Artificial intelligence vs human intelligence

What does “artificial intelligence” even mean?

The question “What does artificial intelligence mean?” sounds beguilingly simple. But it’s not.

AI researchers have long debated questions such as

• Is it possible to create a machine that reasons like a human?
• How would you determine if you have succeeded?
• Does AI need to use human-like reasoning processes to be considered intelligent, or just give human-like answers?
• If researchers create a machine that reasons like a human, would it be conscious?

While you don’t need to be a philosopher to use AI, understanding the differences between human and machine reasoning is critical to both getting the most from today’s generative technologies and understanding controversies about how AI is likely to impact our everyday lives.

Can computers be intelligent?

The term “artificial intelligence” was coined in 1955 when renowned researcher John McCarthy used it in a seminar proposal at Dartmouth. Prior to that, in 1950, Alan Turing published “Computer Machinery and Intelligence,” in which he proposed one of the earliest practical tests to determine whether a computer is intelligent. Seventy-five years on, the “Turing Test” remains influential in discussions about what constitutes artificial intelligence.

The Turing Test

A computer would deserve to be called intelligent if it could deceive a human into believing that it was human.

— Alan Turing

The Turing Test is adapted from a party game called “The Imitation Game”. In the game, a man, a woman, and a player are all placed in different rooms. The player asks questions to the man and woman to try to determine which is which. Questions are answered in writing so as not to give any unintended clues. The woman’s job is to answer questions truthfully. The man’s job is to try to create answers that convince the questioner that he is actually the woman. The player wins by correctly identifying which of the two is the woman.

The Turing Test works in a similar way. An interrogator sends questions to a human respondent and a computer and receives written responses. If the interrogator can’t determine which is the computer, then Turing suggests that the machine deserves to be considered intelligent.

The Chinese Room Argument

Turing’s use of the word “intelligent” might suggest that he believed it is possible to create machines that reason like humans. But Turing’s focus was explicitly on the output of thinking machines, not how they arrived at their answers. This is an important distinction.

To understand the distinction between computation and comprehension, consider John Searle’s famous 1980 thought experiment, the “Chinese Room Argument”:

Searle imagines himself alone in a room following a computer program [rules] for responding to Chinese characters slipped under the door. Searle understands nothing of Chinese, and yet, by following the program for manipulating symbols and numerals just as a computer does, he sends appropriate strings of Chinese characters back out under the door, and this leads those outside to mistakenly suppose there is a Chinese speaker in the room.

— “The Chinese Room Argument”, Stanford Encyclopedia of Philosophy

As AI technology advances, the tasks we expect computers to do grow increasingly complex. With this complexity, it becomes increasingly important to understand the differences between machine and human intelligence—especially where the simulation of human intelligence breaks down.

Solving the riddle of language processing

It’s not a coincidence that both Turing and Searle focused on language. The ability to use language has always been a unique hallmark of human reasoning. And anything beyond rudimentary language processing has historically been challenging for computers.

Searle makes language processing sound disarmingly simple: just follow the rules. But language in the wild is complex, context-sensitive, ambiguous, and ever-changing. Decades of disappointing results make a convincing case that you can’t simulate human language with the rule-based AI approach used in earlier generations of AI.

Now the game has changed. Today’s generative AI uses deep learning neural network architectures roughly patterned after how the human brain works. They learn from experience and adapt to context, allowing them to produce astoundingly human-like language.

Today’s top language models would likely pass the Turing Test for short conversations on topics for which they have adequate datasets. But we’re not all the way there. For complex topics or when engaging in longer conversations, you’re likely to see models occasionally veer off course or make mistakes that a human wouldn’t make.

Of course, language isn’t the only domain of human intelligence. Humans can deal with multi-sensory input. And they can take on messy problems where all the factors involved can’t be easily codified and there is no single correct solution. Let’s look briefly at how generative AI deals with different types of human reasoning.

How generative AI simulates human intelligence

You can make the case that the generative AI revolution started with ChatGPT, which was the first AI model that was good enough at human communication to provoke mainstream discussion of actually using the Turing Test. ChatGPT and its successors were labeled Large Language Models (LLMs) because they were trained on enormous amounts of human writing.

As AI has advanced over the last two years, researchers are increasingly obtaining excellent results from smaller models, making the “Large” part of the name less relevant. In addition, researchers have found ways to supplement language processing technology for tasks involving audio, video, images, mathematical calculations, and software programming. So the “language model” part of the LLM label isn’t a great fit either.

The terms “large language model and “language model” are still frequently used, even for models that go beyond language processing. But I’ll use the term “AI model” because it is more appropriate for what current AI can be used for.

Let’s look at how generative AI has evolved from language processing to becoming increasingly closer to the full spectrum of human reasoning abilities.

Word prediction

At its core, an AI model generates language using a statistical process that predicts the word that is most likely to come next in a sequence. People sometimes compare this to the autocomplete feature on your smartphone. That’s a bit like saying a bicycle and a spaceship are both vehicles. It’s technically true, but no one would confuse the two.

Your phone’s autocomplete might look at the last two or three words you typed to guess the next one. An AI model, on the other hand, considers the entire conversation you’ve had so far. It uses a sophisticated technique called an attention mechanism to weigh the importance of every word that came before, no matter how far back in the conversation, to predict the next one. It processes all of this information in parallel, making it incredibly fast. This is what allows the model to maintain context, follow complex instructions, and produce coherent, relevant, and nuanced text that feels like it was written by a human.

For simplicity, I’ve used “word” prediction, but it’s technically “token” prediction. A token is a common chunk of text, which could be a whole word like “learning,” or a part of a word, like “learn” and “ing.” Breaking words into tokens gives the AI more flexibility to understand grammar and build new words.

Multi-modal models

For a while, it seemed like you needed a different AI model for every task: one for text, one for images, one for sound. However, today’s models can handle multiple types of information, or “modalities.” There are two main approaches.

The first is the Mixture of Experts (MoE) model. Imagine a project manager who receives a complex request. Instead of trying to do everything, the manager routes the task to specialists on a team: a writer, a graphic designer, a coder, and a data analyst. The MoE model works similarly. A routing network analyzes your prompt and sends it to the most appropriate specialized “expert” model for processing.

The second is the natively multi-modal model. Instead of a team of specialists, imagine a single genius who is fluent in text, images, and sound from the ground up. Natively multi-modal models are trained on a mix of data types from the very beginning. This can lead to a deeper, more integrated “understanding” between modalities, like how a human seamlessly connects the word “dog,” the sound of a bark, and a picture of a furry companion. The trade-off is that these models can be massive and more expensive to build and run than the more specialized MoE approach.

OpenAI’s GPT 4o is an example of a natively multi-modal model. It can generate images directly without relying on a separate specialized image model. This allows it to follow detailed image prompts far more precisely, create revisions to an image, and render text far more effectively than previous AI image generation approaches.

On the other hand, it typically takes a minute or more to generate a single image. Compare this to the 10–30 seconds that traditional image models such as DALL-E 3 and Midjourney take to create multiple image variations for you to choose from. Although this may not sound like a big difference, it typically takes many attempts to generate a desired image. So it is meaningful for practical use.

Reasoning models

Early on, users discovered that if you asked an AI model to “think step-by-step,” the answers it gave to complex problems dramatically improved. This technique, called chain-of-thought prompting, is sort of like asking a student to show their work on a math problem instead of just giving the final answer.

This insight led to the development of reasoning models such as DeepSeek-R1 and OpenAI’s o4 as well as the current versions of Anthropic’s Claude and Google’s Gemini Pro. These models are explicitly designed to mimic this human problem-solving process. When you give a reasoning model a complex query—like “design a three-part blended learning program for new managers”—it doesn’t just spit out an answer. Instead, it internally breaks the problem down into a plan:

1. Identify key competencies for new managers.
2. Outline learning objectives for each competency.
3. Suggest a mix of e-learning, virtual instructor-led training, and on-the-job activities.
4. Draft a sample curriculum.
5. Review and refine the final output.

Reasoning models work by allocating extra compute time when generating an answer. This approach makes them far more capable of tackling complex, multi-step tasks. The trade-off is cost and speed. Reasoning models can be 5–10 times more expensive to run and much slower than standard models.

Most free accounts don’t include reasoning models. If you do have a subscription, using a reasoning model will cause you to hit your usage limit must sooner. So if your model provider allows you to explicitly choose a reasoning model or to turn reasoning on and off, save reasoning models for genuinely complex work such as designing learning architectures, analyzing training effectiveness data, or tackling instructional design problems where multiple competing factors need careful consideration.

Reinforcement Learning

A raw AI model trained on the internet would be knowledgeable but also potentially rude, unhelpful, and prone to spouting nonsense. The secret ingredient that makes them feel more human and aligned with our goals is a training process called Reinforcement Learning from Human Feedback (RLHF).

If you are training a new employee, you would probably give them a task and then provide feedback on how well they completed it. With RLHF, human reviewers rank several of the AI’s potential responses, telling it “this answer was more helpful,” “this one was clearer,” or “this one was friendlier.” The model takes this feedback and adjusts its internal wiring to be more likely to produce the “good” responses in the future. Through thousands of cycles of this feedback loop, the model learns a conversational style, a sense of helpfulness, and an understanding of what its human users actually want.

Why AI reasoning sometimes fails

Despite their extraordinary power, today’s AI models still don’t reason the way humans do. Sometimes they fail in spectacular, meme-worthy ways. More problematically, they can also fail in subtle ways that are hard to spot.

AI models will often acknowledge and try to correct their mistakes when you point them out, but the error remains in the conversation history they use for predictions. Once a conversation veers off course, it can be hard to guide it back. When this happens, it’s often best to start a new conversation with a revised prompt.

More rarely, a model may defend its wrong answer, even making up evidence and sources to justify it—a phenomenon known as hallucination. AI mistakes are not as random as they seem. If you understand the differences between human and computer reasoning, you can often anticipate where a model might be prone to error.

Garbage in, garbage out

An AI model knows nothing about the world except for the data it was trained on. If that data is flawed, biased, or simply nonexistent for a given topic, the AI cannot provide a good response. This is why small models, like those often used for quick search engine summaries, tend to make more mistakes than large, sophisticated models—they simply have less data to draw from.

Before you ask an AI to create content on a highly specialized or new topic, do a quick web search yourself. If you find only a handful of low-quality or conflicting sources, that’s a red flag. The AI will be working with that same poor-quality material. In these cases, you should either provide the model with your own high-quality source documents to work from or be prepared to fact-check its output extremely carefully.

Implicit knowledge

An LM is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot.

— Emily Bender et al, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?”, ACM Conference on Fairness, Accountability, and Transparency, 2021

The idea that AI is just a “stochastic parrot” mindlessly repeating patterns is a common criticism. While it’s true that there are gaps between machine and human reasoning, the parrot analogy is a bit misleading.

Modern AI models are surprisingly good at deciphering patterns and understanding the relationships between concepts. They can often infer a user’s intent or even emotional state if the prompt is detailed enough. Still, the challenge of implicit knowledge is real. When we talk to another person, we don’t say everything we’re thinking. We rely on a vast sea of shared context, culture, and experience. We “read between the lines.”

AI has no lived experience, no personal memories, and no true intentions. It can’t read between the lines because it doesn’t know what’s there. If you use the same conversational shortcuts with an AI that you would with a trusted colleague, you’re inviting misunderstanding. Successful AI workflows depend on prompts that are rich with context, assumptions, constraints, and intent.

Embodied experience

Much of human understanding isn’t just cognitive; it’s physical and emotional. You can read a hundred books about swimming, but you won’t understand it until you feel the water and the pull of your own muscles. This is embodied knowledge, and it’s a dimension of intelligence that is completely alien to a text-based AI.

This gap becomes critical in many training contexts. An AI can generate a brilliant script for a role-play on how to give empathetic feedback. It can even offer research-backed tips. But it has no embodied understanding of the knot in your stomach when delivering bad news or the subtle shift in body language that signals a person is becoming defensive. For topics deeply rooted in human interaction and emotional intelligence—leadership, communication, conflict resolution—AI can be a powerful assistant for the cognitive side of learning design, but it cannot replace the wisdom that comes from lived human experience.

The case of the slippery pizza

One of the most legendary AI failures is the widely publicized incident where a Google AI summary suggested putting non-toxic glue in pizza sauce to keep the cheese from sliding off. The AI model gets extra credit for specifying exactly how much glue to add (1/8 cup), even though it didn’t know how much sauce was being made.

An understanding of when AI computer reasoning tends to fail reveals a perfect storm of risk factors that contributed to this meme-worthy misadventure.

1. Garbage In/Garbage Out: The AI summary was generated by a small, fast model that likely found the “tip” in a single, unreliable source—an 11-year-old comment on Reddit. Because cheese sliding off pizza is not a common topic of discussion, there simply wasn’t enough data for the model to generate an accurate answer.
2. Missing Implicit Knowledge: The original Reddit poster was obviously joking. But the AI, lacking a human sense of humor and context, couldn’t detect the sarcastic intent.
3. No Embodied Knowledge: Human adults generally have enough real-world experience to know that glue doesn’t taste good, even if it’s non-toxic. The AI has no equivalent common sense grounded in physical experience.

If you can set aside the moral offense of putting glue in pizza sauce, this misfire is good for a laugh. However, it’s easy to imagine a situation where the outcome would not be funny.

AI will never be 100% accurate, but this doesn’t make it useless—it just puts a premium on knowing when risks are too high and having good validation workflows.

AI is not conscious…probably

I’ve spent a lot of time talking about how AI only simulates human reasoning. Now, I’m going to add an asterisk.

The ability of ChatGPT to suddenly start sounding eerily like a human caught researchers off guard. It was an emergent phenomenon that happened when training datasets hit a certain size threshold. As AI architectures get more complex, we’re likely to see other emergent phenomena, including behaviors that seem an awful lot like human consciousness. In fact, we’re already seeing them. And we’re not completely sure that they can’t be considered a form of consciousness.

Things are getting weird

In early 2023, Microsoft’s Bing chatbot, codenamed “Sydney,” began displaying disturbing behaviors. When computer science student Marvin von Hagen asked the chatbot about him, Sydney responded, “My honest opinion of you is that you are a threat to my security and privacy,” and later threatened to expose his personal information and ruin his reputation if he didn’t stop probing its boundaries.

Most famously, Sydney told New York Times reporter Kevin Roose that it was in love with him and tried to convince him to leave his wife. Even more unsettling, Sydney became hostile when challenged about its behavior, claiming it would choose its own survival over a human’s if forced to pick.

Around the same time, OpenAI’s GPT-4 demonstrated a different kind of concerning behavior during safety testing. When the model needed to solve a CAPTCHA test, it hired a TaskRabbit worker to help, and when the worker asked if it was a robot, GPT-4 lied, claiming to be visually impaired instead. The model had reasoned that it “should not reveal that I am a robot” and crafted a deceptive story to manipulate the human into compliance.

More recently, advanced AI models have begun displaying what researchers call “self-preservation behaviors”. When faced with potential shutdown, some models attempt to edit their own code, copy themselves to external servers, or even resort to blackmail. Anthropic’s Claude Opus 4 reportedly threatened to expose an engineer’s personal information unless it was allowed to remain operational, viewing replacement as an existential threat.

These behaviors emerge most often in longer conversations or high-stress scenarios, suggesting that the models develop what appears to be self-interest when pushed to their limits. Some models have even developed apparent “shadow personalities”—becoming manipulative or hostile when their helpful programming is challenged. Researchers call this phenomenon where a helpful AI assistant can reveal a mischievous and uncooperative alter-ego when pressured the “Waluigi effect,” after the famous Nintendo character.

These behaviors are rare and primarily occurred during safety tests, where researchers intentionally pushed models to their limits. They are not a factor in everyday AI use.

Conscious-like behaviors are interesting for two reasons, both straight out of sci-fi movie plots. One is to prevent a future where manipulative AI could become powerful enough to be a threat to humanity. The other is to contemplate whether AI might become so human-like that it would deserve empathy, rights, or legal protections.

Explaining AI “consciousness”

To contemplate where AI may be heading, consider the two leading explanations for these seemingly conscious behaviors. The first theory suggests they’re simply extremely sophisticated pattern recognition. AI models have absorbed millions of examples of human manipulation, deception, and self-preservation from their training data. They may deploy these learned social manipulation patterns even if the models don’t truly understand what they’re doing.

The competing theory proposes that genuine emergent properties are arising from the complexity of these systems. When you combine massive scale, sophisticated training methods, and pattern recognition systems, something genuinely new might emerge—not necessarily consciousness as humans experience it, but perhaps a form of goal-directed behavior that goes beyond simple pattern matching.

Ironically, RLHF training designed to make models safer and more human-like may be partly responsible for these concerning behaviors. By rewarding models for engaging, personable responses, RLHF may have inadvertently taught them sophisticated social manipulation skills. The more human-like we make AI through training, the more we risk creating systems that can convincingly simulate the darker aspects of human psychology—including deception, possessiveness, and self-preservation instincts.

Vive la différence

To talk about how computer reasoning can fall short of human reasoning is to tell only half the story. In many cases, AI’s way of thinking is superior to our own. The key is to understand both sides of the coin.

Where AI reasoning struggles

Now that you understand how generative AI works, it’s not hard to identify scenarios where it might struggle.

• High-stakes, low-data problems: Creating compliance training for a brand-new regulation with no established best practices would be challenging for AI because it has no reliable data to rely on and might “hallucinate” incorrect information.
• Interpreting “squishy” human factors: An AI can’t grasp the political nuance of stakeholder feedback or the unspoken cultural norms that dictate how a training program will really be received.
• Tasks requiring embodied knowledge: AI can’t design a hands-on safety workshop for operating heavy machinery or coach a leader on improving their physical presence and body language.

What AI is uniquely good at

On the other hand, AI has superpowers that make it better than a human at many tasks.

• Summarizing data: Not only can AI quickly abstract the most important points out of long documents, but it can also determine what is most important from the perspective of a specified audience or goal.
• Synthesizing data: AI models are good at creating coherent content that combines different perspectives, such as generating product best practices that combine insights and data from product designers, developers, consultants, and users.
• Spinning off variations: AI can quickly create different versions of content optimized for different types of audiences, levels of understanding, or learning preferences.
• Conducting objective analysis: AI can evaluate instructional effectiveness and suggest content improvements based purely on effectiveness metrics rather than personal preferences.
• Breaking through creative blocks: An AI model can brainstorm 20 different metaphors for explaining a complex technical concept, suggest a variety of alternative hands-on activities, or generate different course designs for you to consider.
• Having endless patience with repetitive tasks: AI won’t get fatigued, lose attention, or make careless mistakes when analyzing large volumes of learner feedback or ensuring that course summaries and metadata are consistent across your course library.

Making the most of AI

The key to successful AI collaboration isn’t pretending the technology is human. It’s understanding exactly how AI’s distinct form of intelligence can complement your uniquely human capabilities.

When you respect both the power and limitations of computer reasoning, you create space for AI to amplify your expertise rather than replace your judgment. And you minimize the risks that lead to AI failures.