Best practices for adopting AI in corporate training

There’s enormous hype around AI, which can lead organizations to adopt intelligent technology too quickly without understanding all of the risks. For course creators—and anyone planning to augment creativity and innovation with AI—it’s critical to approach such projects with care
and thoughtfulness.

The key best practices for successful AI partnerships fall into these categories:

Keeping humans at the center.
Protecting data integrity.
Adopting tools with future-proof, flexible architectures.

What keeps you up at night about AI?
_{(Select your response to view poll results)}

Creating corporate AI policy
Sourcing the right AI tools
Reskilling and
upskilling for AI
Building the right
AI infrastructure
Job security

Keep humans at the center

Human-centered AI can empower course creators to rapidly
create magical, high-quality learning experiences.

AI alone can’t build accurate, unbiased, and effective training courses. Courses built solely with AI are very low quality, full of inaccuracies, and lacking in instructional design principles that ensure positive outcomes for training.

Organizations can’t overcome these faults without the intimate involvement of skilled instructional designers and e-learning professionals in the course creation process. To ensure that training is high quality and highly effective, these people need to stay in the driver’s seat while using AI as a partner—not a replacement—for their work.

Implemented well, AI can free up course creators’ time to apply their skills to creative learning approaches, since AI allows for rapid iteration and access to source data and images. Instead of replacing course creators with AI, organizations should encourage creators to harness AI to inspire creativity and speed up mundane tasks.

Course creators should be in control of AI—not the other way around. For example, with Articulate AI Assistant, course creators use AI to outline and draft content quickly by prompting to research or using source material. The time savings from reduced manual processes boost productivity and give course creators the freedom to get more creative to achieve learning objectives. AI Assistant enables authors to edit block structure and content, convert blocks instantly, and improve writing and tone—all from within the creator workflow.

Articulate builds its course design tools with this iterative, collaborative process top of mind, using core instructional design principles and frameworks. For example, the first step in generating a course involves AI synthesizing the content and suggesting learning objectives. The learning objectives are guided by core ID frameworks, such as Bloom’s Taxonomy, to support a learner’s ability to recall facts and basic concepts, and to use information in
new situations.

After aligning on effective learning objectives, the Articulate AI Assistant proposes a course outline, which the instructional designer edits, approves, and updates as many times as needed to ensure high-quality content. From there, the AI Assistant iterates hand in hand with the course designer—just as a human assistant would—to generate lessons, knowledge checks,
and videos.

Attributes of human-centered AI

AI accelerates mundane tasks.
AI-generated content is a starting place, not an end point, for course creation.
AI and course creators are working together to generate the best outcomes.
Learning frameworks are built into AI functionality.

Course creators can harness the power of AI to make their work
even better. And what the creators have to remember is that
using AI intelligently and thoughtfully has become an essential
career-building skill.”

Adam Schwartz | founder and CEO of Articulate

Protect data integrity

AI can create a legal landmine for organizations’ proprietary training if they don’t safeguard
data and course ownership, protect confidentiality of content, and mitigate biases.

Ownability
In the absence of contractual and operational assurances about AI approaches, avoid inputting copyrighted or otherwise protected IP or confidential material into AI tools. Why? Inputting this data into an AI tool runs the risk of forgoing protections or IP rights on that information. Such practices might also breach the protections or IP rights of others.

Confidentiality
To prevent others from accessing or using sensitive or confidential information when inputting content AI tools, an organization should
do the following:

Always involve humans in the content creation process.
Review vendor terms to see if they prevent proprietary content from being contributed to other outputs or used to train AI models. Also, check that the terms state that the organization’s content remains the organization’s alone, and is not otherwise usable by the vendor. These restrictions are more likely to be available through paid licenses.
Set up the correct operational and governance practices in the organization, such as an AI Use Policy, to monitor use of the AI tool—including what’s inputted and how it’s used.
Don’t input proprietary content into an AI tool unless there are the correct assurances in place with the vendor in order to protect data confidentiality or prevent content from being used in training a third-party’s AI tool.
Proceed with caution when using AI tools that take in data from a variety of sources. Using content generated by such a tool can put the organization at risk of IP infringement, inaccuracy, and bias.

Bias
AI is only as bias-free as its training data. Purely AI-generated training—that is, AI content that is not paired with human input—will reflect bias and include inaccuracies. While technology will continue to evolve in an attempt to address bias and inaccuracies, it currently falls short.

Organizations should never use AI-generated training without human input at every step to make sure that they don’t share these biases or inaccuracies in workforce training or elsewhere. Human judgment must supplement AI and machine learning systems. These technologies are simply not equipped to decide when and whether bias has been sufficiently eliminated—or that content is wholly accurate and safe for
broad distribution.

Unless assurances are given about ownership, organizations should assume that any generative AI outputs are not “ownable.”
While a “work” that is part human-authored and AI-authored may be copyrightable, AI-generated portions of the work alone are not protected. In the United States, it’s expected that the U.S. Copyright Office, as well as the legal court system, will continue to oppose the protection of any solely AI-generated content.

Which statement do you agree with most?
_{(Select your response to view poll results)}

We have a clear AI data policy in place
We are creating an AI data policy now
We haven't started, but plan to implement an AI policy later
We do not plan to have an AI policy

Articulate’s approach to protecting data:
When customers use Articulate’s creator platform for workplace learning, their data and content remain theirs. They own the rights to any and all content they create using our licensed products, and their content will not be used to train data models outside their company walls.

Best practices for reducing bias and protecting data

Choose AI vendors that don’t use your company’s data to train third-party
AI models.
Be sure that humans check for bias and accuracy throughout the
creation process.
Remember that it’s impossible to copyright content created solely by AI.

Adopt tools with flexible, future-proof AI architectures

AI tools at every layer of the technology stack must have flexible architectures that
can keep pace with the speed of innovation.

Because the field of AI is evolving at an unprecedented rate, there’s a real risk of organizations quickly falling behind if they don’t choose tools that have baked into their very architectures the ability to continuously improve and take advantage of the newest AI developments. In short, all technology tools an organization adopts must have flexible AI architectures that can adapt to the rapid pace of innovation in AI.

To evaluate whether technology tools have the flexibility to evolve, it’s important for organizations to understand the layers of the AI tech stack within the tools they’re evaluating. Imagine the AI tech stack embedded into learning technology as a multilayer cake. Starting from the bottom up, there’s the foundation model then the domain model and finally the governance and application layers.

Take advantage of ever-changing foundation models

The basic function of the foundation model is to understand and generate. There are a number of foundation models tuned for different tasks. Some models are better than others at specific jobs, such as writing code, generating learning content, or creating images. Large language models (LLMs) are excellent at understanding language and generating content. Diffusion models understand language but generate images.

It is absolutely critical that organizations choose tools that can flexibly evaluate, test, and adopt new foundation models as they evolve. Tools should be able to take advantage of new and evolving foundation AI models to generate content, quizzes, images, translations, and videos so that they get better and better every month to save organizations time and offer the highest-quality output.

Incorporate expert and proprietary domain models

The next layer in the technology stack is the domain model. While most foundation models are trained by content on the internet, a domain model is trained by the context, customers, decisions, data, and all activities that make up an organization or an industry.

The more relevant the domain models, the more value an organization will derive from its AI tools. For corporate training, look for AI-infused tools that incorporate instructional design best practices and L&D sources into the domain model. They will generate more-accurate, more-effective, and more-relevant content, producing better results for the organization.

How comfortable do you feel with the current AI tech stack landscape?
_{(Select your response to view poll results)}

Very comfortable—AI is my second language!
Pretty comfortable—I follow
the conversation
A little comfortable—I've tried some AI tools but I don't understand how to use them
at work
Not comfortable—I need more information to get the basics

Make sure the governance layer is robust

Above the domain model layer in the technology stack is the governance layer. This layer lets organizations derive data insights, ensure oversight, and manage security. It’s in this layer that organizations audit, validate, and control data and associated models, continuously monitoring for quality and trust.

To ensure quality and compliance, organizations need to choose AI-infused tools with governance layers that evaluate the quality of the content generated. Ask tool vendors how they know if the quality of AI-generated content is improving or getting worse.

Look for a human-centered application layer

As noted above, a human-centered AI approach augments human capabilities and insights, while automating tasks cost effectively and in ways that unlock and support human creativity. Human-centered applications not only amplify the work of course developers, they also eliminate mundane clean-up work.

Look for course creation applications that not only embed instructional design principles, but also support rapid, constant iteration to produce high-quality outcomes. Best-in-class application layers feel seamless, acting like valuable assistants that help course creators create better courses faster.

Best practices for future proofing the
AI-powered tech stack

Look for tools that let you test, iterate, and adopt new foundation models as they become available.
Adopt tools that integrate domain models that include industry-specific and company-specific data.
Double-check that the AI tools you use have robust governance layers that monitor and evaluate the quality of
AI output.
Choose tools that have a human-centered approach that unlocks creativity and productivity.