The Uncanny Valley of Voice: How Realistic AI Audio Undermines Trust

Written by

Paul Edwards

in

The Uncanny Valley of Voice: How Realistic AI Audio Undermines Trust
The Uncanny Valley of Voice: How Realistic AI Audio Undermines Trust
0:00
–:–

I was listening to a podcast recently from someone whose voice I know well. Nothing was obviously wrong. The content made sense and the delivery was polished. The tone was incredibly close to what I expected.

But it didn’t take long before something felt off. The rhythm was slightly too smooth, and the phrasing felt a little less lived-in. It sounded almost like them, it was incredibly close, but not quite, by the smallest of margins. So I asked them directly: “Just curious. Have you changed something about your podcast? Did you do the last one with AI?”

They replied: “Yes I did. Interesting you spotted it, no one else has. It was a test.”

That exchange highlighted something I think we’re about to run into much more frequently in business communication. This isn’t just a “podcast production” issue. It’s a brand trust issue, because more and more businesses are turning blog posts into audio, summaries, and social content by leveraging AI tools.

However, as we all know. Just because we ‘can’ doesn’t necessarily mean we ‘should’. So, let’s dig into that a bit more.

Just because we ‘can’ doesn’t necessarily mean we ‘should’.

Why I haven’t cloned my own voice.

I’ve sincerely thought about using AI to generate my own voice for podcasts or audio content for a while now, and as someone who’s not a natural extrovert, it’s very easy to see the appeal in that.

Recording audio takes time. You need a quiet space, decent equipment, enough energy, and enough confidence to speak clearly without stumbling over pronunciation. Then there’s editing, cleaning up mistakes (like the endless seagulls that decide to start screeching as soon as I click record), exporting files and a bunch of other time consuming tasks. For a marketing team of one, that’s a lot of friction that can make publishing content very difficult to achieve with the regularity that’s needed.

AI voice tools promise to remove much of that work. In theory, I could write a blog post, turn it into a script, feed it into an AI version of my voice, and publish an audio version, all without speaking into the microphone. Despite the obvious benefits, I still feel uncomfortable with the idea of using a synthetic version of my own voice in public-facing content, not because it’s AI, but because it feels dishonest somehow.

The problem isn’t AI.

An obviously AI-generated audio format, such as a NotebookLM-style discussion, doesn’t make any pretence to be something other than it is and I think that the listener understands that. They know they’re listening to an AI-assisted output based on source material that it’s been given. Even so, where I’ve experimented with this format, I still openly declare the use of AI in the introduction and in the show notes.

That feels cleaner to me. It’s artificial, but it’s honestly artificial, and I think that really matters.

It’s artificial, but honestly artificial.

A cloned voice is different… A cloned voice borrows the trust people already have in the real person. It uses the sound of familiarity, and it carries the emotional signal of someone’s voice, even when the person didn’t actually speak those words aloud. That’s where things get a bit tricky…

Even if the script is mine, if I review every word, even if I fully approve the final result, the delivery still isn’t really me. For some uses, that might be fine but for content built around expertise, reassurance and trust, I think we need to be incredibly careful not to inadvertently damage the trust people have placed in us. Our livelihood literally depends upon it.

Why “almost human” can be worse than “obviously AI”

A clearly synthetic voice is easy for the listener to process. The brain hears it and files it away as artificial. You might like it or dislike it, but there’s no confusion about what it is and it doesn’t cause any distracting thoughts that may cause the message to get lost.

A real human voice is also easy to understand. You hear the pauses, the hesitation, the emphasis, the humour, the rough edges, and the little imperfections that make the person feel present and real. For many small businesses, the personal element is part of the value they buy.

The strange bit is the middle ground. A voice that sounds 90 percent like someone you know can be more unsettling than a voice that sounds completely artificial, because your brain keeps looping over and checking for the missing 10 percent. Something is slightly too smooth. The emphasis lands in the wrong place. The emotion is nearly there, but not quite. The voice sounds like the person, but the human presence is thinner or doesn’t follow patterns that your memory of that voice expects.

This is another incarnation of the uncanny valley problem. Often, uncertainty isn’t always a conscious thought. It can show up as a vague sense of unease or alert on a sub-conscious level, and that’s enough to change how much you trust what is being said. Think back to when you last had a gut feeling that someone was lying to you. There may have been the tiniest of changes in inflection, the length of a pause or a change in pace. It takes that little for you to feel something is wrong, and AI just hasn’t got that right yet. Not enough for us to risk our reputation on it’s output.

Voice carries more than words. When someone listens to you, they’re not just processing information, they’re reading a person. They hear emotional state, confidence, authority, warmth, frustration, humour, enthusiasm, and care, and that’s especially true if you’re a consultant, coach, trainer or anyone whose work depends on people trusting your judgement.

When things don’t “match”, people feel it.

Why “obvious AI” can be more honest

This is why I find myself more comfortable with an obviously AI format than an AI clone of my own voice.

If I used a tool such as Google’s NotebookLM-style to generate a podcast style audio discussion, the format would be clear. It would clearly sound like an AI-generated conversation (albeit an incredibly convincing one) and that gives the listener an honest frame. They’ll likely think: “This is a useful summary of Paul’s article,” rather than: “Paul recorded this podcast.”

The content can still be helpful, provide value, and the listener can still benefit, but the relationship is clearer. The AI isn’t impersonating the human. It’s helping package the material in another format, and that feels like a much healthier line to draw right now.

Where I’d be more careful

There are situations where I’d be even more cautious about using a cloned human voice.

I’d think twice about using it for personal thought leadership, client reassurance, sensitive updates, crisis communication, sales messages, or anything where the listener reasonably expects the person behind the voice to be genuinely present.

If a client is worried about a problem with their website, I would not want to send them an AI clone of my voice explaining what happened. That kind of communication needs human judgement. It needs care, and it needs the little imperfections that tell the other person there is a real human being paying attention.

The same applies to content where your credibility is the point. If people follow you because they value your perspective, your delivery is part of the relationship. Replacing it with something that sounds almost like you might save time, but it might also dilute the very thing people came for.

Use AI to support communication, not to disguise who is communicating.

A simple rule for businesses using AI voice

For me, the rule is simple: use AI to support communication, not to disguise who is communicating. That means AI audio should make content easier to access, easier to repurpose, or easier to understand.

This doesn’t mean every use of AI needs a dramatic disclaimer, but it does mean businesses should think carefully about audience expectations. Would the listener reasonably assume this is a specific, real person speaking? Would they feel misled if they found out it was AI-generated?

Is the voice being used to borrow trust from a real human relationship? If the answer is yes, it’s worth considering your use of AI very carefully.

Most teams don’t need a big AI policy document. They need a few sensible defaults they can actually follow, and someone experienced to sanity-check the grey areas. That’s the kind of thing I help with as a website consultant: making practical decisions that protect trust, without turning content into a laborious compliance exercise.

Transparency isn’t being overly apologetic. It’s a trust signal.

Some businesses worry that disclosing AI use makes the content seem less valuable somehow. Actually, I think the opposite may be true. A simple note like this would be enough in many cases:

“This audio was generated using AI from an article written and reviewed by our team.”

Or:

“This is an AI-generated discussion based on our original guide.”

That isn’t an apology. It’s a trust signal. It tells the audience what they’re listening to, where the ideas came from, and how much human oversight was involved. For small businesses and marketing teams, that kind of clarity matters. You don’t need to avoid useful tools, but you do need to use them in a way that protects the trust you’ve already built. The trust you have accrued over time has a value, and it also has a cost to rebuild.

The trust you have acrued over time has a value, and it also has a cost to rebuild.

My current view

I’m not against AI-generated audio at all. I may end up using it myself. I just don’t think the goal should be to make AI sound enough like me to avoid the uncanny valley. The goal should be to make communication clearer, more accessible, and more honest.

If you take one thing from this, make it this: don’t make your audience guess who’s speaking. If the format is AI, make it obviously AI. If it’s meant to be you, perhaps you should record it yourself.

If you want help drawing a sensible line for your organisation, I can help. You are very welcome to reach out to me via my contact page or by giving me a call on 01903 527927.

Appendix: links mentioned