Workers are not just experimenting with large language models anymore. They are building entire workflows around them - and a new dataset published in December 2025 shows exactly where that trust is being placed, and in what volume.

Paligo, a structured content management platform based in Solna, Sweden, published research on December 8, 2025, analyzing usage data from AIPRM's public library of ChatGPT prompts and GPTs. The dataset covers more than 74.3 million uses across 5,000-plus prompts spanning 56 categories of work tasks. The findings offer one of the clearest empirical pictures yet of how organizations are directing generative AI toward specific job functions - and which categories absorb the most activity by a significant margin.

The numbers are stark. Article and blog writing leads all categories with 35.3 million uses, a figure that is five times larger than the second-ranking category. Nothing else comes close.

Writing at industrial scale

According to the Paligo research, article and blog writing alone accounts for 47.4% of all automation demand captured in the dataset. The most widely used individual prompt in the entire sample - accessed 1.9 million times - is described as a "humanized, SEO-optimized, long-form article writer." That a single template of this kind would accumulate nearly two million uses points to how systematically content teams have moved toward LLM-assisted production.

Writing-related tasks appear more than once in the top ten across all categories. Creative and scriptwriting prompts logged over 1.4 million uses; marketing copy prompts recorded the same. Call-to-action prompts, which generate short persuasive text for web pages and applications, reached 1.2 million uses. Editing and proofreading templates attracted more than 360,000 users, a lower figure that Paligo attributes to the ease of pasting text directly into a chat window and requesting corrections without a structured template. Academic writing - despite representing only 24,106 uses - ranked tenth in the writing subcategory, a figure that carries weight given mounting evidence of student reliance on generative AI for assessed work.

According to the Higher Education Policy Institute, the proportion of higher education students using generative AI to create material for assessed work rose from 54% in 2024 to 88% in 2025.

The pressure to produce written content at scale is not unique to academic settings. Among marketers specifically, SurveyMonkey data cited in the Paligo research shows more than half use AI to optimize content, and over 40% rely on it to create material, brainstorm ideas, support social media strategies, and analyze data. Social media automation prompts accumulated more than 6 million uses in the AIPRM dataset, placing the category third overall after article writing.

Marketing tasks as a category reached 4 million uses. Further down the ranking, link-building and outreach prompts recorded 372,350 uses, audience segmentation 163,215, and media channel selection 69,612. The Paligo analysis describes these as tasks that "traditionally rely on judgment, nuance, and experimentation" - the kind of strategic work that organizations are increasingly asking AI systems to shortcut. Whether that shortcut improves outcomes or compresses them into something generic, the researchers note, remains uncertain.

The image generation layer

A notable share of the dataset reflects a pattern that goes beyond direct text generation. Workers are using LLMs as a prompt engineering layer for other generative media systems rather than as an output tool in itself.

Midjourney-focused prompt templates recorded 2.7 million uses - the highest among all image generation platforms in the sample. One of the most widely used Midjourney helpers, DishPrompt, has over 4,700 uses and generates food imagery prompts from recipe descriptions. Stable Diffusion templates logged 439,545 uses; Leonardo AI followed at 372,762. Newer or more specialized platforms - Sora, Flux, Firefly - recorded lower figures, suggesting that community momentum and established utility outweigh raw model performance in driving adoption at scale.

Audio-first tools such as AudioCraft recorded 40,681 uses, significantly below the visual generation tools. The appetite for AI-assisted image creation appears considerably broader than for audio, at least within this dataset.

The broader context for image generation is sizable. An estimated 15 billion images have been created through generative AI since 2022. Brands including Coca-Cola integrated AI-generated media into major television campaigns in both 2024 and 2025.

Where data, strategy, and customer tasks land

Beyond writing and visual creation, the AIPRM dataset shows meaningful activity in several other categories. Data and analysis prompts - covering tasks from modelled testing to spreadsheet cleaning and data engineering - accumulated more than 2.6 million uses in total. Business planning and strategy prompts, which include tools for startup ideation, product naming, and pricing recommendations, recorded over 2.3 million uses. Customer interaction prompts, covering draft responses and CRM strategy, reached more than 1.3 million uses collectively.

These numbers describe a labor market actively trying to delegate cognitively demanding work - not just formatting or repetitive text - to automated systems.

The broader AI adoption context provides important framing. According to MIT research cited in the Paligo report, 95% of companies have seen zero return on their AI investments so far. A Financial Times analysis of S&P 500 company filings identified a similar pattern: major U.S. firms struggle to articulate concrete benefits from AI even after publicly building hype around AI solutions in earnings calls.

That gap between adoption volume and measurable return shapes how the Paligo findings should be read. High prompt usage indicates intent and experimentation. It does not necessarily indicate productivity gain.

The marketing sector is dealing with its own version of this pressure. Research published by Adobe in January 2026found that 84% of marketers work past scheduled shifts, with 78% doing so nearly five times per month - amounting to roughly 55 extra days of unplanned work per year. Adobe's study of 1,106 U.S. workers across seven industries found that marketers currently automate only 18% of their daily tasks, while estimating that another 29% could theoretically be automated. The structural gap between what is possible and what is actually implemented remains wide.

The quality problem that prompt volume cannot fix

Paligo's research makes a deliberate point that sits outside the usage data itself. The organization produces structured content management software, and its report uses the automation demand figures to argue that technical documentation - a field where accuracy, version control, and audit trails are non-negotiable - is particularly poorly served by LLM-based workflows.

According to the Paligo report, a survey by Terzo and Viz Capitalist of more than 1,700 businesses identified inaccuracy as the top risk associated with AI content generation, ranking above job displacement and cybersecurity concerns. Companies are already rehiring positions that were cut in order to fix errors in AI-generated content.

OpenAI's policy revision in October 2025 - limiting ChatGPT's ability to provide licensed professional advice including medical or legal guidance - demonstrated that even AI providers acknowledge certain content domains cannot absorb the error rates inherent in automated generation.

The Paligo report argues that five structural limitations prevent LLMs from adequately replacing professional documentation work: accuracy requirements that cascade into customer integrations and compliance violations; the need for updates to propagate systematically across hundreds of linked documents when a single product or regulation changes; audit trails and version control that regulated industries require; single-source publishing principles that allow one piece of content to appear across web help, PDF manuals, in-app tooltips, and AI agent knowledge bases; and the fundamental inability to eradicate hallucinations regardless of prompt quality.

As the report notes: "There is no evidence that the rate of hallucinations from LLMs can be eradicated regardless of the quality of a user's prompt."

What this means for marketing professionals

The advertising and marketing technology sector has watched AI automation reshape campaign management at a comparable pace to content production. OpenAI's early-stage advertising infrastructure build-out in late 2025 reflected similar logic: that sufficiently automated systems could compress expertise-intensive workflows into prompt-based interactions. Agentic AI infrastructure dominated advertising industry discussions heading into 2026, with McKinsey tracking $1.1 billion in equity investment flowing into agentic AI during 2024 and job postings related to the technology increasing 985% from 2023 to 2024.

The Paligo dataset suggests that the demand for automation is real and accelerating across all content-adjacent functions. Social media, marketing copy, audience segmentation, and strategic planning are all being pushed into LLM workflows at scale. What the data cannot measure is whether the outputs are being used, corrected, replaced, or quietly discarded.

The rise of agentic AI - systems capable of planning, executing, and completing multi-step tasks without human intervention at each stage - represents the logical next step from the prompt-based workflows in the AIPRM dataset. LiveRamp introduced agentic AI tools for marketing automation on October 1, 2025, describing a shift from chatbot interactions to autonomous agents managing complex workflows independently.

A 2024 study cited in the Paligo report linked overreliance on LLMs among students to measurable weakening of critical thinking, decision-making, and analytical abilities. The same dynamic - skilled workers delegating judgment-intensive tasks to automated systems and receiving plausible-sounding but unverified output - is now playing out at organizational scale across content, marketing, research, and strategy functions.

According to a UNESCO survey of higher education institutions referenced in the Paligo report, more than half of respondents admitted feeling hesitant about using AI effectively in teaching or research, and one in four reported that their institutions had already faced ethical issues ranging from student overreliance to authorship disputes and biased outputs.

A study released by Graphite in October 2025 found that 50% of new written content on the internet is likely produced by AI. Whether that volume of output is improving the information environment for readers - including marketing professionals making strategic decisions - is a separate question from whether it is being produced.

Timeline

Summary

Who: Paligo, a structured content management platform headquartered in Solna, Sweden, conducted the research. The study analyzed data from AIPRM, a platform where users share prompt templates for ChatGPT and other large language models. The findings are relevant to marketing professionals, content teams, documentation writers, and organizations deploying generative AI in workplace workflows.

What: Paligo analyzed more than 74.3 million interactions across 5,000-plus prompts and GPTs in 56 work categories. Article and blog writing leads all categories with 35.3 million uses - five times more than the second-ranking category. Social media prompts recorded 6 million uses; Midjourney-focused templates reached 2.7 million; data and analysis prompts exceeded 2.6 million; business planning and strategy prompts surpassed 2.3 million. The research simultaneously argues that LLMs are structurally unsuited to technical documentation work due to accuracy requirements, lack of version control, hallucination rates, and the need for systematically propagated updates.

When: The research was published on December 8, 2025. The underlying dataset reflects AIPRM usage data current as of November 2025.

Where: AIPRM is a web-based platform where users share and access prompt templates for ChatGPT and other generative AI workflows. Paligo is based in Solna, Sweden.

Why: Paligo conducted the research to quantify where workers are deploying LLMs at scale - and to make the case that high automation demand does not translate to high automation quality, particularly in domains such as technical documentation where errors carry organizational and compliance consequences.

Share this article
The link has been copied!