Ad agencies embrace AI hallucination with the ‘Insights Machine’

But we’ve also learned, sometimes painfully, that there are some things LLMs can’t do so well – not yet, anyhow. For one thing, they’re not great at sticking to a strict code of truthfulness; they often ’hallucinate’ or deliver factually inaccurate information that’s veiled as truth. They also can’t tell us much about the internal, subjective human experience, despite the claims of people like Blake Lemoine – the former Google engineer who was fired after he gushed to the press about the supposed sentience of the company’s LaMDA AI model – most experts would roll their eyes at the suggestion that AI is self-aware.

Thankfully, most mainstream LLMs aren’t allowed to pretend they’re a person; ask ChatGPT if it’s capable of feeling anger, and it will politely tell you that it’s a machine and therefore incapable of experiencing emotion. Was there something else it could help you with?

But what if AI could emulate human personalities well enough to effectively give marketers a direct line of communication with their target audiences? What if, in other words, marketers had a focus group of AI personalities that was constantly at their disposal, through which they could test ideas and glean inspiration?

It was questions like these that sparked the Insights Machine, a new platform from marketing agencies Pereira O’Dell and Plan.Net and which allows brands to create their own AI ’personas’ based on the demographics of key audiences and then interact directly with these personas as if they were interacting with flesh and blood human beings.

Unveiled yesterday, the Insights Machine is accessible to all and comes with four pre-built personas: Christian, Martina, Javier and Katja. Users can ask any persona a question through the form of text prompts, just as they would while using ChatGPT. (The platform, in fact, uses OpenAI’s API.) Each persona’s answers will conform with the short bio that appears below their name. For example, when asked about his happiest memory, Javier, a 28-year-old “Road Warrior” from Denver, recalls the first time he flew off to another city for work. Katja, a mother-of-two “Family Adventurer” from St. Louis, answers that same question by briefly describing a family trip to Paris.

Of course, neither of those events could have ever taken place – no algorithm has ever literally boarded a plane. Each of the personas is designed to play the role of an actual human being; in this way, they differ sharply from other AI chatbots, which, again, are often explicitly programmed to refrain from precisely that kind of behavior.

Insight Machine users can also ask the same question to multiple personas simultaneously and review their answers side-by-side, thus theoretically gaining a better idea of the differences in ’opinion’ between these simulacra of their target audiences.

As its name suggests, the goal of the Insights Machine is not to provide completely factually accurate content but rather to challenge the marketers who are using it to think in new and creative ways.

“This is a machine that will spark insights in our own heads,” PJ Pereira, founder and chief creative officer of Pereira O’Dell, told The Drum in an interview. The idea is not to get to know Javier, the AI chatbot, at a deep level – it’s to gain a clearer sense of what a fictitious but still reasonably realistic late-20s-something male who travels a lot for work might be prioritizing in his day-to-day life (’might’ being the operative word there).

For example, Pereira recalls an occasion during the Insights Machine’s testing phase in which the persona of a woman who owned a dog suddenly changed all of her views on travel once the dog was removed from her persona.

“Holy cow,” Pereira remembers thinking, “If we were just thinking about traveling as a category … we would not have considered pets to be an important part of a persona until we saw that [use] case.” The dog, which was apparently not previously thought to be especially relevant to queries about travel, suddenly became salient.

While hallucination is typically discussed throughout the AI industry as a bug and not a feature – something around which reliable safeguards need to be constructed – Pereira views the phenomenon a bit differently – as a quirk in AI that can benefit human interlocutors. “Hallucinations are important [to] this process,” he says, referring to the use of the Insights Machine. “The fact that it’s unpredictable adds another layer that gets us thinking … that’s where the inspiration comes from.”

Pereira also emphasizes that it’s ultimately up to the marketers using the platform to determine which of the personas’ responses are useful and which ones should be ignored. “The tool is for inspiration,” he says. “The responsibility belongs to a human.”

For more on the latest happenings in AI, web3 and other cutting-edge technologies, sign up for The Emerging Tech Briefing newsletter.