Why Prompt Engineering Shouldn’t Exist

If AI is as advanced as it claims to be, why are we putting so much effort into “engineering” prompts?

Rather than using your normal human voice to ask AI a question or provide an instruction, prompt engineering involves rephrasing questions or instructions to get better results from AI. Instead of asking ‘What US state produces the most tomatoes?’ an engineered prompt may read ‘tomato production state’ to improve result accuracy. 

As information technology advances, AI is increasingly able to recognize more natural language rather than depending on a slew of strategic keywords to produce desired outcomes, but results are still not perfect.

Ideally, rather than training users to engineer prompts in a way that helps AI offer better responses, shouldn’t the focus be to teach AI to answer human-like questions with human-like answers?

Typical Prompt Engineering Use

In the early years of machine learning, tools like Google search required users to  significantly simplify their queries. Search was designed to simplify the process of finding information, but due to the limited computational power and algorithmic tools available at the time, responses relied only on keyword matching and combinations of keywords. Simpler queries outperformed natural questions.

Over the years, users grew accustomed to offering engineered prompts to AI. When using a search or AI tool, users understand simplified prompts often work better and automatically adjust their language when using this kind of technology. This often involves testing various keyword combinations to look for different results – a process that requires more time and energy than would be necessary if using natural language was possible.

The evolution of machine learning and AI has evolved exponentially since the early days of stringing together specific keywords to produce an intended response. Today, prompt engineering and large language models (LLMs) assist with tasks including detailed data analysis, language correction, content production, recommendation generation, and chatbots. While the technology producing these various ‘intelligent’ responses is far superior to what it was just a few years ago, it still has limitations.

Prompt Engineering Challenges

Despite technological progress, effective responses still heavily rely on users’ ability to craft effective prompts. Without the ability to comprehend human nuances in tone, phrasing, and specific language variations, questions asked in different ways may receive inconsistent responses, despite the questions essentially meaning the same thing. 

Certain tones, phrases, and structures tend to provide different outcomes. For example, statistical correlations between polite language and higher-quality, more helpful responses were seen, but not due to the machine’s appreciation of pleases and thank yous. This is instead due to the correlation of massive internet data sets. As such, users learned to input polite phrases into their search, but this tactic did not help the AI to become better at understanding tone or context.

The challenges center around AI’s reliance on engineered prompts. If prompts are adjusted to cater to AI responses, the responses only change as the prompts do, and only in the singular platform used. Moving to another application or using another language, different slang terms, or even different tonalities puts the user back at square one. The user is the one who has to – again – relearn the best way to adjust the prompt rather than encouraging smarter and more accurate responses, making it difficult to replicate processes or scale efficiencies for multiple uses or platforms. 

Rethinking the Value of Prompt Engineering

Instead of focusing on building LLMs that handle natural language queries seamlessly, the industry seems to embrace prompt engineering as the norm. Tests and benchmarks often evaluate AI performance based on engineered inputs. If LLMs are judged on how well they respond to engineered prompts rather than plain natural language queries, they will evolve to expect and rely on those engineered inputs. Users risk locking themselves into a system where prompt engineering is always necessary, drifting away from the true goal of natural and intuitive interaction.

Instead of training users to adapt to AI’s shortcomings, resources could be focused on directly addressing these limitations, paving the way for more robust, human-like systems. Rather than spending hours of research and millions of dollars on determining the best way to present the query, efforts to address the actual weaknesses of LLMs could broaden AI’s ‘understanding’ of various prompts, moving ever closer to the goal of human-like understanding and interaction. 

If this approach does not become the norm, continuous prompt engineering not only blinds users to the dream of realistic responses, but flaws in various systems can remain hidden and prevent further improvements from being integrated. Additionally, LLMs excelling at natural language but offering subpar performance with prompt engineering might go unnoticed in the market if value continues to be placed more heavily on engineered-prompt performance.

Advanced AI Approaches

Improvements to LLMs can reduce reliance on prompt engineering. There is exponential value in perfecting AI models by helping them to understand nuances in tonality and language differentials. The development of more effective models and systems that expand natural language understanding will support users in a way that closely mimics human conversational interactions. As more applications mature to incorporate expanded varieties of prompts, the need to engineer them to cater to specific models can be eliminated.

A possible solution is a system that chooses the best LLM for a given query without requiring the user to do any prompt engineering. Imagine a model that receives the user’s input, understands it as natural language, and then automatically selects an LLM best suited to handle that query. Implementing this kind of “intelligent model selection” (IMS), it could encourage LLM providers to compete on natural language performance. This would push the field towards the original goal – truly natural, effortless query-to-answer interaction.

The Future of AI Without Prompt Engineering

Prompt engineering reflects a fundamental misunderstanding of AI’s purpose. Instead of improving the underlying models to handle natural language queries as intended, users are being trained to craft inputs to coax better results. This normalizes a practice that undermines the initial vision for AI. The future should be about creating more robust AI systems that genuinely understand natural language rather than requiring people to learn how to manipulate prompts.

IMS could realign priorities of these technological advancements and restore focus on natural language capability. By focusing on creating AI systems that genuinely understand human communication, we can move closer to a future where prompt engineering becomes obsolete. Such advancements would ensure that AI fulfills its potential as a tool for effortless, human-centered interaction leveraging natural language understanding and aligning with human values.