2023 Application Cycle Round-Up
The expected, the unexpected, and the AI-generated
2023 marks the seventh cohort of the Civic Digital Fellowship and the third cohort of the Civic Innovation Corps at Coding it Forward. Over the past six years, we’ve seen year-over-year growth of our fellowship cohorts, as well as the expansive reach of our impact across all levels of government in the United States. As we continue to grow, the core of our mission remains the same: creating new pathways into public interest technology for early-career technologists. As we wrap up the application process for our 2023 cohort, we’re reflecting on this application cycle — one that is both familiar and unfamiliar in different ways.
Application Cycle
This year, we received 2,500 applications — another record-breaking year for our summer fellowships. The continued enthusiasm for our opportunities also meant that this year was our most competitive cycle — with only 4% of applicants placed as Fellows in government host offices. As much as it pains us to turn away well-qualified candidates at every turn of the process, we are truly grateful for the opportunity to learn about our candidates and the diversity of their backgrounds, experiences, talents, and dedication to public service they bring to the table. Our team is committed to expanding our programs so that every early-career technologist has an opportunity to serve with their skills.
We also want to acknowledge our dedicated alumni reviewers. Over 50 members of our alumni community volunteered their time and energy reviewing applications with our team, contributing over 1,000 hours to ensure every candidate was considered. Without their support, we wouldn’t have been able to move our process forward with such an unprecedented volume.
A New Assessment
Our application process was also revamped this year. We replaced the behavioral interview with our team with a take-home assessment. We’ve found that live interviews may not be the most effective and equitable avenue for us to capture the qualities and experiences we look for in our Fellows. One-on-one interviews also introduce variability and bias, skewing our pool based on each reviewer’s interview style and evaluation. They’re also time-consuming, with our team spending hundreds of hours interviewing, resulting in a drawn-out application process.
With the expertise of close to 20 alumni, we crafted and refined take-home assessments that resemble what our Fellows typically execute at the beginning of their fellowship — a design and planning process for their summer project. We extracted common scenarios from past projects completed by our Fellows across our four areas of practice — data, design, product, and software engineering — and identified specifications related to each sub-role type (e.g., data science vs. data engineering). Intending to capture each candidate’s communication skills and ability to apply their technical expertise, the assessment also served as a sneak peek into the project experience many of them will soon enter. It also allowed us to evaluate each candidate’s thought process more consistently and conservatively.
ChatGPT and AI-Generated Content
While we anticipated and prepared for new challenges with our newly launched assessment process, the simultaneous advent of ChatGPT posed a new challenge for our team. A week into reading applications, our alumni reviewers noticed essays that contained identical phrases and sentence structures while lacking personal anecdotes. The answers were so generic that they barely addressed our simple application question:
Why do you want to participate in Coding it Forward’s summer fellowship programs? Please describe how this program would help you achieve your long-term professional goals.
We knew that ChatGPT was skyrocketing across many industries and started to wonder — did people use it to write their applications for our programs?
The intuition of our alumni reviewers was correct. Simply copy-and-pasting a few of these essays into the open-source GPT-2 detector by HuggingFace immediately flagged them as “99.99% Fake,” meaning that AI-generated content was detected. As we reviewed more submissions, many more were flagged by our reviewers and subsequently confirmed by this GPT-2 detector. We found ourselves at a crossroads — we want our Fellows to be technical innovators up-to-date with the latest tools. Yet, these essays didn’t represent the candidates in an honest or personable manner — two qualities that are equally important to succeed in public interest technology.
We quickly realized this was a recurring issue and began systematically investigating all submissions. At first, we tried implementing the HuggingFace detector (a pre-trained RoBERTa model) on our own. Unfortunately, its accuracy was subpar, given it’s trained on a previous version of the GPT corpus (GPT-2 instead of GPT-3.5 used in ChatGPT; each version of GPT is improved and trained on an exponentially larger corpus of text than its predecessor, with the newly release GPT-4 now available).
Around the same time, we were fortunate to have come across a new detection model called GPTZero, developed by Edward Tian and his team at Princeton University. We became one of the pilot users of their API service, using their proprietary algorithm to detect AI-generated content. On a very high level, GPTZero helps identify words and phrases written in such a “random” way that an actual writer would never have written them, resulting in a low “perplexity” score (i.e., formulaic and random). This detection scheme showed very promising results from early testing by our team.
Findings
What did we find? A non-negligible number of submissions contained one or more responses with a low perplexity score. Two of the six written components in our application process required a personal touch — needing elaboration on past experiences, future goals, and detailed thought processes. AI-generated responses to these questions lacked nuance and prevented our reviewers from learning about a candidate.
Further analysis helped us pinpoint the systematic differences between genuine, human-written, and AI-generated essays. A brief frequency analysis of the words used in our essays revealed some interesting patterns. In genuine essay responses, we see candidates talk more about their desire to learn, their willingness to contribute to public service, and their goal of giving back to the community. We see words related to aspirations grounded in reality. We see their understanding of the prompt that leads to discussions on requirements, challenges, and next steps. Moreover, many of these essays share personal stories that informed their decision to enter public service.
In contrast, AI-generated essays repeated or rephrased our application prompts, sometimes verbatim. They use buzzwords without elaborations in depth: they mention the phrase “long-term goals” without specifying the actual goals, the phrase “real-world problem in the society” without specifying the actual problem, and the phrase “grow personally and professionally” without specifying how. Many responses were paraphrased directly from our website and info sessions. It is also no surprise that the words like “residents,” “vaccines,” “app,” “booster,” or “city services” were frequently used because they were part of the prompt as well.
Our analysis demonstrated that AI-generated content is grammatically coherent but often lacks personality and details. AI-generated content represents what we prompt it to write rather than our personality and background, which would be reflected in our personal prose.
Takeaways
The 2023 application cycle was rigorous but incredibly rewarding for our team. Although we didn’t expect to spend so much time reflecting on the importance of the written word in a process designed to review applicants for software, data, design, and product fellowships, this experience reminded our team of the reason we decided to revamp our application process in the first place: to better understand how our applicants approach problems, ask questions, and communicate their ideas. These qualities are all inherently human, and their importance is more evident than ever for our team.
As we wrap up our application process for 2023, we remain inspired by our applicants and their continued interest in using their skills to deliver policy, improve systems, and strengthen products on behalf of the American people. Stay tuned as we welcome our latest cohort of Civic Digital Fellows and Civic Innovation Corps Fellows this summer!
