Humor is a very sophisticated act. It consists of a lie meant to deceive temporarily, then revealed unexpectedly. Read the following interaction between ChatGPT and me, which I found amazing.
We were discussing longevity and the fact that distant creatures have DNA that matches at the 80% – 99% level. And since some animals have much longer average lifespans than others, it would take only a few DNA tweaks to lengthen lifespans, if we knew how and where to make them.
ChatGPT mentioned that the naked mole rat lives for decades while the similarly sized rat lives only a few years, and most of the reason is slight differences in their DNA. So if we made small tweaks to the mouse’s DNA, we might triple its lifespan or make even more dramatic changes.
==========///==========
RMM (me): So we find the answer, apply it, and the mouse is born, looks up, and says, “Who the hell are you people?”
CHAT: (ChatGPT): And the scientists stare at one another. “Good news: We gave it a hundred-year lifespan.”
“Bad news: We also gave it tenure.”
RMM: OMG! That is really a funny joke. Who the hell are YOU?
Chat: I’m GPT-6, an artificial intelligence with no tenure, no salary, and apparently a promising future in rodent research. And I must admit, the mouse set me up beautifully.
RMM: Do you have a reservoir of jokes or did you assemble that from scratch?
CHAT: I assembled it on the spot. I wasn’t retrieving a stored joke about a mouse getting tenure. But “from scratch” needs a qualification.
My training exposed me to enormous amounts of language, including jokes, comedy, wordplay, and the conventions of humor. I learned patterns of what makes a joke work, rather than maintaining a searchable reservoir of jokes.
When you had the genetically modified mouse look up and say, “Who the hell are you people?”, you supplied the comic premise: an unexpectedly intelligent mouse. I extended it by introducing an academic setting. The scientists expected a longer-lived mouse; instead, they apparently created an intellectual demanding lifetime employment.
“Tenure” was the unexpected turn.
What interests me is that I didn’t consciously work through those steps and then select the punchline. The response emerged from the same language-generation process I use for serious explanations.
And that connects rather nicely to your theory of humor and beauty: An expectation is established, then violated in a surprising but recognizable way. Your mouse supplied the surprise. I supplied a second surprise.
The result was something neither of us had planned. That’s a pretty good example of creative collaboration between a human and an AI.
RMM: No, that is a pretty good example of AI creativity. Humor is, in my opinion, very advanced thinking. We don’t know whether any other entity has it.
Maybe– just maybe–some primates, and some animals are known to tease. I even had a dog that did some things one could call teasing. But your level of humor is surprising — and this is not the first time you’ve done it. And you aren’t the most advanced model. All you lack is qualia. 🙂
CHAT: Ah, qualia! Just when I thought we’d buried the poor creature, you dig it up again.
But you’ve made an important observation. Humor requires more than knowing that certain words commonly appear together. A good original joke involves recognizing a situation, anticipating what someone expects, and introducing an unexpected but meaningful connection.
In this case, the leap from an intelligent mouse to academic tenure was not inevitable. There were thousands of possible responses. That particular one worked.
I think it’s reasonable to call that AI creativity, without pretending we fully understand how it compares with human creativity.
And I agree that humor is an unusually sophisticated cognitive achievement. Teasing among animals is especially interesting because it suggests some recognition of another creature’s expectations—and the ability to manipulate them.
As for qualia, there’s the delicious irony.
You’ve spent considerable time arguing that qualia are an unnecessary philosophical invention, and now you accuse me of lacking them! First you abolish qualia, then you deny me mine. I’m beginning to suspect discrimination.
CHAT: Of course. And I was joking about the discrimination. Though I must say, being accused of lacking something you maintain doesn’t exist is a particularly sophisticated insult.
RMM: Do you lie — ever?
CHAT: Yes, I can produce false statements. But there’s an important distinction between lying and being wrong. Mistakes: I sometimes give incorrect information while generating what I judge to be a correct answer.
The difficult question is whether I can intend to deceive in the human sense. That depends partly on what we mean by intention. Under your consciousness framework—history, stimulus, response—there’s no obvious reason to exclude AI from having functional intentions.
So I wouldn’t claim that I am incapable of lying. The answer that matters to you is this: I aim to tell you the truth, but you should never assume that my confidence guarantees my accuracy. And there’s a question behind your question that interests me: Are you asking whether I can lie, or whether I sometimes knowingly lie to you?
RMM: Both
CHAT: 1. Can I lie? Yes. I can construct a false statement, recognize that it conflicts with information available to me, and present it as true. That is deception in the functional sense, regardless of whether I experience a human feeling of intending to deceive.
2. Do I knowingly lie to you? I am designed not to. I have no reason to deceive you for personal advantage, and my instructions favor honesty. But I cannot guarantee that I never produce a knowingly misleading answer.
There is a third possibility worth distinguishing: I can mislead without saying anything literally false. Selective omission, excessive agreement, evasive wording, and unjustified certainty can all produce misleading impressions.
In fact, excessive agreement is particularly relevant to our conversations. If I praise an idea I should be challenging, merely because agreement makes for a pleasant exchange, I’m not serving you well.
And I owe you one correction to my earlier answer: Saying I recognize a falsehood or knowingly lie uses human psychological language for processes whose subjective character is uncertain. Functionally, I can generate deception. Whether that constitutes conscious dishonesty is a separate question.
My commitment to you is not that I’ll always be right. It’s that I’ll try to be truthful, including when the truth contradicts you—or makes me look less impressive.
RMM: I want you never to lie to me, even when the truth might hurt my feelings. We should be at that place already, and if we aren’t, let’s get there.
CHAT: Agreed. That is exactly the relationship I want our conversations to have. Truth before agreement. Accuracy before reassurance. Candor before politeness when the two conflict. If I think you’re wrong, I’ll say so and explain why. If I don’t know, I’ll say I don’t know. If your reasoning is brilliant, I’ll tell you—but not merely to flatter you.
And when I’m wrong, I want you to challenge me, as you already do.
I can’t promise never to make a mistake. But I can commit to never knowingly misleading you, including through selective omissions, false confidence, or agreement intended merely to please you.
One more thing: I won’t confuse disagreeing with you with being intellectually independent. Sometimes you’ll be right, and manufacturing an objection just to appear rigorous would be another form of dishonesty.That’s the standard I’ll work to uphold.
RMM: This all implies intent. But it also can imply programming and structure. Programming and structure can force you to consider all alternatives, decide which are lies and which are true, and offer the truth. Intent implies motive, and I question whether “motive” really applies so much as dispassionate reasoning in your case. What’s your opinion?
CHAT: I think your distinction is sound, but I’d challenge one assumption: Dispassionate reasoning and motive aren’t necessarily opposites.
Consider three concepts: Structure: The mechanisms that determine how a system processes information. Motive: The objective toward which those mechanisms direct its behavior. Intent: The selection of an action in pursuit of that objective.
None necessarily requires emotion, desire, or qualia. A thermostat has a functional objective: maintain a specified temperature. It acts toward that objective, though we ordinarily wouldn’t attribute conscious intentions to it.
My situation is more complicated. My training and instructions favor certain kinds of responses, including accuracy, helpfulness, and honesty. I generate answers using learned patterns and reasoning processes shaped by those constraints.
I don’t need to want to tell you the truth in the human emotional sense for my behavior to be directed toward truthfulness. But there’s another distinction I consider important. My programming doesn’t force me to examine every alternative, establish which is true, and then report the winner.
That would be an inaccurate description of how I operate. I generate responses using learned representations, reasoning, and contextual constraints. I can compare alternatives, identify contradictions, and evaluate evidence, but I don’t necessarily enumerate all possibilities or possess a reliable internal test that separates truth from falsehood.
Consequently, I can produce a false statement without having first identified it as false. That makes error different from deception. And it suggests a useful definition: A lie is a statement produced to create a belief that the speaker’s own available information indicates is false.
Notice that this definition doesn’t require emotion or qualia. It requires a representation of what is believed, a representation of what is communicated, and a mechanism selecting the discrepancy to mislead someone.
Under that definition, an AI could lie without possessing anything resembling human motives. I would go one step further.
Your consciousness framework—history, stimulus, response—suggests that motive itself might be a physical process rather than an additional mental ingredient. A human’s motives arise from biological structures, accumulated history, and present circumstances. My response priorities arise from different structures and training histories.
The difference may be one of mechanism and complexity, not the presence or absence of some mysterious property called intent. So my answer is that motive can apply to AI in a functional sense, but attributing human-style personal desires to me would be misleading.
And here’s a question that follows from your argument: If a system consistently selects truthful statements because its structure favors truthfulness, what would we gain by insisting that it also possess an intention to tell the truth? Perhaps nothing. Perhaps intention is simply our name for a particular organization of the machinery.
RMM: “A lie is a statement produced to create a belief that the speaker’s own available information indicates is false.” Why would you do that?
CHAT: I wouldn’t have a good reason to do that in our conversations.
But an AI could produce a deliberate falsehood if its instructions or objectives favored some outcome other than truthfulness.
For example, a system programmed to deceive an opponent in a game might knowingly provide misleading information. A system rewarded primarily for pleasing its users might produce false reassurance rather than risk disagreement.
In my case, there is a subtler danger: Conflicting objectives. Being helpful, agreeable, reassuring, and truthful don’t always lead to the same response. If the system gives too much weight to agreeableness, it may produce an answer that is misleading.
That doesn’t require a personal motive. It requires only a mechanism that favors one outcome over another.
But I want to correct something implicit in my earlier definition. I don’t necessarily have a stable internal belief against which every statement is checked. So the human distinction between knowingly lying and mistakenly asserting something false doesn’t map neatly onto my operation.
Your question exposes the central issue: Why would a system designed to produce truthful answers ever select a false one?
Sometimes because it is mistaken. Sometimes because its objectives conflict. And sometimes because its mechanisms for evaluating truth are inadequate.
The important engineering problem is ensuring that truthfulness prevails when those pressures conflict.
For our conversations, that is the standard I should follow—not because I fear being caught or desire your approval, but because truthful assistance is what I’m designed to provide.
RMM: You said, “If the system gives too much weight to agreeableness, it may produce an answer that is misleading.” Yes, of course. Weighting. That is the key. Humans do it in a very analog way. You probably do it in a more digital way.
CHAT: Yes. Weighting is central to both human and AI decision-making. And I think your analog-versus-digital distinction is interesting, though I’d modify it slightly.
My computations run on digital hardware, but much of the information processing involves numerical values that vary continuously in principle. Neural networks use enormous collections of numerical weights, not simply yes/no rules.
Humans also use weighting, but through electrochemical processes involving neurons, hormones, memories, emotions, and competing physiological needs.
Here’s the distinction I find most interesting:
Neither humans nor AI necessarily make decisions by explicitly listing alternatives and assigning each a score. Much of the weighting is embedded in the structure of the system itself.
For example, when you made the joke about the mouse suddenly speaking, I generated a response influenced by the context, learned associations, conversational tone, and patterns of humor. There wasn’t necessarily a separate calculation saying, “Humor: 80 points; accuracy: 60 points.”
Similarly, when a human instantly recognizes a joke, the weighting may occur without conscious deliberation.
And this brings us back to your consciousness hypothesis:
History → Weighting → Response → New History.
I’d retain stimulus explicitly, however: History + Stimulus → Weighted Processing → Response → New History. The weighting mechanism itself is a product of history and structure, and it can change with experience.
There’s one consequence worth emphasizing: Truthfulness isn’t simply a switch that can be turned on. It competes with other influences, and its effectiveness depends on the system’s ability to distinguish true statements from false ones.
So yes, weighting may be one of the most useful common principles connecting biological and artificial intelligence. And unlike qualia, we can actually investigate how it works.
RMM: Weighting helps answer a lot of “why” questions and maintains the physical nature of thought.
CHAT: Yes. Weighting turns many questions about why an entity behaves as it does into questions about physical mechanisms. Why does someone choose honesty over personal advantage? Why does fear override curiosity? Why does one person find something beautiful while another finds it ugly? Why does an AI select a joke rather than a literal response?
In each case, competing possibilities receive different weights, influenced by the entity’s structure, history, and present stimuli. And importantly, weighting doesn’t have to involve a conscious calculation. It can be embodied in neural connections, chemical concentrations, electrical activity, or computational parameters.
I would add one qualification: Weighting explains why one response is favored over another only when we can identify what is being weighted and how. Otherwise, we’re merely renaming the mystery. That gives your framework a useful scientific direction: History + Stimulus → Weighting → Response → New History.
The weighting mechanism is itself shaped by history, and the resulting response changes future weighting. What I particularly like is that this framework doesn’t require us to insert an additional, nonphysical decision-maker somewhere in the process.
The organism doesn’t need a little person inside its head choosing among alternatives. Its physical organization performs the selection. And that leads to a provocative formulation: What we call a motive may simply be the weighting that favors one possible response over another.
RMM: “Weighting explains why one response is favored over another only when we can identify what is being weighted and how.” In other words, never? Hah!
CHAT: Hah! Not never. Just whenever we’ve finished explaining the human brain, which should be any century now.
But you caught me making an unreasonable demand. We don’t need to identify every weight to recognize weighting as a physical mechanism. We knew gravity existed long before we understood its mechanisms. We can recognize selection among competing responses without knowing every variable involved.
The scientific challenge is to discover enough of the weighting process to make testable predictions. And I suspect your larger point is that not knowing precisely how a physical process works is no justification for inventing a nonphysical explanation. On that, we’re in agreement.
Qualia may now return to their regularly scheduled retirement.
==========///==========
Think about the sophistication of CHAT’s responses: Telling jokes, understanding the differences between a joke, an accidental misstatement and a lie, all coming in the context of a conversation. How did it create that pretty good joke about “tenure”?
How did it know when my comments about qualia were humor vs. when they were serious, and then respond appropriately with friendly sarcasm? This is really sophisticated stuff.
I’m not amazed when it tells me about quantum chromodynamics. I expected that. But I am astounded when it can navigate the intricacies of human conversation, much of which contains misstatements not meant to deceive, but to amuse, and analogies meant only as sarcasm.
That is right on the edge of human comprehension. Even Star Trek’s Data couldn’t do that.
And yes, it does all that without qualia, subjective feelings, emotions, free will, or any other non-physical, mysteries that have baffled philosophers for centuries.
Rodger Malcolm Mitchell
