understand?

Don't diss my parrot.

Some claim that AI is becoming smarter than humans. Others claim AI is only autocomplete on steroids. Does AI understand? Ultimately, the answer depends on your definition of “understand”.

“I understand” suggests an ideal but unrealistic state where there are no more questions, no unknowns, no misunderstandings. I cannot imagine many fields of study that are fully understood. Increasing our understanding of a subject, ironically, only increases the number of unanswered questions. But that does not make “understanding” a meaningless word.

Imagine that I am explaining how honey bees navigate to an engineering student. I ask this student, “Do you understand?”, and her response is “I understand”. That simple response can have one or more meanings. I order them below from weak to strong.

1. I accept what you say is true because you are famous (celebrity-endorsed social learning)

2. If asked to explain what you said, I could repeat it back verbatim (parroting)

3. I understand what you said only to the extent I understand the metaphors used in your explanation (metaphorical understanding)

4. It all makes sense! I am having an eureka moment. I feel an epiphany. (affective understanding)

5. I accept what you say is true because I recognize your expertise (mentor-advised social learning)

6. I understand well enough to apply your information in innovative and useful ways (applied understanding, e.g., understanding how to use a cell phone)

7. What you said, while new to me, is consistent with and builds on my existing understanding (bootstrapped or ratcheted understanding)

8. I believe I could repair the system you describe (empirical or clinical understanding, e.g. understanding how to repair a cell phone)

9. I believe I could build the system you describe (model-based understanding, e.g., understanding how to build a cell phone)

All of these senses of “understand” apply to me. As an engineer in training, career, and temperament, I focus on #7 and #9 and am motivated by #4. I also respect the opinions of mentors (#5) and know the difference between them and celebrity endorsements (#1). I am going to argue that AI satisfies only #2 [1] but it does so better than humans. It is fair to describe ChatGPT as autocomplete on steroids.

If I asked you to write an essay comparing Romeo and Juliet with West Side Story, you might notice similarities in their structure.

• Romeo is to Juliet as Tony is to Maria

• Capulets are to Montagues as the Jets are to the Sharks

• Juliet is to her nurse as Maria is to Anita

If I ask the same question to Google, I get almost 2 million hits and lots of essays on this topic. If I ask ChatGPT or Bing AI, I get a very reasonable essay. However, it synthesizes that essay from those 2 million hits that Google returned. AI is parroting an answer. It is not creating a single original idea. Using only the text of Romeo and Juliet and the screenplay of West Side Story, ChatGPT could not produce a meaningful essay. It does not understand the meaning of words, only the statistical relationships between words.

A Google engineer claimed that Google's AI chatbot LaMDA was sentient [2], in part, because it said it enjoyed “spending time with friends and family”. First, LaMDA doesn't have any family or friends. Second, it does not feel pain or joy. It does, however, repeat what it has “learned” from other humans. I resist calling LaMDA a glorified parrot because that disparages parrots. Parrots have family, friends, and they avoid pain. Parrots are capable of mimicry (aka parroting) but they can learn on their own in ways that LaMDA or other LLMs cannot. Parrot intelligence is real, not artificial.

IBM commercialized Watson, the AI program that championed Jeopardy!, into IBM Watson Health in 2015 to assist physicians in the diagnosis of diseases. On January 2022, IBM sold the unprofitable business to a private equity firm. Critics claimed that “its supercomputer-aided analysis of health data merely compiled existing knowledge without producing new insights” [3]. Game shows like Jeopardy! are based on providing simple answers to short questions. Making sick people healthy is complex and context-dependent. People may get the same finite set of diseases, but everyone gets sick differently, so treatment needs to be tailored to each patient. Doctors need to memorize a lot of facts, but they also need to be creative. AI is not creative.

Should you worry if AI will replace you in your workplace? If your job requires judgement or creative problem solving, you are safe…for now.

Should you use LLMs in your work? Sure! Every job has elements that are tedious and derivative. I love how my grammar checker uses AI to suggest alternative wording for those sentences that I write in a passive voice or are unnecessarily wordy.

But remember that AI is derivative. It only looks backwards, not forward.

---

[1] AI captures statistical regularities in the data used to program it. If you consider statistical regularities as metaphors for the laws of nature, then one could argue that AI enacts a form of metaphorical understanding [#3]. Like parroting, metaphorical understanding is an extremely weak form of understanding. The differences between these forms of understanding is elaborated in The Book.

[2] "Google fires software engineer who claims AI chatbot is sentient", The Guardian, July 23, 2022, https://www.theguardian.com/technology/2022/jul/23/google-fires-software-engineer-who-claims-ai-chatbot-is-sentient, downloaded 10/22/2023.

[3] “AI Dream Fails”, Science, Vol 375, Issue 6579, January 28, 2022.