Joined July 2020
- Tweets 1,073
- Following 632
- Followers 2,571
- Likes 505
595 Photos and videos
Pinned Tweet
Apr 23
Americans disagree on a lot. AI regulation isn't one of them.
In our latest report deep dive, Tyler Bench and Ben Leff unpack what bipartisan agreement on AI means for candidates, and why we're likely just scratching the surface of what AI regulation will look like.
Full video and report linked in comments.
1
1
2
449
Jun 11
Go behind the scenes with the Verasight team at
@AAPOR 2026.
Jake Rothschild presented Verasight's contribution to The Voter Poll by SSRS, and Ben Leff presented our latest synthetic sampling research.
The team shares what drew them to high quality survey data and what researchers are asking about on the exhibit hall floor.
1
258
Jun 10
Congrats to Chris Faricy (@SyracuseU) and Christopher Ellis on a new @PSJ_Editor article, published with Verasight survey data.
Benefiting from a program warms people to it depending on partisanship and who their party sees as deserving.
Read the full article: doi.org/10.1111/psj.70110
1
59
Jun 10
@USATODAY featured Verasight data and commentary from our CEO Ben Leff this morning in @rachelbarber_'s coverage of the May inflation report.
A national Verasight poll of 3,000 U.S. adults found 80% say annual inflation above 3% is "unacceptable."
Full article below.
1
1
1
65
Jun 9
Congrats to Hwayong Shin of @WashU on a new @polcommjournal article, published with Verasight survey data.
By showing how the "truth-balance dilemma" stems from the perception gap between journalists and audiences, the study underscores the reputational risks that truth-seeking journalism faces in polarized environments.
doi.org/10.1080/10584609.202…
1
79
Jun 8
At the topline, synthetic survey samples look accurate. By party, they don't.
An AI standing in for real respondents misses by 6.6 points overall, but the error nearly doubles within party groups: the model pulls everyone toward 50/50, understating how partisan voters really are.
New synthetic sampling study from @gelliottmorris, @pete_enns, and Benjamin Leff.
1
3
172
Jun 7
A verification method is only useful if the resulting sample still represents the population.
Three design choices set Verasight apart:
1. Recruitment and authentication rely on widely-held credentials.
2. Samples are balanced on a wider set of demographic variables than is standard in the industry.
3. Respondents are verified before, during, and after taking a survey, rather than at any single point.
Read more about Verasight's approach to verification. verasight.io/post/verasight-…
50
Jun 5
Congrats to Matthew Boyce (Texas A&M) and @RebeccaKatz5 (Georgetown) on their new @Vaccines_MDPI open-access study, fielded on the Verasight panel.
55.8% of Americans support tying the child tax credit to age-appropriate polio immunization; 20.9% oppose.
Read the full study:
doi.org/10.3390/vaccines1405…
1
57
Jun 5
Can you train AI to simulate how real people would answer a survey? After a year of research, our answer is still no.
The video walks the data pipeline we built, and the prompts we used.
Watch the video to learn more about the synthetic pipeline we built, then read the full report from @gelliottmorris, @pete_enns, and Benjamin Leff.
1
109
Jun 3
Congrats to @ZoeyRosen and Will Livingston on their new @RMetS study in Meteorological Applications, fielded on the Verasight panel.
Adding health warnings to winter forecasts didn't shift how people judged their risk or planned to protect themselves, but self-efficacy and response efficacy did.
doi.org/10.1002/met.70180
1
46
Jun 3
We built an AI "digital twin" and tested it against human survey respondents.
Across 16 questions it still missed by 6.6 points on average. The gap held across every model and method we tried. New study from @gelliottmorris, @pete_enns, and Benjamin Leff. Link in reply.
1
2
105
Jun 2
Can AI "digital twins" replace human respondents?
To find out, we built an agent pipeline that samples respondents from Census microdata, attaches attitudes from tens of thousands of Verasight interviews, and generates responses.
Full report and methodology👇
Jun 2
As companies begin to sell synthetic (AI-generated) survey responses, new research with @gelliottmorris and Ben Leff shows why you should be skeptical: verasight.io/reports/can-ai-…
1
92
Verasight retweeted
New report out today with my friends at Verasight writing up a couple new studies I designed on AI “digital twins” in polling. The upshot is basically these tools are very inaccurate unless you have raw polling data to build on (which defeats the purpose!) verasight.io/reports/can-ai-…
2
3
11
4,031
May 29
Require a government ID to take a survey, and the people who decline aren't random.
They skew toward people who are more skeptical of institutions and more concerned about data privacy.
That biases the sample, especially on research about trust and politics. verasight.io/reports/id-veri…
1
1
135
May 28
In Verasight's testing, an LLM standing in for survey respondents labeled about 20% of Trump disapprovers as approvers, and made the same error in reverse.
Our research keeps finding it: LLMs cannot replace human respondents. Synthetic samples are predictions, not measurements.
1
121
May 27
Verifying real respondents shouldn't be a single checkpoint. Verasight verifies identity up front (SMS verification, voter-file matching), then measures quality across every survey using multiple signals.
Verasight doesn't require respondents to upload a government ID to participate; it's a rare backstop to let a flagged-but-real person confirm themselves and keep going.
Two independent studies found near-zero AI agents in our data. Read our new blog post to learn more. verasight.io/post/verasight-…
50