Human and Societal Considerations in AI: When Robots Meet Reality

The AGI Fantasy vs. Narrow AI Reality

The term artificial intelligence was first coined in 1955 at the famous Dartmouth Conference. When the term was first coined, it referred to what we now call artificial general intelligence or AGI - the ambition to create intelligent agents capable of learning any intellectual task that humans can do.

Herbert Simon said in 1965: "machines will be capable within 20 years of doing any work a man can do." 55+ years later, we know this still isn't true.

Yet today it's still a common misperception that when we use the term AI, we're referring to artificial general intelligence. The reality is that all current applications of AI fall under what's called Narrow AI - the ability to accomplish specific pre-learned problem-solving tasks.

AI models (narrow AI) are genius savants - they're incredible on certain tasks that they are trained on, but they can't generalize to new tasks (yet).

The Dog vs Cat Problem: Why AI Can't Think Like Babies

Let's look at an example. Suppose we build a machine learning model to classify breeds of dogs. We might feed it a picture of a golden retriever and a labrador retriever and the model would correctly classify each dog. For some humans, this might be difficult unless you really understood dog breeds. Yet for a machine learning model, this is generally considered an easy task.

Now suppose we fed our model a picture of a cat. As humans we intuitively know this is a cat. Even a two-year-old toddler would correctly recognize this as a cat. Yet our machine learning model trained to recognize dog breeds has no idea what this is. It would likely try to classify it as some type of dog.

As UC Berkeley Psychology Professor Alison Gopnik said: "One of the fascinating things about the search for AI is that it has been so hard to predict which parts would be easier or harder. It turns out to be much easier to simulate the reasoning of a highly trained adult expert than to mimic the ordinary learning of every baby."

How Humans and AI Think Differently

AI models don't get hangry. Here are the key differences:

Memory and Data Processing

Humans: Very limited memory and ability to process data. Given hundreds of thousands of rows of data, we can maybe remember and process a couple dozen rows.

AI Models: Almost unlimited ability to process vast amounts of data. Hundreds of thousands, millions, even billions of data points? No problem.

Decision Factors

Humans: Our decisions are influenced not only by data but also by emotions, biases, and even physical state. Studies show court judges make significantly different decisions depending on whether they're deciding before or after lunch. Our hunger and blood sugar levels impact our decisions.

AI Models: Make consistent decisions no matter what time of day or state. Not influenced by external factors like emotions or physical state.

Mental Models and Reasoning

Humans: We lean on mental heuristics and rules of thumb. We make decisions that fit within our concept of how the world works, so the narrative makes sense to us.

AI Models: Don't have a worldview or concept of how the world should work. They make decisions based solely on data and patterns they're trained on. But AI models can still make biased decisions when trained on biased datasets.

Common Sense and Novel Situations

Humans: We have a unique ability to apply common sense reasoning to novel situations. When we encounter something new, we can apply reasoning we've learned over time.

AI Models: Unable to reason beyond the data they're trained on. Their ability is limited to the specific task they're trained for. When you try to extend an AI model to understand new situations, it will fail unless explicitly trained on that type of situation.

Automation vs Augmentation: Replace or Enhance?

Automation: When AI Replaces Humans

Automation refers to replacing humans with AI systems to perform routine tasks or make standard decisions. A McKinsey study of 800 different occupations found that 60% could have more than 30% of their activities fully automated by AI.

Fields Being Heavily Automated:

Warehouse Logistics: Largely dominated by intelligent robots and AI systems managing operations.

Factory Lines: Increasingly automated, including quality control through cameras and computer vision models detecting defects.

Transcription and Translation: AI models increasingly replacing human workers for speech-to-text and language translation.

Customer Support: Intelligent chatbots replacing human customer service workers.

McKinsey estimates that 15% of the global workforce (roughly 400 million jobs) could be displaced by automation by 2030. Some job categories will significantly shrink while others will see rapid growth, requiring accelerated workforce skill shifts toward digital skills, programming, design, and even building ML models.

The upside: labor productivity growth is expected to rise from roughly 0.5% (2010-2014) to an average of 2% due to increased automation and workforce upskilling.

Augmentation: When AI Enhances Humans

An alternative to automation is augmentation - complementing human intelligence rather than replacing it.

Key Advantages of Human-Computer Collaboration:

Complementary Skill Sets: Humans excel at understanding new situations, context, and causation. AI models can process vast amounts of data and identify fine patterns within large datasets.

Human Control: Humans stay in the driver's seat and can use AI as a tool to complement their intelligence when needed.

The Cyborg Chess Example

In 1996, IBM's Deep Blue defeated chess grandmaster Garry Kasparov. This led to the emergence of "cyborg chess" where human players compete but each can be assisted by a computer.

Studies found that cyborg pairs (human + computer) outperform both the best humans alone AND the best AI systems alone. The complementary nature of human common sense reasoning and context understanding, combined with the computer's data processing abilities, creates superior performance.

Two Forms of AI Augmentation

Triage: AI as First Pass Filter

The idea is using an AI system as a first pass. When the AI can make clear decisions, it doesn't need human review. When there's high uncertainty, it goes to a human expert.

Insurance Underwriting Example: ML models create a first pass to flag low-risk users (clear them for policy) and very high-risk users (don't underwrite). Users in the middle get passed to human underwriting experts.

Radiology Example: Model makes first pass looking for pneumonia in chest X-rays, identifying clear cases of pneumonia or no pneumonia, but referring uncertain cases to human radiologists.

Decision Support: AI as Advisor

The AI model provides recommendations based on data, but the human expert applies common sense reasoning and context understanding to make the final decision.

Investment Example: ML model processes vast amounts of historical stock data, but human investors make final decisions by applying reasoning and understanding of broader market conditions.

Medical Diagnosis Example: AI processes historical data on similar cases and treatment outcomes, providing recommendations to doctors. But doctors engage in dialogue with patients and understand the full symptom picture to make final diagnoses.

Building Trust: If Users Don't Trust It, They Won't Use It

No matter how good your model is, it will not be right 100% of the time. In fact, it's likely to often be wrong. How do you get users to trust it regardless?

Some users who are technology savvy may be eager to try AI systems. But others who are not as tech-savvy may be resistant to using automated models. If users don't trust your model, they won't use it.

Five Strategies to Inspire Model Trust

1. Communicate Performance Metrics

Track and communicate two types of metrics:

Output Metrics: Model performance metrics like mean square error for regression or precision/recall for classification.

Outcome Metrics: Business impact metrics, often in dollars saved/earned or time saved.

Use metrics to inspire trust by: Communicating business outcomes achieved by previous users ("on average, users save X million dollars per year"). Tracking and sharing model metrics specific to the current user's experience.

2. Present Confidence and Uncertainty

Consider a hurricane prediction model. You could output a discrete result like "Level 3 storm" or provide probabilistic output showing probabilities for each level.

If you predict Level 3 and it turns out to be Level 4, users might completely lose trust. But if you show "55% chance Level 3, 24% chance Level 4" and it becomes Level 4, users may lose some trust but won't completely disregard the model since it indicated the Level 4 possibility.

3. Provide Explanations Behind Outputs

When models make incorrect predictions, transparency into how the prediction was reached helps users understand why it differed from reality.

Approaches for providing explanations:

Use interpretable models that are easy to understand (linear regression vs neural network)

Indicate which primary features the model is considering

Provide simplified approximations for complex models like neural networks

Offer counterfactual explanations ("if your credit score was 10 points higher, we would have increased your limit by $2000")

4. Acknowledge Model Limitations

Sometimes the best option is simply acknowledging limitations rather than giving incorrect answers.

Back to our dog breed classifier encountering a cat: Rather than guessing what dog breed it is, have the system respond with "no answer" or indicate that the model isn't trained to classify this picture. Even better if you can provide an alternative solution path.

5. Incorporate Human in the Loop

Have human quality control of model outputs so humans can flag potential issues before they reach users. This is particularly important early in product rollout.

But use this strategy carefully! You shouldn't try to guess when the model is right or wrong.

The Weather Forecasting Cautionary Tale

Using human experts in the loop to help course correct when the model is wrong, but you have to make sure that you don't use this too judiciously.

In an example from the instructor's previous role, an automated weather forecasting system had difficulties with certain extreme situations. They decided to allow human meteorologists to review and edit outputs before sending to clients.

It turned out these meteorologists were making too many edits - not just in edge cases where the model struggled, but in many scenarios where human intervention wasn't needed. The automated system actually outperformed the human-in-the-loop model overall because the humans were over-editing.

Change Management: Getting People to Actually Use AI

AI products almost always require changing somebody's workflow or user behavior since they involve automation. This creates significant disruption to user workflows.

Examples of Workflow Disruption:

Unlocking smartphones with facial recognition instead of manually entering passwords

Automating customer service departments with intelligent chatbots, completely replacing human customer service

When it comes to change management at companies, you may have to overcome additional challenges - we're talking about allaying fears related to job security.

The Change Management Process

Phase 1: Proactive User Education

Early in rollout, proactively communicate why users should adopt the system and what value it will create for them.

Critical: Focus on specific tangible benefits for individual users, not vague aggregate benefits for the organization. As humans, we think "what's in it for me?" Be sure to explain what's in it for them - specifically them, not what's in it for their company.

Phase 2: Deployment and Training

Follow a step-by-step deployment roadmap rather than single-shot deployment to everyone at once.

Start with early adopters in a small group. Gather feedback from early users over time. Begin rolling out to others based on that feedback. Provide training so everyone, even those least comfortable with new technology, can build confidence.

Train the Trainer Approach: The system provider trains early adopters to become trainers, then those early adopters train the rest of the organization.

Phase 3: Monitor Adoption

Build in methods for monitoring usage of your product. Compare adoption rates to initial targets/expectations. If usage is lower than targets, this indicates need for further education or training.

Account for user adoption rate as one of the metrics you need to monitor and flag as an issue if there's low usage.

Key Takeaways: The Human-AI Partnership

Humans excel at understanding and learning from context and applying common sense reasoning to new situations. AI systems have great ability to process vast amounts of data and use patterns within that data to reach decisions.

Because of these complementary skillsets, we can take advantage through augmentation approaches where we're augmenting human decision-makers with AI systems rather than replacing them entirely.

The strategies for performing augmentation include triage systems and decision support systems. But successful AI adoption requires thoughtful change management, proactive user education, and building trust through transparency about both capabilities and limitations.

The future isn't humans vs. machines - it's humans WITH machines, each doing what they do best.