Consulting AI before a major personal decision vs deciding without it
Last reviewed 2026-05-11
Evidence quality 4.0/5
Eight-dimension review score against the
quality rubric
. Each dimension scored 1–5.
D1 Source verification
4/5
D2 Source authority & independence
5/5
D3 Regret-rate accuracy
2/5
D4 Source comparability
2/5
D5 Gilovich pattern
4/5
D6 Prose quality
5/5
D7 Caveat completeness
5/5
D8 Sample quality
5/5
Average4.0/5
Proxy data — no direct regret survey exists for this decision. Rates are derived from satisfaction scores and access-barrier data rather than questions that directly asked about regret. See caveats below.
Action regret
Consulting AI before deciding
22%
~22% of those who consulted AI for a significant decision report a harmful or problematic outcome (proxy; 11% received unsafe health recommendations; 21–49% of AI medical responses are rated problematic in controlled studies)
US adults who consulted AI chatbots for health, financial, or life decisions
October–December 2025
Inaction regret
Deciding without AI input
28%
~28% of those who skipped AI consultation may have missed useful information that could have improved their decision (proxy; 46–59% of AI health users report concrete benefits; 66–90% of AI financial users found it worthwhile)
US adults who made significant decisions without consulting AI
2025–2026
% who regret this choice
Consulting AI before decidingDeciding without AI input
22%28%
balanced — Roughly balanced — both choices carry similar regret.
Related decisions
Semantically similar decisions — same territory, different trade-offs.
About 1 in 4 US adults has now consulted AI for health information or advice, according to a West Health/Gallup panel of 5,660 adults surveyed in late 2025. Among those users, 46% felt more confident asking their providers questions afterward, and 59% used AI to prepare before a doctor visit — concrete stated benefits that non-users forgo. But the same survey found that 11% of AI health users reported receiving unsafe recommendations, and a parallel UCLA/BMJ Open study rating 250 AI responses to medical questions found 49.6% were problematic to some degree — mostly delivered with confidence and few caveats, making them difficult for users to identify as unreliable. An MIT Media Lab study published in NEJM AI documented that participants systematically overestimated AI medical reliability and could not distinguish AI-generated from physician responses, even when the AI response was inaccurate.
The regret arithmetic here is genuinely ambiguous, which is unusual in this dataset. The action-side risk (22% proxy, bounded by 11% unsafe-recommendation rate and 49.6% problematic-response rate) and the inaction-side opportunity cost (28% proxy, based on AI users’ reported benefits) are close enough that the entry is classified as ‘mixed’ rather than clearly inaction-dominates. How AI is used matters more than whether it is used: supplementing a scheduled doctor visit with AI research before attending is a different action category than using AI as a triage replacement for a symptom that warrants evaluation. The former has low action-risk and meaningful information benefit; the latter has higher action-risk and may cause harmful delay. The 14 million Americans who skipped a provider visit based on AI advice represent the higher-risk end of the use spectrum, though some of those skipped visits may have been genuinely unnecessary.
The honest summary of this entry’s evidentiary state: the AI consultation decision is too domain-specific, use-case-dependent, and rapidly evolving to generate a stable regret-pair estimate. Financial AI consultation self-reports are overwhelmingly positive (Wells Fargo: ~90% found results worthwhile), which would pull the inaction-regret figure up substantially if the financial domain were weighted equally with health. Medical AI consultation carries real documented risk of inaccurate confident advice in a domain where acting on wrong information has direct health consequences. The mixed classification reflects both the genuine uncertainty and the heterogeneity of “consulting AI” as a decision — something between a research tool and an advisor, with properties of each and the disclaimers of neither.
Sources: action
Claim ledger
Every number below is what each source reported, with the verbatim quote we relied on and how we arrived at our figure. Click any link to verify directly.
[1]West Health / Gallup — Millions of Americans Now Consult AI Before, After, and Sometimes Instead of Seeing a Doctor
Reference source
11% of US adults who used AI for health information reported receiving unsafe recommendations; 14% (~14 million) skipped a provider visit based on AI advice
Excerpt
“"11% of AI health users reported receiving unsafe health recommendations from AI. 14% of recent AI health users skipped a provider visit based on AI advice. 46% felt more confident asking providers questions after using AI."
”
Source data from
2026-01-01
Accessed
2026-05-11
Calculation
West Health / Gallup panel survey of n=5,660 US adults, fielded October–December 2025, margin of error ±2.1pp. This is the most methodologically rigorous large-sample US survey on AI health consultation outcomes. The 11% unsafe-recommendation rate is the primary action-regret proxy: it represents the share of users who received advice that was later identified as unsafe, which is a necessary precursor to regret even if regret is not directly measured. The 14% who skipped a provider visit based on AI advice is a behavioral-consequence measure; it does not directly map to regret (some may have correctly assessed that no visit was needed). The 22% action-regret estimate is bounded by the 11% unsafe-recommendation rate (lower bound) and the broader 21–49% problematic-response rates from academic accuracy studies (upper bound), centered at approximately 22% to reflect that problematic responses do not always produce regrettable outcomes.
[2]CIDRAP (University of Minnesota) reporting on UCLA / BMJ Open study — AI Chatbots Provide Poor Answers to Medical Questions Half the Time
Reference source
49.6% of AI responses to medical questions were rated as problematic in a blinded evaluation of 5 major chatbots (30% somewhat, 19.6% highly problematic)
Excerpt
“"49.6% of AI chatbot responses to medical questions were rated as problematic — 30% 'somewhat problematic' and 19.6% 'highly problematic.' Chatbot responses were consistently given with confidence and certainty, with few caveats or disclaimers."
”
Source data from
2025-02-01
Accessed
2026-05-11
Calculation
UCLA / BMJ Open study of 250 total questions across 5 major chatbots (ChatGPT, Gemini, DeepSeek, Meta AI, Grok), 10 questions each across 5 medical categories, data collected February 2025, published 2026. The 49.6% problematic-response rate establishes the upper bound of action-side risk for health-domain AI consultation. The key finding about confident framing with few caveats is the mechanism that converts problematic responses into potential harm: users cannot easily identify which responses are in the unreliable half. Used here to anchor the upper bound of the 22% action-regret estimate; the lower bound is the West Health/Gallup 11% unsafe-recommendation rate from real-world self-report.
[3]NEJM AI / MIT Media Lab — Patients Cannot Distinguish AI-Generated from Physician Responses
Peer-reviewed
Participants rated high-accuracy AI responses as significantly more valid and trustworthy than physician responses — and could not identify inaccurate AI responses as such
Excerpt
“"Participants could not distinguish AI-generated responses from doctor responses. They rated high-accuracy AI responses as significantly more valid, trustworthy, and complete — but critically, they also overvalued AI responses when accuracy was low, suggesting systematic overtrust."
”
Source data from
2024-01-01
Accessed
2026-05-11
Calculation
NEJM AI / MIT Media Lab study on AI medical response evaluation. The overtrust finding is the mechanism explanation: users systematically overestimate AI reliability in the medical domain, meaning the problematic-response rates from accuracy studies translate into real-world harm more readily than they would if users could identify unreliable responses. This supports using the problematic-response rate (not just the unsafe- recommendation rate) as the relevant risk bound.
Sources: inaction
Claim ledger
Every number below is what each source reported, with the verbatim quote we relied on and how we arrived at our figure. Click any link to verify directly.
[1]West Health / Gallup — Millions of Americans Now Consult AI Before, After, and Sometimes Instead of Seeing a Doctor
Reference source
46% of AI health users felt more confident asking providers questions afterward; 59% used AI to research before a doctor visit
Excerpt
“"46% of AI health users felt more confident asking providers questions after using AI. 59% used AI to research before a doctor visit. 71% were motivated by wanting answers quickly; 71% wanted additional information."
”
Source data from
2026-01-01
Accessed
2026-05-11
Calculation
West Health / Gallup panel, n=5,660, October–December 2025. The 46% who felt more confident asking providers questions is a concrete stated benefit of AI consultation — a benefit that non-users forgo. The 28% inaction-regret proxy is derived conservatively from the proportion of AI users who report concrete preparation or information benefits (46–59%), adjusted downward to reflect that non-users may have obtained similar information through other channels (search, calling a nurse line, reading drug inserts). No direct "do you regret not consulting AI for this decision?" survey was identified. This is the most data-sparse side of the entry.
[2]ABA Banking Journal (reporting on Wells Fargo and TD Bank surveys) — Bank Surveys Find Consumers Increasingly Turning to AI for Financial Advice
Reference source
~90% of US adults who used AI for financial decisions and acted on the advice said results were 'profitable or worthwhile' (Wells Fargo 2026); 55% of adults use AI for financial management decisions (TD Bank 2026)
Excerpt
“"19% of U.S. adults used AI for financial advice; 38% among Gen Z. Two-thirds of those who used AI acted on its suggestions. Approximately 90% of those who acted said the results were 'profitable or worthwhile.'"
”
Source data from
2026-04-01
Accessed
2026-05-11
Calculation
Wells Fargo consumer survey and TD Bank consumer survey, both reported April 2026. Sample sizes and full methodology not disclosed. The 90% "worthwhile" self-report is the most positive available outcome figure for AI financial consultation but is highly susceptible to self-serving bias: people who acted on AI advice and lost money are less likely to report the action as worthwhile, but are also less likely to be included in a consumer satisfaction survey. Used here as directional corroboration that AI financial consultation produces positive self-reported outcomes at high rates, supporting the inaction-regret proxy. Not used as a primary figure due to methodological opacity.
Caveats
PROXY MEASUREMENTS THROUGHOUT. No survey has directly asked individuals "do you regret consulting AI for this decision?" or "do you wish you had consulted AI before deciding?" Both sides are constructed entirely from adjacent data: outcome quality studies (for action-side risk), reported benefits (for inaction-side opportunity cost), and behavioral measures (provider visit skipping rates). The regret_delta of -0.06 is too small to reliably classify this as inaction-dominates; it is classified as 'mixed' because the evidence does not clearly support either pattern. The action-side risk varies dramatically by domain and by how the AI is used: supplementing a doctor visit (low risk, high benefit) vs. replacing one for a serious symptom (high risk). The academic accuracy studies (21–49% problematic response rates) reflect AI performance on medical questions specifically — the highest-stakes domain. Financial and general life decisions likely have lower problematic-response rates but fewer published accuracy benchmarks. The NEJM AI overtrust study is a particularly important caveat: the mechanism that converts problematic responses into harm is users' inability to detect them, and confident framing with few caveats (documented in the UCLA/BMJ Open study) systematically conceals this uncertainty. "Using AI" is not a single action: reading an AI summary to prepare for a doctor visit is categorically different from acting on an AI diagnosis instead of seeing a doctor. This entry aggregates these heterogeneous uses, which limits its precision. The Wells Fargo and TD Bank self-report data is opaque on methodology and should be treated as directional only. AI capabilities are improving rapidly; accuracy and reliability benchmarks from 2025–2026 will likely differ from those applicable 12–24 months later. This entry has a shorter shelf life than most in the dataset.
