For many science fiction fans the computer voice on the Enterprise is iconic, and is, in fact, “the” voice of any computer. Majel Barrett deliberately gave the computer a stilted, monotone, mechanical voice to reflect the fact that it was a non-sentient machine. But also, it reflected the technology at the time. Voice synthesizers had been around since the VODER, first debuted at the 1939 World’s Fair. They produce an electronic-sounding definitely artificial voice. Many people are also familiar with this electronic voice through Stephen Hawking’s voice synthesizer.

But now we have the technology to create artificial speech that sounds convincingly human – the voice sounds natural and not electronic and contains all the inflection and nuance of natural speech. Sesame is one example, and you can try the demo for yourself. The app adds pauses, breaths, “ums”, and other fillers to sound more natural. Essentially it is now simply not possible to tell an AI generated voice from a human generated voice. I recently rewatched Star Trek TOS and paid close attention to the futurism in the series and was struck by how dated all concepts of AI were in the show. We can already produce natural-sounding human speech, so why is the Enterprise’s computer so mechanical 200 years in the future?

However – there is another explanation (this is obviously a retcon having nothing to do with the original intent of the creators). More recently it occurred to me that we might consider the mechanical nature of the computer voice a deliberate feature, not a bug of technological limitations. In the episode Tomorrow is Yesterday the ship’s computer had recently been overhauled on Cygnet XIV and was given a female personality with a more natural sounding (albeit humorously annoying) voice. We can therefore say that the technology existed and the mechanical voice was at least partly a deliberate choice. At this point you are probably thinking, “So what?”

Well, one of the features of LLM-based chatbots is that they are getting increasingly good at imitating human interactions through language, whether with text or speech. This creates the powerful illusion that they are sentient and feeling, when they clearly are not. I know this point can be debated – how can we be sure they are not sentient? I maintain they are simply not built to be sentient. As an example, the same technology can be used to make pictures, but no one argues that Midjourney is sentient. So for the sake of discussion let’s stipulate LLMs are not sentient. 

Why, then, program them to essentially pretend to be sentient? AI companies will argue that this makes them easier and more comfortable to use, but the cynical interpretation is – this makes our products more addictive to our customers. I get why they might want to perfect the technology of mimicking human language and speech for specific applications, but is this really optimal for everyday use?

Let me give you an example. I partly use various LLMs because I want to learn their capabilities and I want first-hand experience when I speak and write about this important technology. One thing I do is use LLMs as a partner in various creative projects. I know you can affect LLMs’ behavior by the prompt and other settings, but in the default mode the feedback it gives is not only sycophantic, it is gratuitously humanized. At one point, in response to my suggestion, the LLM replied, “I just got chills!”. Did it? Obviously not, so why did it say so?

I would argue that gratuitously humanizing an LLM’s interaction is inherently deceptive. And sometimes this deception works, in that it convinces people that a non-sentient AI is actually sentient. People form relationships with LLMs and form emotional attachments as if they were sentient. This is nothing short of a hack into human psychology – our brains are literally not equipped to distinguish things that are actually alive from things that simply act as if they are alive. We can intellectually understand the difference, but emotionally we respond to things acting alive as if they are. This is arguably a manifestation of AI companies making their products addictive rather than useful.

The fix, therefore, could be for AI companies to voluntarily, or for them to be influenced by legal precedent or regulations, essentially not do that – to make LLMs sound like non-sentient computers rather than people. In this way you would instantly know if you are talking to an AI or a person. You would not get sucked down an emotional rabbit hole of feelings for an AI. It would not be so tempting to assume a language model is sentient. Even 200 years in the future a computer voice might sound mechanical as a deliberate choice, even a matter of law, rather than a limitation of technology.

I doubt this is going to happen. That genie is already out of the bottle. But I increasingly feel as if it should happen, or at least I would like to see some AI companies start to follow this policy or at least offer it as an option, perhaps even the default option. At the very least I think that LLMs should not sound gratuitously human, should not pretend to have emotions that they do not have, and should adopt a “personality”, if you will, that is easily identifiable as an LLM. There is likely a happy medium. Computers do not need to adopt a Star Trek style mechanical voice, or use words like “affirmative” instead of “yes.” But I would like to see a serious attempt to develop a style that is comfortable yet clearly not human. Ditch the um’s and fake breath sounds. Ditch the fake emotions. LLM’s might need to adopt the equivalent of the mid-Atlantic accent popular in early TV and film – a sterilized affectation that is comfortable for people but does not try to mimic people.