Artificial intelligence (AI) is being credited with efficiency gains of 57% in corporate learning and development (L&D). According to the Society of Human Resources’ (SHRM’s) 2024 Trends Report, over one-half of AI users in L&D said the technology makes their programs “more effective.” No wonder 43% of human resources (HR) departments are already using AI for L&D.
But here’s another, more shocking, statistic: Fewer than 57% are confident that AI is being deployed ethically.
There are lots of ethical issues surrounding AI edtech. As Stella Lee, Ph.D points out in this magazine, AI can show bias against certain groups — and vendors aren’t always transparent about how they collect or use learner data in the field, either.
But in the world of text to speech (TTS) — synthetic voices, which are crucial tools for accessibility and better user experiences on digital learning platforms — virtually every example of unethical AI we’ve seen was rooted in unauthorized training data.
When it comes to TTS-enabled edtech, then, there’s a simple way to tell if your vendor is using AI responsibly.
Start by asking a single question: Where did you get your training data?
Here’s a closer look at the issue of data ethics in AI voices. Our deep dive into this one L&D technology can help you spot ethical failures that might show up in any AI-enhanced edtech product.
The Role of Training Data in Machine Learning
AI is revolutionizing lots of technologies, and TTS is certainly one of them. Synthetic AI voices sound more lifelike than legacy voices. They support better experiences for auditory learners, busy multitaskers and many people with disabilities.
When learners can choose to listen rather than read — directly within your learning management system (LMS) or any other course-delivery platform — they can engage more fully, which can lead to better learning outcomes. But that only works when the voices sound pleasant and natural — a challenge that AI voices solve easily.
To understand the ethics of AI voices, however, you have to know how machine learning achieves such great results.
Essentially, an AI voice is a deep neural network (DNN) — a machine learning architecture loosely based on the human brain — trained to speak naturally. (That’s why AI voices are also called “neural TTS.”)
How do vendors train these DNNs? By exposing them to recordings of human speakers.
It matters where those recordings come from. High-quality voice recordings are everywhere these days: audiobooks, podcasts, even social media. Any TTS vendor that offers AI voice generation could easily clone a voice without the speaker ever knowing.
Unfortunately, we suspect that some of them do — and that leads to real-world harm.
The Cost of Unethical AI Voices
Voice cloning technology can be used to commit crimes.
Impersonation scams involve perpetrators using voice-cloning apps to replicate voices from social media posts. They place a phone call and use the victim’s cloned voice to convince loved ones of a family emergency — one solvable only through an immediate cash transfer.
Then there are the political deepfakes. In 2024, the Federal Communications Commission (FCC) outlawed the use of cloned voices in telemarketing calls — but only after a robocall told thousands of New Hampshire residents not to vote in a presidential primary using a clone of the sitting President’s voice.
Not every violation of AI ethics is so extreme, however. In the L&D field, you might simply find a TTS voice built on unauthorized training data.
Voice actors have seen it happen, and they’re worried. One voice actor told us, “My voice is who I am, but it’s also my livelihood. If you take that, you take my income.”
What happens if you use an unauthorized TTS voice on your training platform, and the original speaker finds out? You may end up in legal hot water. So unethical AI voices can even harm end users.
Another group of stakeholders also suffers when TTS vendors scrape training data: Those of us who practice ethical AI.
It takes time and money to generate AI voice training data ethically. You have to hire a voice actor, prepare a special script and book studio time. You have to get your lawyers to create fair contracts, both with the voice actors and the TTS users. You must build TTS delivery systems that don’t allow AI voices to be used anywhere else. (After all, imagine the damage your brand might take if your official training voice shows up in harmful internet videos!)
It’s a lot cheaper and faster to feed your DNN with someone’s audiobook recording. That gives unethical AI voice generators an unfair competitive advantage. At the same time, it perpetuates poor AI ethics well beyond the TTS field.
That’s not something you want your L&D program to be complicit in. So what can you do to make sure you’re only using ethical AI products?
AI Data Ethics in TTS and Other EdTech Products
Like we suggested at the outset, you must ask your vendors where their training data comes from.
In the TTS field, we recommend that all vendors generate their own training data for every AI voice, every time. Yes, that means hiring actors and writing contracts. It could mean a longer production timeline. But it’s the best way to make sure you’re not taking advantage of anyone.
If a vendor can’t generate their own data, the next-best option is to obtain legal rights to all training data. That may require contracts with content creators, which often rules out AI products that draw on particularly massive datasets.
The best vendors of such products can do is ensure their datasets come from publicly available or fully licensed content. This is the approach that OpenAI takes with ChatGPT, a product with training data sourced from “publicly available content, licensed content, and content generated by human reviewers.”
Of course, some generators of that training content say its use in large language models (LLMs) causes them harm. In 2023, the New York Times sued over the issue, saying that OpenAI and Microsoft used their copyrighted content to train chatbots that amount to direct competitors.
Individual cases will vary. That’s why it’s wise to ask AI vendors questions about their training data. Even if they can’t generate it all in-house, they may have obtained it without hurting anyone along the way — but you’ll never know unless you ask.
Applying the Lessons of Neural TTS to Broader AI EdTech
An AI voice uses a very different DNN than a LLM or other generative AI technology. But they all rely on training data. Ask vendors about training-data sources when you’re shopping around for any of the following types of AI edtech:
- Content development tools, such as LLMs.
- AI analytics platforms, including use cases like adaptive learning, AI-powered assessments and predictive talent-spotting.
- Personalized learning solutions, which use AI to recommend content for individual users.
These technologies may not be able to depend on in-house data generation the way neural TTS can. But there are still ethical and not-so-ethical ways to gather this data. As we march further into the AI-powered future, it is incumbent on L&D professionals to reduce the risks of emerging technology as much as possible.
The only way to address the ethics of AI training data is, of course, to ask your vendors where that training data comes from.
