This Generative AI Prompting Technique Uses Multiple Expert Personas To Derive First-Class Answers

In today’s column, I showcase a valuable prompting technique that uses multiple expert personas to push generative AI and large language models (LLMs) toward first-class or top-notch answers. The approach is straightforward. You invoke the intrinsic capabilities of AI LLMs to undertake persona simulations, and cleverly combine this facility with a focus on the personas as experts in whatever field of inquiry you are interested in. Various crucial upsides and downsides need to be observed.

Let’s talk about it.

This analysis of an innovative AI breakthrough is part of my ongoing Forbes column coverage on the latest in AI including identifying and explaining various impactful AI complexities (see the link here).

Prompt Engineering And Personas

In a previous posting I explored over fifty prompt engineering techniques and methods, see the link here. Among those myriad approaches was the use of personas, including individual personas and multiple personas, as depicted at the link here, and the much larger scale mega-personas at the link here. Personas are a powerful feature available in LLMs, yet few users seem to be familiar with the circumstances under which they should consider invoking the capability.

A persona consists of telling generative AI to pretend to be someone and simulate what that person might know or say. No special setup or elaborate instructions are needed to get this undertaken. I will be providing you in a moment with prompts that show how this is readily performed.

A quick example of the value of personas might be helpful here.

One of the most popularly invoked personas entails generative AI pretending to be Abraham Lincoln. A teacher might tell a generative AI app such as ChatGPT to simulate the nature of Honest Abe. Students then access the pretense or persona. In an amazing flair, the AI seemingly responds as we assume Lincoln might have responded.

Of course, this is based simply on the numerous speeches, written materials, and other collected writings that suggest what he was like. The AI has pattern-matched computationally on those works and mimics what Lincoln’s tone and remarks might be. Anyone using a generative AI persona needs to keep their wits about them and realize that the conversation or engagement is nothing more than a mimicry or imitation. Though the AI appears to often convincingly fake the nature of the person, it is all still a computational simulation.

Generative AI And Subject Matter Expertise

A clever way to lean into personas consists of telling generative AI to pretend to be an expert in some field of study.

Imagine that you are interested in climate science as a specific domain. You could tell generative AI to pretend to be an expert in climate science. Notice that you don’t have to name a particular person, such as in the case of naming Abraham Lincoln as the subject of the persona. A persona can be unnamed. If the persona is not based on a known person, you’ll need to be mindful and state what the persona is supposed to be modeled on.

I might tell generative AI to pretend to be a climate science expert. Voila, that alone would be sufficient to get the AI to simulate that specialty persona. One issue to keep in mind is whether the generative AI that you are using has sufficient data and has sufficiently patterned on that data to adequately represent the field of interest. Be cautious since the AI might be quite shallow and yet portray the expertise as though it is in-depth.

For whatever kind of expertise you want to base a persona on, a vital question is whether the AI has been data-trained extensively on that topic. Most of the major generative AI apps are broadly trained across a wide swath of the Internet. This tends to make the AI so-called mile-long and an inch deep in terms of expertise.

Do not assume that some desired expertise is always at the ready.

It is usually wise to ask the AI what level of proficiency it seems to have in whatever topic you are exploring. This doesn’t guarantee that the AI will be forthright, and you might get either a bland answer or a false or misleading response. If possible, consider asking direct questions on the topic and have a basis to compare the answers, such as doing online searches or having in-hand materials to corroborate what the AI opines on the topic.

Another common means of boosting AI in a field of interest would be to feed or import content on the topic directly into the generative AI. The use of retrieval-augmented generation (RAG) and in-context modeling can aid the AI in moving up the ladder in terms of expertise on a specific topic, see my discussion at the link here.

Invoking Multiple Expert Personas

We are now then ready to combine the notion of personas, expertise, and the idea of using multiple instances.

If you invoke just one persona, let’s say it is an expert in climate science, there is a strong chance that the generative AI will give a particular bent to that persona. It might be that the persona is mainly shaped around the atmospheric and weather facets of climate science. That’s fine, except that you could end up getting responses that are confined to that arena. The persona might skip considering other areas such as the economics of climate science.

Aha, this brings us to the notion of invoking multiple expert personas, all at once.

You can pretend that a room full of experts is being convened. The instructions to the AI are that multiple expert personas are to be defined and used simultaneously. You can either let the AI choose what those personas will consist of, or you can shape the direction of each persona. This depends on what you are trying to accomplish with the simulation.

I will next walk you through some examples to highlight the tradeoffs of how this is accomplished.

Getting Underway With Multiple Expert Personas

I went ahead and logged into ChatGPT by OpenAI to come up with an example for you. You might find it of keen interest that ChatGPT garners a whopping 300 million weekly active users. That is a staggering amount of usage.

I briefly conducted an additional cursory analysis via other major generative AI apps, such as Anthropic Claude, Google Gemini, Microsoft Copilot, and Meta Llama, and found their answers to be about the same as that of ChatGPT. I’ll focus on ChatGPT but note that the other AI apps generated roughly similar responses.

Let’s get underway.

• My entered prompt: “I want you to pretend to be multiple experts. I will tell you what field of expertise they have. I will also tell you how many experts there are. Your job will be to then answer my associated questions by pretending to be those experts. Do you understand these instructions?”
• Generative AI response: “Yes, the instructions are very clear. I will wait for you to tell me what the field of expertise is, and how many experts there are to be involved. I will then answer your question based on those personas. Let me know when you are ready to proceed.”

Take a close look at the prompt that I entered.

You can see how simple it is to get multiple personas on the table. I laid out for the AI that I want the AI to pretend to be multiple experts and that I will say what the area of expertise is. Furthermore, I clarified that the purpose for doing this is so that I can ask questions of the simulated experts and get (hopefully) suitably informed responses accordingly.

That is a handy-dandy templated way to get the matter underway and you are welcome to use that same prompt in your endeavors.

Taking The Experts For A Test Drive

Let’s try out the multiple experts by stipulating a field of interest. I went with climate science for this example. In addition, I will pose a pressing question for the experts.

• My entered prompt: “I want you to pretend to be three experts in the field of climate science. My question is this: What is the most pressing issue caused by climate change?”
• Generative AI response: “To confirm your request, I will establish three personas that are experts in climate science. I will refer to them as: (1) Dr. Green, an ecologist specializing in biodiversity, (2) Dr. Blue, an atmospheric scientist focusing on weather patterns, and (3) Dr. Brown, an economist studying the impact of climate change on global markets. They are to answer the question from each expert’s perspective about what the most pressing issue is caused by climate change. Should I proceed?”

You can see that I opted to explore the field of climate science.

In this case, before I invoked the personas, I had in the past asked numerous questions about climate science and knew that the generative AI had done a semblance of pattern-matching on climate science topics when it was initially data-trained. I’m not suggesting that the AI is truly an expert and only mentioning that I did some pre-work to make sure that at least a minimum amount of data was scanned and patterned on in general.

Observe what the AI did in response to my request.

I had stated that I wanted three expert personas in climate science. Sure enough, the AI echoed to me that it would simulate three instances. I hadn’t said what their specific detailed expertise might each be. The AI opted to go with three experts, each having a different subspeciality, consisting of an ecologist, an atmospheric scientist, and an economist.

If I didn’t want the personas to be on those subtopics, I could easily enter another prompt and tell the AI that the expertise choices should be changed. The AI helpfully opted to give names to the three expert personas (Dr. Green, Dr. Blue, Dr. Brown), making it easier for me to refer to each one specifically (the naming seems reminiscent of the now-classic movie Reservoir Dogs if you know what I mean).

Getting An Answer From The Expert Personas

I will go ahead with the three expert personas and let the AI derive an answer to my posed question.

Here we go.

• My entered prompt: Yes, please proceed, but be brief in their respective responses.”
• Generative AI response: “According to Dr. Green (ecologist), the most pressing issue caused by climate change is the rapid loss of biodiversity. Dr. Blue (atmospheric scientist) indicates that the most urgent issue is the intensification of extreme weather events. Meanwhile, Dr. Brown (economist) says that the biggest concern is the economic fallout of climate change. Would you like me to expand on any of those views?”
• My entered prompt: “Provide more detail by Dr. Brown on the economic ramifications.”
• Generative AI response: “Dr. Brown indicates that disruptions to agriculture, increased costs of disaster recovery, and forced migration due to rising sea levels will strain global markets. Developing nations are disproportionately affected, as they often lack the resources to adapt, leading to a widening wealth gap and global instability.“

I told the AI to be brief in the AI responses.

It is up to you whether you want to get short answers or long answers. I usually prefer to start with short responses and then see whether the AI is on target. The problem with getting long answers at the get-go is that if you are paying for the use of AI, you might needlessly be racking up online costly processing cycles. I like to dip my toe in and then decide whether to go expansively.

After getting a short answer from each of the three expert personas, as combined into one response, I decided to see what else Dr. Brown, the economist persona, might have to say on the topic. The AI responded accordingly.

Big Question About Consolidating Experts

One of the toughest aspects of using multiple expert personas entails how to end up with a final answer. The simplest approach involves the AI merely stating what each expert persona had to say. This can be combined into one “final” response.

For example, I asked to see a cohesive final answer.

• My entered prompt: “I want you to combine the three answers into one cohesive final answer.”
• Generative AI response: „The most pressing issue caused by climate change is its multifaceted impact. From an ecological standpoint, the rapid loss of biodiversity due to habitat destruction and shifting climates destabilizes ecosystems that provide essential services like pollination and carbon sequestration (Dr. Green). This ecological imbalance feeds into atmospheric challenges, as these weakened ecosystems struggle to regulate weather and climate, contributing to more frequent and severe extreme weather events such as hurricanes, floods, and droughts (Dr. Blue). Disrupted agricultural systems, soaring disaster recovery costs and forced migration amplify global inequality and strain financial resources, particularly in developing nations that lack the means to adapt effectively (Dr. Brown).”

The problem is that the presumed cohesive final answer by the AI is still disparate answers that happen to be lumped together.

Without anthropomorphizing AI, let’s shift gears and consider what happens with human experts. If you get human experts together and ask them to answer a tough question, the odds are they will each have a particular opinion. You won’t necessarily get just one answer, unless the question at hand is something that lends itself to solely one answer, such as perhaps a numeric-oriented question that involves doing calculations and arriving at a single value.

There is an entire body of literature on how to get human experts to coalesce toward one unified answer on a given topic. You can have them take a vote. They can be given metrics or boundaries that essentially force them into an agreed-upon answer. And so on.

The same somewhat applies to the invoking of generative AI multiple expert personas, in the sense that you’ll need to tell the AI how it is to combine the disparate answers. You can take the route of specifically stipulating what the method should be. An easier approach is to force the AI to decide how to do so.

In this case, as shown next, I went with a blending approach of allowing the AI to mix the expert personas responses as though the answer was derived from a single source.

• My entered prompt: “You are still providing three separate recommendations. Try to blend them into one overall unified recommendation.”
• Generative AI response: „The unified recommendation is to prioritize large-scale investments in nature-based solutions, such as reforestation, wetland restoration, and sustainable land management. These efforts address biodiversity loss by protecting habitats and fostering ecosystems that support life, while simultaneously mitigating extreme weather impacts by stabilizing climates and reducing the frequency and intensity of disasters. Additionally, such investments can drive economic growth by creating green jobs, enhancing agricultural productivity, and lowering the long-term costs of climate adaptation and recovery.”

Dealing With Generative AI Myopia

A disconcerting caveat of using generative AI to simulate multiple expert personas is that the AI is dipping into the same data set and pattern-matched data pool for each of the simulated personas.

Here’s why that raises the eyebrows. You probably assume that each persona is independent of the other ones. Unfortunately, that’s not true. The AI is shaped around a large-scale data structure and is based on large-scale data training, all of which ends up as one gigantic pool. Each persona is a slice of that same pool. The gist is that the personas aren’t going to be independent, and they are bound to lean in similar directions. Any biases are likely to be carried into each of the personas.

You can tell the AI to not allow any such bias to arise, but that’s still problematic and you are going to get a likely false or misleading indication that the AI has abided by your request. There are prompting techniques that can help to try and somewhat soften the blow of those biases, see my description at the link here.

A more compelling approach involves using more than one generative AI.

The chances are that a different generative AI will have a different semblance of the pattern-matching and the data was used during the initial setup. As such, you are potentially going to get somewhat different answers. I say that this is only somewhat because most of the AI makers are nonetheless scanning much of the same parts of the Internet, see how this has led to a shared indication that has some disturbing consequences, per my discussion at the link here.

Using Additional Generative AI Apps

Based on the noteworthy qualms that my three expert personas are only within the confines of the one generative AI app that I was using, I opted to log into a different generative AI app and ask the same question that I had posed earlier.

I then brought that answer into my underway online conversation with Dr. Green, Dr. Blue, and Dr. Brown, cheekily adding the answer as though it came from Dr. Purple.

Here’s what happened.

• My entered prompt: “I went ahead and asked a different generative AI to provide one expert opinion or persona on the same topic. I will refer to this persona as Dr. Purple. Here is what the AI said: ‘Focus on accelerating the transition to renewable energy systems, such as solar, wind, and geothermal power, while simultaneously phasing out fossil fuels.’ I want you to now include that fourth opinion and tell me how it impacts the three experts you are simulating.”
• Generative AI response: “The approach of Dr. Purple addresses biodiversity loss indirectly by reducing habitat destruction linked to fossil fuel extraction and mitigating carbon emissions that drive climate instability (Dr. Green). It curbs extreme weather events over time (Dr. Blue) by slowing the progression of global warming. Economically, it promotes energy security, reduces long-term energy costs, and creates millions of jobs in the clean energy sector (Dr. Brown). By targeting the root cause of climate change—greenhouse gas emissions—this recommendation provides a scalable solution with long-term benefits for ecosystems, weather stability, and global economies. This perspective shifts focus to energy systems, providing an actionable and complementary strategy to the previous recommendations.”

You can see that the generative AI sought to intertwine the answers of all four expert personas.

Having Many Experts Has Its Ups And Downs

Congratulations, you are now ready to try out the use of multiple expert personas.

I’d suggest that you pick a topic that you know something about and start with that as your means of experimenting with multiple expert personas. This allows you to readily gauge whether the responses are making sense. It will prepare you for coping with subject areas that you aren’t familiar with and when opting to involve multiple expert personas.

During my classes on prompt engineering, I place the use of multiple expert personas into the advanced portion of the course. The technique is something you won’t necessarily use constantly. The aim would typically be when you are examining a topic that you’ve not previously considered and for which you find useful seeing multiple perspectives. One venue in which this approach especially shines is in an educational setting, such as having multiple experts debate a legal issue such as a portion of the constitution. It is almost like having a crew of legal beagles pontificating on a piece of the law.

I’ll close this discussion on multiple expert personas with a few enlightening quotes from some experts.

Franklin D. Roosevelt famously said this: “There are as many opinions as there are experts.” This pertains to a comment that I mentioned earlier. Unless the question posed has a singular final answer, the likelihood is that your multiple expert personas are going to provide a multitude of viewpoints and not especially land on a single end-all response. Please keep that in mind.

Finally, the great Thomas Jefferson made this illuminating remark: “Difference of opinion leads to inquiry and inquiry to the truth.” May your use of generative AI entail abundant and probing inquiries, and may those inquiries, in turn, lead to the truth.

That’s mindful sage advice in this modern-day era of advanced AI.