Six rounds of TechTonic Justice + Data for Progress research.
Round 01
Main concerns about government use of AI
Mar 6–9, 2026N=1,242 likely votersMaxDiff
Tested: main concerns about government use of ai
Public services harder to use+6.3
One mistake harms millions+5.3
Misleading the public with fake media+3.8
No one takes responsibility for harm+2.6
Wrongful fraud flags or benefit cuts+2.1
What this suggests. Concrete consequences and service failures outranked abstract technical warnings.
MaxDiff is a relative preference NET score, not a percentage. Positive and negative values reflect comparison within this round.
Round 02
Which broader AI harms feel most serious?
Mar 20–23, 2026N=1,193 likely votersMaxDiff
Tested: which broader ai harms feel most serious?
Insurer AI denies doctor-recommended care+15.7
Data centres strain water and power+5.2
Government outsources programs to tech firms+3.7
Benefits, housing or jobs cut without accountability+1.3
Tracking data without meaningful consent+0.5
What this suggests. The insurer denying doctor-recommended care was the standout relative concern.
MaxDiff is a relative preference NET score, not a percentage. Positive and negative values reflect comparison within this round.
Round 03
Who controls consequential AI decisions?
Apr 24–27, 2026N=1,167 likely votersAt least somewhat upset · %
Tested: who controls consequential ai decisions?
Health insurance companies68%
Landlords67%
Private companies running public programs67%
State government officials66%
Big Tech companies64%
Employer53%
What this suggests. Even comparatively trusted institutions drew unease when placed in charge of consequential decisions.
Percentages are shares of respondents under the question wording shown. Ranges preserve the source's split-sample or version variation.
Round 04
Concrete harms and lived consequences
Jul 2–5, 2026N=1,164 likely votersExtremely or very concerned · %
Tested: concrete harms and lived consequences
Doctor-recommended treatment denied73%
Social Security or disability cut off72%
Falsely accused of benefits fraud71%
Charged more for groceries71%
Job rejection with no human involved70%
Forced to deal with AI instead of a person69%
What this suggests. Specific losses in care, benefits, prices and human recourse generated high concern. The separate forced-choice priority is a different measure.
Percentages are shares of respondents under the question wording shown. Ranges preserve the source's split-sample or version variation.
Round 05
Which three-word response feels right?
Jul 10–13, 2026N=1,219 likely votersMaxDiff
Tested: which three-word response feels right?
Prevent, Protect, Repair+24.7
Stop, Protect, Repair+7.1
Prevent, Stop, Fix+2.7
Stop, Protect, Fix-4.3
Prevent, Stop, Treat-7.5
Prevent, Mitigate, Repair-22.6
What this suggests. Prevent, Protect, Repair led the comparative language test across partisan groups.
MaxDiff is a relative preference NET score, not a percentage. Positive and negative values reflect comparison within this round.
Round 06
What rules are people willing to support?
Jul 24–27, 2026N=1,058 likely votersSupport · % or reported range
Tested: what rules are people willing to support?
Meaningful penalties77–81%
Off switches79–80%
Safety proof before release79–80%
High-stakes bans70–80%
Safety proof before use77%
Community boards with legal authority74–75%
What this suggests. The full policy package had 72% support. Ranges reflect versions or split samples and should not be treated as single estimates.
Percentages are shares of respondents under the question wording shown. Ranges preserve the source's split-sample or version variation.
Source: TTJ Message Testing Cheat Sheet, September 2026. Rounds 1, 2 and 5 report relative MaxDiff scores; other rounds report percentages. Question wording, samples and methods vary across rounds.
Follow the themes
Follow the themes.
03 / Practical use
Use the research.
Choose a resource or build a message grounded in the findings.
Practical resources
Swipe to explore all three resource groups →
Movement organisations
Message guidance and concise references for organisers and advocates.
