Back to AI & Tech

Where people draw the line on AI in healthcare

Source reportMethodology

Overview

AI health tools are part of the landscape, but the strongest signal in this cluster is caution around high-stakes care.

Adults report limited recent use, low mental-health disclosure to AI tools, and little willingness to pay for a researcher-developed AI nutrition chatbot.

Topline

62.9% would only trust a human provider to adjust medicines.

For which of these tasks would you only trust a human healthcare provider versus an AI agent?

  • Adjusting your medicines 62.9%
  • Addressing your concerns about your medicines 57.5%
  • Emotional support 54.8%
  • Telling you whether your medicine is necessary 53.3%
  • Interpreting your lab data, medication or health record information 50.9%
  • Providing dietary advice 31.8%

2026 · base n 1,000 · +/- 3.3%

2

View source data

Human providers still hold the boundary

Human-provider preference is clearest around medication decisions, with 62.9% saying they would only trust a human provider to adjust medicines.

Majorities also reserved medication concerns, emotional support, and medicine necessity decisions for human providers.

Topline

55.7% reported no recent AI-based health-tool use.

In the past 3 months, what have you used AI-based tools (e.g., ChatGPT) for related to your health?

  • None of the above 55.7%
  • Understanding a health condition, symptom, or diagnosis 27.0%
  • Learning about medications or treatments 19.8%
  • Diet, nutrition, or meal planning 15.4%
  • Physical activity or lifestyle improvement 12.5%
  • Preparing for or following up on a medical visit 11.6%

2026 · base n 1,000 · +/- 3.3%

2

View source data

Additional supporting data from this section.

Topline

75.6% would not pay for an AI nutrition chatbot.

What is the maximum you would pay monthly for an artificial intelligence (AI) nutrition chatbot developed by researchers?

  • $0 / I would not pay for this 75.6%
  • Up to $5 10.1%
  • $5 to $10 8.2%
  • $10 to $20 4.4%
  • $20 to $30 1.5%
  • More than $30 0.3%

2026 · base n 1,000 · +/- 3.3%

2

View source data

Recent use remains limited

A 55.7% majority said they had not used AI-based tools for health in the past three months.

Another 84.4% said they had not disclosed mental health concerns to AI tools such as ChatGPT, and 75.6% said they would not pay for an AI nutrition chatbot developed by researchers.

Views of AI consciousness add context

Views on whether large language models have consciousness or sentience are split across rejection, future possibility, current belief, and uncertainty.

Methodology

Full methodology
Mode
Verasight panel recruited via random address-based sampling, random person-to-person text messaging, and dynamic online targeting
Population
US adults age 18+
Field dates
2026-05-01 → 2026-05-04
Base (unweighted)
1,000
Margin of error
+/- 3.3%
Module
2
Sponsor
Verasight
Weight variable
weight
Weighting targets
age, race/ethnicity, sex, income, education, region, metropolitan status

Sources

[5]

Citation

SBM Omnibus Survey #2026-049, fielded May 1-4, 2026, N=1,000 US adults age 18+, +/- 3.3%.

https://reports.verasight.io/reports/sbm-2026#q-2-8

Verasight survey methodology

How Verasight conducts surveys.

This page describes the Verasight general survey contract, separate from how the Data Library packages it. Each wave's specific field dates, sample sizes, and module breakdown are listed in that wave's report.

Mode
Verasight panel recruited via random address-based sampling, random person-to-person text messaging, and dynamic online targeting.
Population
US adults age 18+.
Sample design
Surveys are run as omnibus or single-topic waves. Omnibus waves are split into modules with their own respondent set, typically around one thousand respondents per module.
Field window
Each wave specifies its own field dates. Most omnibus waves field across roughly two weeks.
Weighting
Per-module weighting to CPS targets including age, race and ethnicity, sex, income, education, region, and metropolitan status.
Partisanship benchmark
Pew Research Center's NPORS benchmarking surveys, three-year running average.
Vote benchmark
2024 presidential vote population benchmarks.
Margin of error
Typically about plus or minus 3.4 to 3.6 percent per module at standard module sizes. Question-level MoE is recomputed when a base shrinks materially below the module baseline.
Reporting
Every wave is published as a standalone report at verasight.io/reports with full instrument and methodology.
Transparency
Verasight is a member of the American Association for Public Opinion Research Transparency Initiative.

Wave-specific methodology, full weighting variable lists, and verbatim instrument text live in each report at verasight.io/reports.