I tried out SyntheticUsers, so you don't have to

Using AI as a replacement for interviewing actual users is a brilliant idea if you want to look like you made an effort, but are really looking to fill the page with superficial, stereotyped info.

“Syntheticusers.com uses advanced AI and natural language processing to generate synthetic personas that can mimic real human behavior.”

I decided to test it out. First an overview of a recent study we conducted with 14 real humans in Oakland:

We worked with the Oakland Unified School District to understand what low income and immigrant parents need in the process of finding a school for their child. The paper titled “Not Another School Resource Map: Meeting Underserved Families’ Information Needs Requires Trusting Relationships and Personalized Care” was published in the ACM CSCW 2022 conference and received a recognition for contribution to Diversity and Inclusion.

Initially, we set out to study how the “Oakland school finder” dashboard can be made more accessible and usable for our target population. In the course of our work though, we learned that there was a mismatch between the information families need and the information that is readily available online.

We found that the parents in our sample were seeking a high quality education for their child, but faced challenges in reaching this goal, including finding useful and relevant information about the schools available, and making choices between schools that balanced access to well-resourced schools against other considerations like safety, inclusion, and convenience. Also, families face an overwhelming number of options, which exacerbates the challenge of finding relevant and helpful information. We found that a number of “parent advocates” — people who work at the school district, individual schools, or local non-profits — support parents during the enrollment process and play a big role in filling this information gap. Parent advocates provided relevant and concise information to families by first building a trusting relationship with them, learning about their personal circumstances and priorities, then connecting them with the right information or resources. There are over 100 schools in Oakland, so much of this work involved reducing information overload for families. In addition, this support often extended beyond enrollment to address other survival needs like housing, food, and healthcare.

So, overall, we found that the school dashboard was not providing the information that the participants needed, or in the words of one participant:

“it’s geared towards parents who are looking for certain things. It’s geared toward a middle-class, upper middle-class aesthetic”

And instead, parent advocates’ strategies focused on personal relationship building and trust and highlighted the importance of attending to individual families’ circumstances and priorities.

---

So… let’s see what SyntheticUsers would say about this problem. You can see the full interview transcripts here.

SyntheticUsers first asked me to define my users.

“Getting Started? We can help!

This is where you can add your audiences. Get specific, for instance: 34 year old worker in the fintech industry, married, with two dogs and a love of rugby.”

So I typed in: Low income, immigrant parent in Oakland who does not speak English.

Next, it asked me to define my problem: How do you find out information about which school to send your child to?

Next, it asked for a proposed solution to the problem: A dashboard that provides information about school locations, test scores, attendance statistics.

After clicking submit, a loading screen popped up that displayed the following sentences, one after the other:

Identifying potential interviews and checking availability … Scheduling interview … Conducting interview, taking detailed notes … thanking participants

Let’s set aside for a minute the ridiculousness of pretending that actual interviews are being scheduled and analyzed (sigh), I received back a report with 6 interviews. Of these 6, 4 said that they lived in Oakland, one lived in San Pedro Sula, Honduras, one in Karachi, Pakistan, and one in Tehran, Iran. Let’s also set aside the factual inaccuracies and focus on the substance of the interviews. I’ve summarized the main findings from the SyntheticUser “interviews” below:

• Participants said that the dashboard would be “very effective,” “very useful,” and “very helpful” in solving their problems (6/6) and rated it on average 3.4/5.

• Challenges to the dashboard idea raised by participants are:

  • limited internet access

  • not enough digital literacy

  • support for languages other than English

  • keeping information reliable and up to date

  • providing information beyond test scores such as extra curricular and safety records

---

Our own in-depth interviews over months with real parents found that the dashboard was a misguided solution and that what was really meeting our participants' information needs was through trusting relationships that met them where they were and went beyond information about schools. On the other hand, the main results from SyntheticUsers “interviews” are exactly what should be expected from a pattern synthesis engine¹: stereotypically likely challenges for low income immigrant families that lack substance about their actual real lives. This is not a challenge that might be improved with more fine-tuning or better models. A pattern synthesis engine is fundamentally unaware of anything in the real world, and can not produce any "insights" beyond whatever patterns exist in its training data. Priming it by telling it to act as though it's someone else does not change this basic fact.

Finally, apart from the fact that misguiding users by acting as though a pattern synthesis engine can actually be “interviewed” is wrong, I’m also not sure why it’s even remotely useful. The whole point of spending the time to interview people and then spending a lot more time analyzing the large amounts of data gathered, is the ability to connect with them, build trust, dig deeper, ask them to share stories, and learn about their feelings and emotions. Pattern synthesis engines have none of those. They can generate a lot of text that make it look like they do, but there’s no substance there. In fact, basically the same information can be extracted from ChatGPT by simply asking it the question directly:

• Imagine a low income, immigrant parent in Oakland who does not speak English. They are looking for information about which school to send their child to. A potential solution to this problem is a dashboard that provides information about school locations, test scores, attendance statistics. Create a list of challenges that the dashboard idea may face.

• ChatGPT response (summarized here, full response below²):

  1. Language barriers: Since the target audience does not speak English, the dashboard would need to be available in multiple languages, which could increase the development cost and maintenance of the platform.

  2. Technology Access.

  3. Understanding of the Data: The information provided in the dashboard may be overwhelming or confusing for parents who are not familiar with the school system in Oakland. Providing some guidance on how to interpret the data could help overcome this challenge.

So my main takeaways are, do not do this: it does not provide any substance about actual people’s lives beyond what is statistically most likely (see stereotyping), is wasting energy and producing unnecessary carbon, and is unethically misleading people to think that pattern synthesis engines can be scheduled for an interview and even thanked after.

1. “pattern synthesis engine” is a much more accurate term to describe the AI models we see today such as LLMs. I first read it here.

2. There are several challenges that a dashboard designed to provide information about schools to low-income, immigrant parents who do not speak English in Oakland may face. Some of these challenges include:

1. Language barriers: Since the target audience does not speak English, the dashboard would need to be available in multiple languages, which could increase the development cost and maintenance of the platform. Moreover, it may require professional translations to ensure that the information provided is accurate.

2. Technology Access: The target audience may not have access to technology or the internet. This could make it challenging for them to access the information provided on the dashboard, making the dashboard less effective.

3. Understanding of the Data: The information provided in the dashboard may be overwhelming or confusing for parents who are not familiar with the school system in Oakland. Providing some guidance on how to interpret the data could help overcome this challenge.

4. Data Accuracy: The accuracy of the data provided in the dashboard could be questioned, which may reduce the trust of parents in the system.

5. Overcrowding: Some schools in Oakland may be overcrowded, and the dashboard may not provide any information on this aspect. This could lead to parents enrolling their children in schools that cannot accommodate them, resulting in a waste of time and resources.

6. School ratings: School ratings based solely on standardized test scores can be misleading and may not provide a complete picture of a school's quality.

7. Availability of information: The information available on the dashboard may not be updated or may be incomplete, making it difficult for parents to make informed decisions.

8. Social Stigma: Some parents may feel hesitant to use the dashboard due to social stigma associated with receiving help or accessing public resources.

Addressing these challenges requires careful consideration of the user needs, adequate resources, and collaboration with community organizations to ensure that the dashboard is effective and accessible to the target audience.