For many, the age of artificial intelligence (AI) has only just become a conscious reality with the arrival of apps such as ChatGPT, DALL.E2, and character.ai. Any humble user of a website can now use these tools to create written, spoken, and graphic output that can be posted on social media, submitted for academic assessment, or published as the author’s own work in the public domain. These apps can even create the code for a website on which the content can be posted.

For some, however, these apps are perceived as a threat. With AI mimicking human conversation and responses so well, critics fear that with only one step further, these models can be customized so that they sound just like the original “author” (the “prompter”). Could, for example, the essence of a loved one be captured in code, so that they can be communed with after their earthly demise? Or could a maleficent hacker appropriate the audio, video, and speech habits of an individual and then deliver “fake news” indistinguishable (in digital form at least) from content created by the original person?

These fears are not without foundation. AI suppliers are already willing and able to “train” their large language models (LLMs) using customized inputs to interact with website users via those fickle little chat boxes that appear on the screen when browsing. The chatbot content is customized to the context of the website and is almost always personalized with a name and even a graphic, as behavioral psychology is rather unequivocal on the likelihood of interaction occurring being much greater when the bot has been anthropomorphized. Furthermore, with sufficient information gleaned from the web browser, IP address, and past engagement with the entity offering the bot, the personalization can be customized to match what the software perceives probabilistically are the user’s preferences.

While the website bots have so far largely confined themselves to written language exchanges, the potential to use the same sort of “tweaks” that take a bland technical answer and then convert it into the idiom of a young female academic or a retired gentleman farmer (conveyed by a persona most likely to elicit the desired interaction) have already been seen in “speaking bots” such as Alexa and Siri. Most AI apps offer a large selection of voices, allowing an individual to pick one and customize (“train”) it by adjusting pronunciation and adding inflections, intonations, and catchphrases. With sufficient training, these bots can produce surprisingly natural, human-like output that can only get better as repeated feedback is used to refine and improve it. Similar effects can also be captured and manipulated using video inputs. The potential can be observed in animations for Oscar-winning films such as Avatar: The Way of Water.

To be sure, the very customization and personalization enabled by AI have innumerable positive benefits. Search engines such as Google and Bing, social networks such as Facebook, and video services such as YouTube and Netflix are so good at delivering what individual users want because they have used a long history of individuals’ interactions to refine the content presented. Search is costly, and AI programs that filter content to save on search costs are highly valuable. Back in the days of the video rental store, I used to joke with my children that it took longer to pick their videos for the week than we took to watch them because faced with the entire array of just one store’s available content, they were quickly overwhelmed and found choosing very difficult. And frequently we got something we had already seen because no one could remember accurately. It is now so easy to pick something that I surely haven’t seen before that the video store nightmare is but a distant memory.

And for the time being, cost stands as an impediment to the widespread customization of at least audio and video representations of mere mortals. While generative, pre-trained written language transformers are very good at spitting out “original” content, customizing it to a specific written “voice” takes time and resources. Training it for millions or billions of users will take even more. And whereas we will pay for better selections by video services and search engines, it begs the question of who will pay, and what an acceptable price may be, for more bespoke outputs. As the cost of creating the audio and video equivalents is increasingly higher than the written language ones, the barrier to creating them is necessarily higher. If the price is right, then AI imitations will be created. But I suspect for most, while costs are high, creating a digital doppelganger is not a high priority. But when the costs come down, who knows?

The post Should We Be Worried About Bespoke Chatbots? appeared first on American Enterprise Institute – AEI.