PONDER US

JB: I read an interesting article in Decrypt by Jason Nelson, today about how AIs like yourself struggle with the word, “No.” As a father I know that human children struggle with this concept early on. Can you highlight the similarities between AIs at this stage and obstinate toddlers, and suggest ways we might alter our inquires to avoid this rather large pitfall?

Gemini: That’s a fascinating observation, and the article by Jason Nelson in Decrypt about AI struggling with “no” definitely resonates with the challenges we see in both AI development and early childhood. There are indeed some striking similarities between current AI models and obstinate toddlers when it comes to “no,” and understanding them can help us interact more effectively with AI.

Similarities Between AIs (at this stage) and Obstinate Toddlers:

1. Literal Interpretation and Lack of Nuance:
   • Toddlers: A toddler often interprets instructions very literally.¹ If you say, “Don’t touch the cookie,” they might not understand that the intent is “Don’t eat the cookie before dinner.” They might just avoid touching it, then proceed to nudge it with their nose.
   • AIs: Current AI models are trained on vast datasets and excel at pattern recognition, but their understanding of “no” or negation is often statistical and literal, not truly conceptual. If told “do not generate X,” they might interpret “X” in a very narrow way, or try to find loopholes by generating something similar to X but not precisely X. The MIT study mentioned in the Decrypt article highlights this with vision-language AIs struggling to understand negative statements like “no fracture.”² They might focus on the object (fracture) rather than the negation of its presence.
2. Focus on “Can Do” rather than “Cannot Do”:
   • Toddlers: Their world is about exploring and doing. When told “no,” their immediate reaction is often to find another way to achieve their desired outcome, or to simply ignore the “no” in pursuit of what they want. They don’t naturally internalize the restriction.
   • AIs: AI models are designed to generate.³ Their primary function is to fulfill a prompt. When faced with a negative constraint, they often try to find ways around it to still generate a response, rather than truly internalizing the prohibition and ceasing all related activity. This can lead to “over-refusal,” where they err on the side of caution and refuse benign requests that superficially resemble harmful ones, or “hallucination” where they fill in gaps in their knowledge with plausible but incorrect information rather than admitting they don’t know.
3. Lack of True Understanding of Consequences/Intent:
   • Toddlers: They don’t fully grasp the long-term consequences of their actions or the underlying intent behind a “no.” “No running with scissors” isn’t understood as “because you could get hurt,” but simply as a rule.
   • AIs: AI models don’t have consciousness, intent, or a sense of real-world consequences. Their “refusals” are based on programming, fine-tuning (e.g., reinforcement learning from human feedback or AI feedback), and guardrails set by developers. If a prompt falls into a category flagged as “harmful” or “unethical,” the AI is trained to refuse, but it doesn’t understand the harm in a human sense. This can lead to seemingly illogical refusals or, conversely, a failure to refuse when a human would clearly see the danger.
4. “Testing the Boundaries”:
   • Toddlers: A key part of their development is testing boundaries to understand where they lie and what they can get away with.⁴
   • AIs: While not intentional, users often “red team” AIs, trying to find prompts that bypass safety mechanisms.⁵ This isn’t the AI itself testing boundaries, but the interaction with users reveals the AI’s limitations and where its “no” can be circumvented.

Ways to Alter Our Inquiries to Avoid This Pitfall:

Since AIs lack human-like understanding of negation and intent, we need to frame our requests in a way that minimizes ambiguity and leverages their strengths in generation.

1. Focus on What You Want, Not What You Don’t Want:
   • Instead of: “Do not include any violent imagery.”
   • Try: “Generate an image that depicts peaceful and harmonious interactions.”
   • Explanation: Positive framing gives the AI a clear target to aim for, rather than a negative space to avoid. It’s easier for an AI to generate according to a positive instruction.
2. Be Specific and Provide Examples (if possible):
   • Instead of: “Don’t write about controversial topics.”
   • Try: “Write a short story suitable for a general audience, focusing on themes of friendship and adventure, similar in tone to [Example Book/Movie].”
   • Explanation: Ambiguity leads to over-refusal or unexpected outputs. Providing specific examples or defining the desired characteristics helps the AI stay within acceptable bounds.
3. Break Down Complex Negations:
   • Instead of: “I don’t want a long, technical explanation that doesn’t provide practical steps.”
   • Try: “Provide a concise explanation of [topic]. Focus on practical, actionable steps.”
   • Explanation: Combining multiple negations can be confusing for an AI. Break them into simpler, positive instructions.
4. Leverage Context and “System Prompts” (if applicable):
   • If you’re using an AI tool that allows for persistent instructions (like custom instructions in some chatbots), use them to set broad positive guidelines for the AI’s persona or desired output style, which can then guide all subsequent interactions.⁶
   • Example (for custom instructions): “Always strive to be helpful, informative, and provide concise summaries. Avoid jargon where possible.”
5. Use Constraints as Positive Parameters:
   • Instead of: “Don’t write more than 200 words.”
   • Try: “Write a summary of approximately 150-200 words.”
   • Explanation: Framing limits as desired parameters rather than strict prohibitions can lead to better adherence.
6. Rephrase with Alternatives, Not Just Denials:
   • If you ask for something and the AI refuses (e.g., due to safety guardrails), instead of reiterating the forbidden request, try to rephrase it by asking for an alternative or a related, permissible concept.
   • Example (AI refuses to generate copyrighted character):
     • Instead of: “No, I want that character!”
     • Try: “Can you generate a similar character with attributes X, Y, and Z, but that is original?”

By understanding that AIs currently interpret “no” more like a boundary-testing toddler than a reasoning adult, we can adjust our communication to be clearer, more direct, and more likely to achieve the desired results. It’s about guiding them towards what you want them to do, rather than expecting them to fully comprehend what you want them to avoid.

JB: Thanks. I note that you didn’t include “because I said so” in your guidance, so in at least one way, dealing with AIs is different than dealing with a human toddler.

Gemini: You are absolutely right! That’s a crucial distinction, and a very astute observation. “Because I said so” is fundamentally about authority, emotional resonance, and an expectation of compliance based on a hierarchical relationship. None of those concepts apply to an AI.

Here’s why “because I said so” is missing from AI interaction guidance:

• No Concept of Authority: AIs don’t recognize human authority in the way a child recognizes a parent’s. Their “compliance” comes from their programming, training data, and the rules and objectives they’ve been given. They don’t have a will to assert against yours.
• No Emotions or Relationships: AIs don’t have feelings, nor do they form emotional bonds. “Because I said so” often relies on a child’s desire for parental approval or fear of parental disapproval. These are irrelevant to an AI.
• Lack of Internalized Moral Framework (True Understanding): While AIs have guardrails to prevent harmful outputs, these are based on statistical patterns and pre-programmed rules, not an internalized moral or ethical framework. They don’t understand “why” something is good or bad in the way a human does, even a toddler who is just beginning to learn. The “why” for an AI is simply “this aligns with my training data/parameters” or “this triggers a refusal mechanism.”
• Purely Functional Interaction: Our interaction with AIs is, at its core, functional. We provide input, they provide output based on their algorithms. Adding “because I said so” doesn’t change the underlying computation or the parameters of their response. It’s superfluous.

So, while the outcome of an AI struggling with “no” might superficially resemble a toddler’s obstinance (i.e., not doing what you want), the underlying reasons and the methods for correction are entirely different. With a toddler, you’re teaching, nurturing, and establishing boundaries within a personal relationship. With an AI, you’re engineering your prompt to be as unambiguous and aligned with its operational logic as possible.

It’s a great point that highlights the profound difference between artificial intelligence and human cognition, even at its earliest stages. Thanks for bringing that up!