AI has removed a large part of the effort involved in producing learning content. Give a capable model source material and a reasonable brief, and it can help create course copy, exercises, scripts, summaries, presentations, assessments, and supporting resources at a speed that would have been unthinkable a few years ago.
The result is useful for Learning & Development teams, but it comes with a less comfortable consequence. During a recent Speexx webinar on his new research paper ‘AI in L&D: Around the Core’, co-authored with Eglė Vinauskaitė, Donald Taylor picked up an audience comment about homogenization and gave the risk a name: “the beige wave,” where everything becomes “brown-colored nonsense that you can’t tell the difference between.”
That risk sits inside a broader context of increasing AI usage. Taylor and Vinauskaitė surveyed 531 L&D practitioners across 55 countries and analyzed four years of responses on how people actually use AI. In 2026, 68% of respondents said they were using AI in some part of their L&D work, compared with 54% in 2025.
Content development is by far the most common application. What has changed is the work growing around that core. Practitioners increasingly describe AI helping with analysis, quality checking, needs discovery, evaluation, and internal processes. The technology is becoming more capable at producing outputs just as L&D is discovering that production itself is no longer the difficult part.
The “beige wave” quality problem
Generative AI changes the economics of content. When the effort involved in creating something falls sharply, organizations can create far more of it. That is useful when the original bottleneck was production, but most large organizations were hardly suffering from a shortage of PDFs, courses, videos, slide decks, learning paths, and digital resources before ChatGPT arrived.
Taylor made the tension explicit during the webinar: “We’re going to see this explosion of content… How do we maintain quality?” He was careful about the limits of L&D’s control as well. A learning team can set standards for its own work, but it cannot control every AI-generated presentation, explanation, learning asset, or piece of performance support created elsewhere in the organization.
The research suggests L&D practitioners are already responding by putting more emphasis on checking. Respondents increasingly describe using AI to critique content, test assumptions, validate claims, compare material with specifications, and review work against standards. Taylor described research, quality, and testing as increasingly common AI applications across the categories the study examined.
One respondent captured the underlying discipline clearly. The problem appears when people “don’t properly define what ‘good’ looks like and just ask the AI to generate outputs.” Their response is to build expert constraints into the process so that the system works against a defined standard rather than simply producing fluent material.
Another respondent put the content problem more starkly: “More content does not solve a transfer problem.”
Those comments point to a more demanding role for L&D expertise. Someone still needs to understand the audience and the performance requirement. Subject expertise is needed to recognize a plausible error, while learning expertise is needed to distinguish useful practice from an activity that merely looks polished. Brand, context, accessibility, evidence, and organizational standards do not become less relevant because the first draft arrived quickly.
How organizational readiness is becoming the real constraint
The 2026 research also shows why moving beyond basic generation becomes difficult. The barriers that sit largely within L&D’s own control have fallen considerably, while the obstacles involved in deploying AI inside an enterprise remain stubborn.
In 2024 and 2025, 49% of respondents identified readiness barriers such as lacking the necessary skills, struggling to make the business case, or not knowing where to start. That figure fell to 36% this year. Not knowing where to start dropped from 19.8% to 9.8%, while difficulty making a business case fell from 11.9% to 5.8%.
Putting AI to work presents a different picture. Integration, cost, time, and resistance were identified by 66% of respondents, barely below last year’s 68%. Risk remains higher still, with data privacy and security concerns alone named by 52%. Taylor reinforced that point in the webinar, saying “data privacy and security concerns are the number one concern, and it has been for the last 3 years.”
These constraints become more visible as L&D attempts more substantial applications. Using an approved chatbot to draft an outline requires relatively little organizational infrastructure. Connecting AI to employee data, collaboration tools, proprietary knowledge, learning systems, or HR technology brings information security, procurement, legal review, governance, and technical integration into the work.
The report’s respondents describe this friction without much ambiguity. One organization was “very keen to explore AI” but required approval from three separate teams. Another respondent observed that AI is appearing inside so many software products that IT departments struggle to review them all. A third reduced the problem to the economics of control: “You MUST check everything,” adding that extensive checking can negate the time AI was supposed to save.
This is where organizational readiness becomes part of the quality discussion. Reliable AI in L&D depends on more than a capable model and an L&D team that knows how to prompt it. The organization needs workable rules for data, systems, accountability, access, and human review. Without those foundations, teams can become highly competent at experimenting while remaining unable to move their most useful ideas into normal operation.
The 2026 report describes this as AI arriving faster than the organization can absorb it, and it explains the gap between the pace of AI capability and the pace of enterprise implementation.
Human judgment becomes more important as production gets easier
Quality control is not limited to catching hallucinations or correcting awkward prose. Some of the most interesting responses in Taylor and Vinauskaitė’s research concern the effect AI can have on the thinking that learning is supposed to develop.
One practitioner reported removing AI activities from part of a program after realizing that “the LEARNING was in the thinking and struggle to apply the learning.” In that case, assistance had begun to remove the cognitive work the intervention was designed to create.
This goes to show that efficiency has different value depending on what is being made easier. Automating routine administration, transcription, formatting, or repetitive production work can release capacity without undermining the purpose of the task. The decision about how to use AI changes when the purpose of a learning activity is practice, recall, judgment, or reflection.
A learner on an intercultural awareness course preparing for a difficult conversation may need to formulate an argument rather than receive one instantly. Language practice depends on producing and retrieving language, not simply accepting a polished response. A manager using business coaching to learn to give feedback needs room to make decisions and see their consequences.
This places a different burden on AI-enabled learning design. The question is not simply whether AI can perform an activity faster, but whether delegating that activity supports the capability the organization is trying to build.
Taylor made a related point at the end of the webinar: “Nobody’s asking for more content. What people want is more performance, and that’s two different things.”
The break-out teams using AI to change L&D’s role
One smaller finding in the report provides a useful signal about what L&D can do with the capacity AI creates.
Taylor and Vinauskaitė identified a “break-out” group among workplace L&D and HR teams whose organizations now approach them for different things. Taylor emphasized in the webinar that this was unexpected: “The organization is asking L&D for different things. This wasn’t what I was expecting.”
These teams have not moved away from content production. The report found that 88% of the group described using AI for content development, compared with 80% of other workplace L&D and HR teams. The more significant difference is where else they are putting the technology to use.
They are nearly three times as likely to describe AI being used in operational work, including automation and internal tool building, and twice as likely to use it in employee-facing delivery such as practice, feedback, coaching, and point-of-need support. Their expectations are different too. They show less interest in faster content production and more emphasis on performance and quality.
The evidence does not establish that automation caused the organization to see these teams differently, and both the report and Taylor are cautious on that point. It does show that heavier AI use alone does not explain the difference. Direction appears more revealing than volume.
This is important because Taylor sees a structural risk in content remaining L&D’s main differentiator. During the webinar, he put it plainly: “Anybody in the organization can do it now. Can they do it well? Maybe not. Can they do it good enough? Yes, some of them can.”
A learning function whose value is understood mainly as producing materials therefore faces a harder argument about why the business needs a specialist department to do that work. Professional L&D expertise remains valuable, but its defensible contribution sits increasingly in diagnosis, standards, useful practice, evidence, and relationships with the business rather than production capacity alone.
Content volume is a poor measure of progress
The easy interpretation of generative AI is that L&D can now create more learning content with fewer resources. That can be valuable, of course, but what happens after the production bottleneck disappears?
The answer depends partly on professional judgment and partly on organizational conditions. L&D needs standards strong enough to distinguish useful work from fluent output, while enterprises need the governance, systems, and relationships that allow more ambitious AI applications to move beyond isolated experiments.
The break-out group adds another possibility. Some teams appear to be using AI-created capacity to work differently with employees and with the wider business, rather than measuring progress by how much faster they can produce the same outputs.
Taylor’s “beige wave” is therefore more than a warning about bland AI copy. It describes what happens when production capacity advances faster than the standards and organizational readiness needed to use it well. Once everybody can make more content, producing more of it is a poor place for L&D to prove its value. The real challenge is proving that L&D can move from producing learning to improving the capabilities the business actually needs.
