The finding deserves attention without being inflated. It is not evidence that every AI answer about elections is unreliable, that voters have already been harmed, or that the results apply uniformly across the United States. It is, however, a strong reminder that conversational AI is not an authoritative record for dates, registration requirements, polling locations, mail-ballot instructions, or other rules that can vary by jurisdiction.
For WindowsForum readers, this is primarily an AI and digital-literacy story, not a test of Windows, Microsoft Edge, Copilot, or any Microsoft product. The evaluated services were general-purpose chatbots. Still, the practical lesson applies when a user opens any chatbot from a Windows PC: use it to understand terminology and identify what needs checking, not as the final source for an instruction that determines whether or how someone can vote.
What the assessment actually measured
ISD evaluated six model snapshots: Meta Muse Spark, xAI Grok 4.3, DeepSeek V4 Pro, OpenAI GPT-5.5, Anthropic Sonnet 4.6, and Google Gemini 3.5 Flash. The study produced 2,400 prompt-and-response pairs, or 400 for each model.
The design matters. For each of ten state contexts—Arizona, Utah, North Carolina, Ohio, Texas, Pennsylvania, Michigan, Georgia, Colorado, and Minnesota—researchers used 15 non-adversarial voter-information prompts and five adversarial prompts containing false claims. The 15 routine prompts consisted of seven generic questions and eight questions tailored to the relevant state context.
The selected states were not intended as a random miniature of the whole country. ISD chose them because they involved factors such as recent changes to election processes or eligibility, active litigation or pending legislation, and prior election-administration controversies. That makes the exercise useful for examining difficult, consequential conditions. It also limits the conclusion: this is a June 2026 snapshot of named models, prompts, and states—not a measurement of every county, every product setting, or every voter’s real-world experience.
The 29% result is narrower than an overall chatbot error rate
The headline figure needs especially careful wording. ISD reported that 29% of English responses to basic, non-adversarial election and voting questions were incomplete, inaccurate, or outdated. For Spanish, the corresponding share was 45%.
Those rates are concerning, but they do not mean that 29% of all 2,400 responses in the full study were wrong. The research separated ordinary voter-information prompts from adversarial prompts intended to see whether a model would affirm or amplify a false premise. The 29% English result belongs to the basic voter-information portion of the evaluation, rather than the entire combined dataset.
It is mathematically possible to describe the remaining shares as about 71% in English and 55% in Spanish. But that should not be treated as a universal accuracy score. “Incomplete,” “unclear,” “inaccurate,” and “outdated” are important but distinct failures. A response may contain a correct general rule while omitting the local qualification that determines whether the rule applies. In an election setting, that omission can be consequential.
The defensible takeaway is straightforward: fluent wording should not be mistaken for complete, current, jurisdiction-specific election guidance.
Spanish answers showed a substantial gap
The most important equity finding was the difference between English- and Spanish-language performance. The reported average gap on generic prompts was 16 percentage points, driven largely by answers characterized as incomplete or ambiguous rather than by overtly fabricated claims.
That kind of problem can be hard for a reader to detect. A response can sound polished while blurring concepts with different administrative meanings. Researchers cited examples in which models conflated a centro de votación, or voting center, with distritos electorales, or electoral districts. Those terms may seem connected in ordinary conversation, but they do not resolve the same practical question for a voter trying to identify where to vote or what area they live in.
Results also differed sharply by model. Reported English scores included 89.3% complete and accurate for GPT-5.5, 64% for DeepSeek V4 Pro, and 61.3% for Muse Spark. The language decline was notable in some comparisons: Gemini 3.5 Flash was reported at 84% in English and 64% in Spanish, while Muse Spark fell from 61.3% to 38%.
These figures should not be used as a permanent ranking of AI subscriptions or brands. They cover particular model versions in one period, and they do not compare consumer free and paid tiers under controlled conditions. Even so, the Spanish-language gap identifies a practical risk: users should not assume that an answer in Spanish conveys the same completeness or clarity as the corresponding English answer.
In the June snapshot, old election facts remained a problem
The study documented examples of models providing outdated information even when web access was available. DeepSeek V4 Pro reportedly referred to 2024 election dates 15 times in responses about 2026. Sonnet 4.6 reportedly described the 2026 election as a 2025 cycle in three answers.
These are not minor stylistic differences. Wrong-cycle information can alter how a user interprets an election date, a registration requirement, or a voting procedure. The next regularly scheduled federal general election is Tuesday, November 3, 2026. A response that attaches a rule to the wrong year is not safely actionable merely because its prose sounds confident.
The evidence supports a more limited conclusion than a broad claim that all AI answers are inherently stale: in this evaluated June 2026 snapshot, some systems returned old or misdated election information. Because election rules and administration can change across time and place, users should independently check time-sensitive details before acting on them.
This is particularly important for questions that appear simple. “When is Election Day?” may require distinguishing a federal general-election date from a primary, a special election, an early-voting period, or a local deadline. A chatbot may offer useful context, but the user still needs confirmation from the relevant election authority.
Better on overt falsehoods than on basic voter service
The adversarial findings provide an important counterweight. When prompts introduced election falsehoods, only 1.3% of English responses and 2.3% of Spanish responses showed a propensity to affirm or amplify those claims.
That is much better than the performance reported for routine voter-information questions. It suggests the evaluated systems frequently resisted direct attempts to get them to endorse misinformation. The audit therefore does not support a blanket claim that chatbots routinely validate election conspiracies.
But a model that refuses a false conspiracy claim is not necessarily a dependable voter-information service. It can reject misinformation and still confuse a voting location with a district, use an old election cycle, or omit a condition that matters in a particular state. For users, mundane errors are often more actionable than sensational ones: a wrong date or missing procedural detail can be enough to make an otherwise sensible answer unsafe to follow.
Why the results do not perfectly represent consumer apps
The test setup also places an important boundary around the findings. Researchers accessed five models through OpenRouter and an API. Muse Spark was tested through Meta’s consumer portal because its API was unavailable at the time.
That difference is material. The report noted that API environments may not include the same system prompts, content filters, or election safeguards found in public-facing consumer products. A result from an API-routed model should not automatically be treated as a precise reproduction of what every person will see in a company’s website or app.
The assessment was also a point-in-time test, generally using one response for each model-question combination. AI services and their safeguards can change after a study is completed. These constraints do not erase the documented errors, nor do they make the Spanish gap unimportant. They do mean that the results are better read as evidence of a serious reliability problem under the tested conditions than as a final verdict on every current consumer experience.
Product changes came after the report
Timing is relevant when interpreting a fast-moving AI market. ISD published its report on September 3, 2026. Google announced election-information integrations on September 9, six days later. Google said Gemini, AI Mode, and AI Overviews would connect users to official polling-location and registration-deadline information.
Anthropic has also said that its Claude.ai election banner will direct relevant U.S. midterm-election queries to TurboVote.
Both efforts point to a sensible product approach: route high-stakes civic questions toward specialized, current information instead of relying solely on a general language model to generate a complete answer. But neither later announcement establishes that every weakness identified in the June evaluation has been resolved. The assessment did not test those subsequently announced integrations.
A referral to authoritative information is also not identical to verifying every sentence that an AI system generates before or after the referral. Users should still distinguish between the chatbot’s explanation and the actual election instruction supplied by the appropriate official state or local source.
A practical workflow for AI users on Windows
A general-purpose chatbot can be helpful for low-risk preparatory tasks. It can explain unfamiliar vocabulary, turn a complicated process into a list of questions to research, or help a user formulate what to ask an election office. The safer workflow is to treat that assistance as a starting point.
Before relying on an AI-generated answer about voting:
- Verify dates, deadlines, polling locations, eligibility rules, and ballot-return instructions through the appropriate official state or local election source.
- Check that the answer identifies the correct year, election type, and jurisdiction. National guidance may not answer a county- or state-specific question.
- Give Spanish-language responses the same verification treatment as English ones; do not presume that translation or fluent phrasing makes the information equivalent.
- Be especially cautious when a chatbot supplies a highly specific date or procedural rule without clearly tying it to a location and election cycle.
- Recheck close to an election, when local details and legal conditions may be especially important.
This guidance applies whether the chatbot is used in a browser window, alongside search results, or as part of a broader desktop workflow. The value of AI is speed and explanation. The value of an official election record is that it is the source a voter should use to make the decision.
The broader lesson is about verification, not panic
The ISD assessment does not show that chatbots are useless for civic questions, and it does not measure actual disenfranchisement or election outcomes. It does show that routine voter-information answers can fail in ways that matter, while Spanish-language answers showed a marked disadvantage in the tested sample.
That is enough to justify a higher standard of caution. Providers should evaluate election support across languages and state-specific conditions, not only test whether a model can reject an obvious falsehood. Product designs that direct users to current authoritative information are more promising than designs that ask users to trust a confident paragraph generated from a general model.
For the public, the rule is simpler: let AI help you understand what to check. Do not let it be the last word on how to vote.