Our Audience Feedback Survey Was Overrun By Bots. Here Are 5 Lessons We Learned.

Humanoid robots working with laptops and headsets Credit: Shutterstock
SciFri Findings is a sequence that explores how we perceive the impression of science journalism, media and programming on our audiences. Sign up for our e-newsletter to get the most recent experiences!

Getting suggestions from those that we serve has all the time been an integral a part of Science Friday’s goal in making science extra accessible. Audience analysis isn’t any totally different— I’m all the time curious to know what worth does the work we do present our audiences? Where can we make issues higher? How can we now have deeper engagement and impression? In June 2023 we launched an viewers survey throughout a number of platforms (radio, social media, newsletters, donors, and many others). This viewers survey was knowledgeable by in depth interviews on radio programming performed in Fall 2022. I excitedly waited because the survey went into the world, hoping 200-300 folks would care sufficient to finish it.
The pleasure and the shock because the numbers began trickling after which monsooning in— 1, then 100, 500, 2500, all the way in which as much as over 6800! Our Director of Audience, Ariel Zych, and I began enjoying with the information–we needed to make use of ChatGPT to theme and summarize among the qualitative knowledge. We copied some open responses into Chat and began noticing one thing odd right here and there within the knowledge we had been pasting over.
There, in our personal uncooked knowledge, had been a damning variety of clearly AI-generated responses, shamelessly self-disclosing with responses to open choice questions like “As an AI language mannequin, I would not have private preferences…” Scrolling extra by the information it grew to become clear…AI bots had struck our research, HARD! We checked out one another and couldn’t assist however chuckle on the irony. Here we share some classes realized (and a few we had forgotten) after we wiped away our tears and began cleanup.
Tips for Your Next Online Survey

Use survey software program that has a CAPTCHA: “Completely Automated Public Turing check to inform Computers and Humans Apart ” or CAPTCHAs are questions we now have all probably seen. These packages differentiate people from bot respondents. Many on-line software program firms present CAPTCHA choices for surveys however just for paid subscriptions.
No CAPTCHA? Trap ‘em: You might not have the funds for licensing survey instruments with a CAPTCHA characteristic. Trap questions are an alternative choice to a CAPTCHA that may assist present some protection in opposition to bots. They are used to establish respondents who will not be taking note of survey questions (e.g. somebody selecting “Strongly Agree” for all questions). A entice query can take many types, together with a query to establish an object in an embedded image, a immediate to sort particular phrases right into a textual content field, and many others. Once the information is collected, you possibly can filter out any respondents with incorrect solutions. Trap questions not solely shield in opposition to bots, but in addition unhealthy actors similar to trolls with an agenda, or those that don’t truly know in regards to the product or program however who need to obtain the money incentive. By offering a small variety of entice questions, you possibly can guarantee your goal audiences are those offering you with good knowledge, and remove the remainder.
Trap query used as a part of Science Friday radio programming survey. Credit: Nahima Ahmed/Science Friday
We integrated a entice query through the design part of our viewers survey. Participants had been requested to establish Science Friday’s host, Ira Flatow. Answer selections included solely different male science journalists and communicators so that each one choices could possibly be viable choices and restrict the variety of bots/unhealthy actors within the knowledge. We used one of these entice query as a result of we needed to survey present audiences who ought to know the host, not new audiences. This one step eradicated nearly 20% (N=1357) of our pattern!

More cash, extra bots, extra issues: Cash is king within the survey world. Participants are sometimes rewarded with money or present playing cards for every accomplished survey. Even the prospect for a lottery incentive has proven to extend response charges for on-line analysis. We selected to offer a $50 e-gift card lottery incentive to stability the size of the survey and encourage extra viewers members to finish it. Money is nice, however with extra money comes larger incentives for bot creators, unhealthy actors, and trolls to take part for money alone. We rapidly realized $50 was so much to supply for a ~12-13 min survey. It made me suppose: How can I make sure that to worth the individuals time whereas nonetheless ensuring I get the knowledge I want? Next time, we’ll contemplate decreasing the edge of our money incentive. Perhaps it might have been restricted to $25 as a substitute? If this didn’t yield sufficient individuals, possibly a second recruitment waive can be so as? In the long run, notably for viewers surveys, we’d contemplate providing different issues of worth, similar to merchandise or free occasion tickets as a substitute. Non-cash provides may scale back the variety of folks curious about simply being paid for survey completion. It may also present worth by giving individuals tangible supplies and/or deeper engagement along with your group.
Segment audiences: Whenever possible, use totally different utm or referral hyperlinks for various recruitment pathways to your surveys. We used totally different hyperlinks for every platform (i.e. Twitter, (*5*), Donors) to know the place site visitors was coming from, search for variations within the preferences between audiences, and to seize the potential universe measurement for our pattern. We had greater than half our respondents come from Facebook, which is disproportionately larger than we normally see for surveys. Generally, we discover our radio audiences to be the biggest referrals so seeing so many come from Facebook was a pink flag. Additionally, segmenting audiences can establish any unusual patterns within the knowledge. For instance, when you’ve got beforehand surveyed audiences, it’s possible you’ll have already got demographic knowledge to test in opposition to new knowledge. If your group primarily serves older adults, and see that your survey consists of solely younger individuals the information could also be compromised. Consider whether or not it could possibly be the subject of the survey, recruitment, or if this anomaly is a possible bot.

Cleaning Up The Data
After a couple of laughs and tears, I had the duty of determining how precisely to scrub up the jumbled mess of information we had. With a filtered dataset (thanks entice query!), I began cleansing the information utilizing

Impossible timestamps: Responses submitted inside the identical second of one another had been eliminated. Many of probably the most suspicious responses had been submitted with practically the identical time stamp late at night time (12-3 am) or early morning (4-7 am) that are unlikely instances for our US-based audiences to finish surveys.
Obvious AI language: I had a variety of open-ended questions for the survey. Any responses that had very apparent language (“As an AI language mannequin, I would not have private preferences…”) had been eliminated.
Non human sounding responses: Some of our open-ended questions included asking why individuals most popular sure broadcast codecs. We eradicated any responses that didn’t sound genuine to an viewers voice. For instance, “Live name can improve the viewers’s sense of participation and loyalty…” It is uncertain that an viewers member can be discussing loyalty.
Human-sounding, however similar, open responses: There had been some responses that repeated usually. This consists of phrases like “It can create memorable moments for each the host and the viewers” and “Maintained the authenticity of this system”. It was extremely unlikely that a number of particular person respondents used the very same phrasing.

Designing viewers centered content material is an inherently inclusive course of. Audience surveys are a possibility to take heed to the wants and considerations of our audiences. Surveys are only one instrument we use to assist collect viewers suggestions at Science Friday. When all of the cleansing was mentioned and finished, we had been nonetheless left with 1200+ survey individuals in our pattern! This was considerably larger than the 200-300 we initially anticipated. As on-line analysis continues to develop, so does the potential for AI bots. I’m appreciative of getting found new methods to enhance my follow even when it value me hours of labor and a few new grey hairs.
Your voices have formed our present. From left to proper, a younger viewers member asks a query at SciFri Live in San Francisco, Ira stands on stage in Salt Lake City, one other younger listener asks a query at SciFri Live in San Antonio. Credit: Alexander Lim/Benjamin Altenes/Cindy Kelleher/Science Friday

Meet the Writer

Nahima Ahmed
About Nahima Ahmed


Nahima Ahmed is Science Friday’s Manager of Impact Strategy. She is a researcher who likes to cook dinner curry, focus on identification, and helps the staff perceive how tales can form audiences’ entry to and curiosity in science.


Recommended For You