It’s no secret that AI tools have quickly become an important and ever-developing asset for millions of workers looking to streamline their professional processes. Generative AI’s integration into many of our existing workplace tools, through something as simple as a chat window, has allowed workers to casually adopt the technology. However, this informal-feeling integration gives way to a harder truth: most workers are not proficient in AI prompting.
A recent study from Adobe Firefly scored over 1,000 text-to-image prompts and found that the average person earns a “C” grade, scoring an average of 57 points out of 100. On top of this, not a single respondent scored an “A” grade–a result that signals a collective skills gap between how workers prompt and what will give the best result.
This skill gap is incredibly important to address, and will only become more important as AI tools become further integrated into the workplace. The difference between a casual and more proficient AI user no longer has to do with access to the technology but the discipline an employee brings to the tool and how they maximize task efficiency. This study suggests that most workers are bringing the wrong instincts to their prompting.
The strange psychology of conversing with machines
One of the most revealing findings from the study has to do with the ways workers perceive the machine on the other end of their prompt. Many users are speaking with their AI tools as though they were humans, bringing common human social habits into their prompts: something that the data reveal has no bearing on output quality. A direct example of this can be found in capitalization: about 1 in 7 workers think that typing their prompts in ALL CAPS will produce better AI output. This pattern divides sharply across gender lines, with men being 80% more likely than women to believe that shouting at their AI tool improves its response.
Using polite phrasing when speaking to AI is another human element workers are taking into their prompting, especially in specific industries. The study found that those working in the finance and banking field said “please” to their AI tools 43% of the time, closely followed by workers in education, transportation and logistics at 42%. While there’s nothing wrong with general courtesy and politeness, evidence is clear within the study that these pleasantries don’t necessarily lead to better AI output. This difference in belief compared to real output quality boils down to a misunderstanding of what AI tool systems actually respond to.
What makes this finding so interesting is not that workers aren’t prompting optimally yet, but the why behind the disconnect. Workers are matching the speech styles they use within their own human conversations because of the lifelike, chat-based format offered to them by AI tools. Speaking to the tool with a caps-locked command or a polite format are means for workers to apply emotional color to their request, when the AI tool itself isn’t processing that aspect of the task. Being able to understand this mismatch is one of the first steps workers can take to begin prompting for efficiency.
Prompt confidence doesn’t equal prompt competence
The study’s data also sheds light on a gap workers are experiencing between their perceived prompt proficiency and true prompting skill levels. Gen Zers have reported the highest confidence level of any group, with 54% of Zoomers rating their own prompting skill at four out of five. However, their average score level landed at 56/100.
This disconnect arose across other divides also: men were 15% more confident than women in their prompting skills but produced prompts that scored only 5% better. Despite these divides, it appears the prompt playing field is more level across the workplace than one might expect–with entry level workers and directors both scoring 55% proficiency. In other words, prompting skill correlates to habit and discipline more than confidence or seniority.
Skill gaps made by omission instead of true error
The ability to discern what needs to be included or omitted from a prompt might be the biggest differentiator between an introductory and advanced prompter. The study revealed that the biggest prompting problems workers are facing have more to do with what people are leaving out from their prompts than what they actually include. When a prompt lacks a detail, the AI tool will fill that gap with its own assumption, creating a final product that may miss the mark.
The most common omissions found in the study were soft, framing elements: output tone, a concrete example of desired style, or a defined role the AI tool should adopt when creating their response. Even though these facets may feel like they only garnish a prompt, omitting them has real impact on output and user experience, especially in a visual medium like image generation.
Trending towards more intentional prompting
The most encouraging aspect of this study is that the fixes users need to make to their prompting doesn’t require technical skill or expensive training. The most proficient prompters approached prompting as a learned skill–asking for a complex product in smaller, more deliberate steps rather than one large request. These prompters have built habits around reuse as well, saving successful past prompts to build on and revisit later for similar requests. The study’s highest performers also treated fact-checking their output as a crucial part of their process rather than an occasional step. Advanced techniques like asking the AI how to prompt it have also begun to emerge.
What we can take away from this study is a true look at workers’ developing AI skillsets. Workers as a group are collectively at an early stage of learning these skills, and are actively looking to improve our prompting and AI partnership within their work. The organizations and individual workers that will find the most success going forward are those who treat prompting like a unique skill rather than a casual conversation: using specific, deliberate and ever-improving techniques to work alongside these tools.
