Imagine a decade from now when Artificial Intelligence (AI) is even more powerful than it is today. You casually mention to your AI assistant, “I need a new outfit for a party,” and return home to find a full-scale runway show set up in your living room. AI-powered robots strut down the catwalk, showcasing the latest haute couture designs, all tailored to your exact measurements.

While this scenario might seem over-the-top, it’s a hyperbolic representation of how AI operates today, just on a much smaller scale. My experience suggests that because AI doesn’t ask clarifying questions, it often expends far more energy than necessary. In an era where concerns about AI’s energy consumption are widespread, establishing a less wasteful approach is crucial.

The Critical Gap: Inability to Ask Clarifying Questions

AI lacks a fundamental capability: the ability to ask clarifying questions. In essence, “they don’t know what they don’t know.” As a result, AI systems often rush ahead, expending significant resources on tasks they might have approached differently had they paused to ask a few questions first. This gap in AI’s capabilities is both crucial and concerning.

Background: Daily Interactions with AI

As someone deeply immersed in the world of AI—coding it, blogging about it, and dedicating my spare time to studying the latest advancements—I’ve observed this issue firsthand. While AI’s tendency to confidently state inaccurate information (known as hallucinations) is well-documented, I’m less concerned about these than many others. My perspective is shaped by comparing AI not to modern computing, which is largely infallible, but to fallible humans. I believe we’ll need to acclimate to some degree of imperfection to fully harness AI’s speed and flexibility.

However, I’m increasingly troubled by AI’s overconfidence. Three recent experiences highlight this issue:

1. Code Modification: While working on a large piece of code (a task that would have taken days just a few years ago), I asked AI to make a routine change—replacing one form of input with another. The AI complied but also removed a significant amount of related code, going beyond my request. Detecting and correcting this unwanted change took far longer than making the modification myself would have.
2. Proofreading: When testing Anthropic’s Claude 3.5 Sonnet for proofreading, it initially seemed to perform adequately. However, a closer examination revealed that the AI had not only corrected typos but also condensed the material by about 70%, losing much of the original detail in the process.
3. Adding Links: Similarly, when I asked Sonnet to add HTML links to this document, the result came out about 50 words shorter. Had it just added links, the word count should have been the same. Instead, it decided to “clean up” some text along the way.

In all three cases, the issue wasn’t necessarily what the AI did—in some contexts, these actions might have been desirable. The problem lies in the AI making assumptions about my intentions without asking for clarification or even informing me of its decisions.

Why This Is Concerning

AI is progressing at an exponential rate. A year ago, it was only helpful when fed small sections of text or code. Now, with improved logical reasoning capabilities and expanded context windows, AI can handle thousands of lines of text or code. While claims of AI building entire products are largely exaggerated, we’re not far from a future where AI will manage large systems for us.

The Importance of Clear Communication

One might argue that the solution lies in providing clearer, more detailed instructions to AI upfront. If typical human interactions require back-and-forth communication, why would we expect human-machine communication to be any different? While I strive to give AI as much relevant information as possible, human communication often involves implicit assumptions or knowledge gaps.

As with humans, more communication isn’t always better. Often additional communication just adds noise and distracts from a good signal. The same is true in AI prompting. Extraneous information can actually make AI prompts less effective. For instance, providing too much contextual information about a coding project’s history or unrelated features might lead the AI to incorporate irrelevant elements into its solution, complicating rather than simplifying the task.

A 50% Solution: Asking AI What It Needs to Know

To address this issue, I’ve experimented with asking AI, “Is there anything else important for you to know before starting this task?” For example, when requesting Claude 3.5 Sonnet to convert a UI interface from Streamlit to Next.js, I posed this question.

The response was a lengthy list of mostly good, specific questions, such as whether I wanted to maintain database integration or switch to local files. However, it also included vague queries like, “Are there any particular performance optimizations or best practices you’d like me to focus on?” While not entirely unhelpful, such broad questions offer limited practical value. The problem is that the list it gave me back was far too large to be worth my time. A few important and specific questions would have been better. In short, explicitly asking AI if it needs more information does lead to better outcomes but would be more beneficial with a smaller, more targeted list of the most critical issues.

Conclusion: The Path Forward

The inability of AI to ask clarifying questions represents a significant limitation in its current capabilities. This gap not only leads to potential inefficiencies and misunderstandings but also raises concerns about AI’s ability to handle complex, nuanced tasks as it becomes more integrated into our daily lives and work processes.

To address this issue, several approaches could be considered:

1. Developing AI systems with built-in mechanisms for recognizing uncertainty and prompting for clarification.
2. Creating more sophisticated user interfaces that facilitate a more natural back-and-forth between humans and AI.
3. Training users to provide more structured, comprehensive inputs when working with AI systems.
4. Implementing safeguards that require AI to confirm major decisions or actions before proceeding.

As AI continues to advance, bridging this critical gap will be essential for creating more efficient, reliable, and truly intelligent systems. By addressing AI’s inability to ask clarifying questions, we can work towards a future where AI not only processes information but also engages in meaningful dialogue, leading to more accurate and useful outcomes.