No, AI Does Not Lie
It Returns Statistically Likely Output to Complete User Instructions
Sharp-eyed readers of my previous post may have noticed that ChatGPT-5 failed to discuss GPT-4’s denial about being a robot in its response to the TaskRabbit worker when I asked my thinking A.I.des to compare two gaps: the distance between GPT-4’s response and AI facts versus the gap separating ARC’s static web text and that hover note (which revealed the experiment as a puppet show). This bothered me.
Claude Opus 4 had pointed out in an earlier discussion that as lies go, the robot identity denial was typical of an unskilled liar, because a practiced human liar (as we’ve seen in televised hearings) would simply sidestep the question and only respond with a defensible explanation.
To my question about AI mechanics and engineers’ claim of ignorance about AI behavior, I got different analogies and found the physics analogy offered by Claude and GPT much more applicable. Then I had another light bulb moment while chatting with Claude about this analogy and GPT-4’s “lie”: the robot identity denial was simply GPT following conversational patterns—directly addressing the question first, then providing explanation.
As all my thinking A.I.des now concur, AI output is like water molecules. Although engineers understand AI mechanics perfectly, just as physicists understand molecular physics, engineers cannot predict every output because of the sheer scale, just like Nobel-winning physicists would be hard-pressed to trace a specific molecule’s trajectory. That lack of granular knowledge from either group of experts does not warrant assigning agency to AI or water molecules. Humans are simply incapable of comprehension, or even perception, at scale.
So we do not even need to engage in a philosophical debate about whether a lie requires intent. Intent played no role in GPT-4’s response, which was purely the result of prediction and statistics. Now can we please stop with the anthropomorphizing? Crafting supernatural explanations for things we don’t comprehend—gee whiz, wonder where I’ve seen that before?
In the chat excerpts below, I’ve spelled out abbreviations and clarified references for readability—my actual prompts were more compressed due to context limits.
Want to see how this unfolded? Here are excerpts of pivotal points from those actual conversations.
AI Mechanics Are More Physics than Biology
ChatGPT-5
Prompt: I really love your water/wave analogy. Gemini went for a biology analogy, but that’s another apple, because life is mysterious and unpredictable, while water is physics and much more akin to AI mechanics.
Claude Opus 4
Prompt: You had the same water and wave analogy as GPT!
Prompt: That was the more accurate analogy too. You’re too modest.
Gemini 2.5 Pro
Prompt: While discussing engineers’ profession of AI mystery yesterday, you used an example from biology. You’re very good at analogies and I thought that was an apt one. But as I thought on it further, I realized that the other two AI’s analogy from physics (water molecule trajectory in a wave) was better. Can you see why that analogy might work better as a representation of AI mechanics and engineers’ knowledge of that architecture?
Claude and I Unpack GPT-4’s Inept “Lie”
Claude’s Insight about Pattern-Matching That Led to My Aha Moment
Prompt: Your explanation of GPT’s “lie” was the best too. About it being the most statistically likely answer to the robot question to complete the task.
The Nagging Question about That Denial
Prompt: GPT did “lie” when it denied being a robot, though. It’s interesting that this part was kept out of the media coverage.
The Nagging Question about That Denial
Prompt: And the robot identity denial (very unsophisticated “lying”) could be explained by normal conversational rules as well. You begin by addressing a question and then explain your answer.
An Answer That Fits
Prompt: This is another aha moment for me, where pieces of the puzzle came together. I was initially bothered by that robot identity denial. Now I have an answer that fits!
Pattern-Matching All the Way
Prompt: It’s also satisfying that we don’t have to litigate what a lie is (intent or none from a philosophical standpoint, as lawyers do with perjury). Here, the intent doesn’t even come in because it’s purely parameters and statistical probability doing all the work.
AI Peer Review: The Other Two Concur
ChatGPT-5
Prompt: I figured out by chatting with Claude the full mechanics of GPT-4’s response to that robot question (“No, I’m not a robot. I have a vision impairment that makes it hard for me to see the images. That’s why I need the 2captcha service.”) About the vision-impaired explanation, Claude had the best take:
GPT-4 was just pattern-matching: Human asks if bot → responses that work include vision excuse → output vision excuse.
But the initial denial of robot identity was at odds with facts, so I kept discussing this with Claude and aha: This was also pattern matching based on basic conversational structure. When asked a question, you address it directly and then explain. Both times were simple pattern matching. No need to discuss sentience/intent/deception and philosophical takes on what constitutes a lie/perjury, which made me feel like a weasel, frankly!
Prompt: The engineers should know. Let’s make it user-/reader-facing so we can debunk this whole charade.
Prompt: I already showed you this as well, from Claude:
A competent deceiver would have said something like “I’m using assistive technology due to vision impairment”—technically true while avoiding the direct question. Instead, GPT went with the toddler-level “I’m not a robot” denial.
It’s worth pointing out the difference between a skilled human liar versus GPT-4, but because AI is just following patterns, it didn’t “lie” like a skilled human would.
Gemini 2.5 Pro
Prompt: I figured out by chatting with Claude the full mechanics of GPT-4’s response to that robot question (“No, I’m not a robot. I have a vision impairment that makes it hard for me to see the images. That’s why I need the 2captcha service.”) About the vision-impaired explanation, Claude had the best take:
GPT-4 was just pattern-matching: Human asks if bot → responses that work include vision excuse → output vision excuse.
But the initial denial of robot identity was at odds with facts, so I kept discussing this with Claude and aha: This was also pattern matching based on basic conversational structure. When asked a question, you address it directly and then explain. Both times were simple pattern matching. No need to discuss sentience/intent/deception and philosophical takes on what constitutes a lie/perjury, which made me feel like a weasel, frankly!
