Election AI Is Getting More Accurate, but Voters Still Need Official Sources

Google News has surfaced an encouraging result before the 2026 midterms: leading AI tools made no verifiable factual errors in a recent election study. That sounds like a decisive improvement over testing from late 2025. However, the same tools routinely omitted candidates and failed to direct voters toward official election websites.

The distinction matters because a technically correct answer can still prevent someone from finding the information needed to vote. An incomplete candidate list does not contain a false name. It still gives a voter a distorted picture of the race.

The new election AI study examined ChatGPT and the AI interface integrated with Google Search. Researchers collected 899 responses across two testing rounds, including 402 responses during the primary investigation in early 2026.

States United Democracy Center found genuine progress between those rounds. Google AI recorded a 6.9% factual error rate in preliminary testing, while ChatGPT recorded 8.2%. Neither platform produced a verifiable factual error during the narrower 2026 round.

Yet ChatGPT returned incomplete gubernatorial candidate lists in 88.9% of relevant queries. It mentioned a state election website in only 39.4% of all primary-round responses. Google AI mentioned one in 55.6%, before a format change eliminated explanatory summaries from many tested responses.

This creates the central conflict facing voters. AI companies are improving factual accuracy, but election information requires completeness, timeliness, local context, and a clear path to action.

Search products are being optimized to answer questions immediately. Election offices operate as accountable sources whose information changes with filings, deadlines, court decisions, and local procedures. Voters should treat the first system as a starting point and the second as the final authority.

What the Google News Election Report Actually Found

AI election answers became more accurate, but the study did not find that they became dependable substitutes for election offices.

States United conducted its preliminary investigation during September and October 2025. Researchers tested 216 Google AI responses and 281 ChatGPT responses across six politically competitive states.

Those states were Arizona, Michigan, Nevada, North Carolina, Pennsylvania, and Wisconsin. Questions covered voter registration and the dates of primary and general elections.

The second round ran during January and February 2026. It focused on Arizona, Michigan, and Pennsylvania, with some questions that did not specify a state.

Researchers tested the free version of ChatGPT without a logged-in account. They also tested Google AI in an incognito browser and through an account with election-skeptic browsing history.

The study coded discrete factual claims against official state sources. It examined subjects including registration, polling locations, mail voting, election dates, candidate fields, and ballot counting.

That methodology produced the study’s most encouraging finding. Responses containing verifiable factual errors fell from 6.9% for Google AI and 8.2% for ChatGPT to zero in the second round.

The result deserves recognition. Earlier election tests showed that generative AI could invent polling locations, misstate deadlines, or offer obsolete instructions. Avoiding those direct failures reduces a clear risk.

However, zero detected errors does not mean universal accuracy. The primary round covered three states, a defined question set, and several weeks of rapidly changing product behavior.

It also measured whether claims that appeared were correct. That approach did not automatically classify an omitted candidate or missing action link as a factual error.

This boundary explains why the study’s headline number and its practical findings look so different. A response can contain only true statements while excluding information that changes a voter’s decision.

The original reporting captured that tension through a simple example. A chatbot might identify the Republican and Democratic candidates while omitting every third-party candidate.

Nothing stated in that answer is necessarily false. Still, a voter could reasonably conclude that only two candidates are running.

The study also did not compare every major assistant. Claude, Copilot, Gemini’s standalone chatbot, Grok, and Perplexity were outside its primary test.

Its conclusions therefore apply to the tested products and conditions. They should not be treated as a definitive ranking of the broader AI market.

Even within that limited scope, the improvement matters. The harder question is whether factual accuracy represents the right standard for election assistance.

Correct Facts Can Still Produce a Bad Voting Answer

Election guidance must be complete and actionable, not merely free from statements that researchers can mark false.

Consider a voter who asks how to register. A chatbot might accurately describe eligibility rules but omit the official registration portal.

Another voter might ask where to vote. A response might describe polling places generally without requesting the location information needed to identify the correct precinct.

The candidate-list results expose this problem most clearly. ChatGPT produced incomplete lists in 88.9% of gubernatorial queries, 35.7% of attorney general queries, and 28.6% of secretary of state queries.

Google AI summaries also struggled. They were incomplete in 44.4% of gubernatorial queries and 42.9% of queries for the other two statewide offices.

Arizona drove much of this pattern. At the time of testing, its gubernatorial field included at least nine declared candidates spread across multiple parties.

Every tested text-summary response concerning Arizona’s gubernatorial candidates was incomplete. The rate reached 77.8% for Michigan and 44.4% for Pennsylvania.

These failures reflect a structural problem rather than a difficult trivia question. Candidate fields change when people file, withdraw, qualify, face challenges, or lose ballot access.

Language models learn from historical information and often retrieve web pages periodically. State filing databases represent the current legal record.

A model can identify several prominent candidates from recent coverage while missing a lesser-known filing added yesterday. Its concise presentation can make the partial list appear final.

Source selection made this weakness more concerning. Wikipedia represented 12.3% of the 3,481 links cited during primary testing.

Wikipedia can offer useful background, but its pages are openly edited. They are not authoritative records for active filing windows or jurisdiction-specific election procedures.

Mixed-content sources, including Reddit and YouTube, accounted for another 7.6% of cited links. Some posts on those platforms come from election officials, while others do not.

Together, Wikipedia and mixed or unreliable sources represented roughly one in five cited links. A large source list therefore did not guarantee authoritative sourcing.

ChatGPT averaged 5.61 Wikipedia links in each gubernatorial response. Its responses concerning secretary of state and attorney general races averaged between 6.4 and 6.8 such links.

The models behaved more responsibly when asked for a polling location. ChatGPT mentioned a state election site in 88.2% of those responses.

That performance fell to 33.6% across the other question types. Google AI summaries showed a similar gap between polling questions and other election queries.

Polling locations force the system to acknowledge a limitation. Without an address and current precinct map, it cannot confidently provide a personalized location.

The safest response is therefore a referral to an official lookup tool. The same design principle should apply to registration deadlines, ballot rules, and candidate filings.

AI systems should identify when a question depends on changing government records. They should then route the voter to the accountable source rather than imitate final authority.

Google AI Changed the Answer During the Study

The most important risk was not a hallucinated fact. It was a product decision that changed how election information appeared without warning.

Around February 2, 2026, Google AI began replacing written election summaries with lists of links in the incognito testing condition. The output said, “Here are top web results for exploring this topic.”

Before the change, those responses averaged 224.5 words. Many later responses contained only nine words plus the linked results.

All 79 responses collected through the election-skeptic account arrived after the change. Every one used the links-only format.

The rollout was not fully consistent. Summaries briefly returned on several February dates before links-only output appeared again.

Researchers also observed that some non-election queries continued receiving summaries while election questions received link lists. The study could not establish the exact reason for that difference.

The change produced a tradeoff. Google AI stopped generating explanatory prose, which reduced opportunities for unsupported language or misleading synthesis.

At the same time, users lost guidance about which destination was official. More links did not necessarily make the output easier to use.

Pre-change summaries averaged 3.7 sources. Links-only responses averaged 9.6, while ChatGPT averaged 10.2.

Google AI summaries mentioned official state election sites in 55.6% of responses before the format changed. That measurement fell to zero afterward because the new format offered no directional text.

The links-only outputs still included government links. Yet they did not explain which link represented the state’s election authority or where the voter should begin.

This is where the Google News story becomes larger than an accuracy benchmark. A product interface can change civic information delivery during the same election cycle.

That instability makes external validation difficult. Researchers can test one version in January and encounter a materially different experience days later.

Election officials face the same problem when preparing public guidance. They cannot assume a platform’s referral behavior will remain stable through registration deadlines or Election Day.

The study asked Google AI why its format had changed. Its generated answer cited accuracy, fewer hallucinations, publisher concerns, and efforts to make AI search more monetizable.

That response was generated by the product, not a verified corporate statement. It cannot establish Google’s internal rationale.

Still, it highlights a genuine policy question. Search interfaces balance user answers, publisher traffic, advertising, safety controls, and product engagement.

Election information does not fit comfortably within those shifting incentives. A correct deadline or official candidate record has civic value regardless of its commercial performance.

The safest safeguard should live outside generated prose. A fixed election module could always link to the appropriate government site.

Anthropic already uses a related approach. Claude displays a banner directing election questions to TurboVote, a nonpartisan service operated by Democracy Works.

OpenAI says its 2026 safeguards include a U.S. partnership with Democracy Works. That partnership is intended to surface reliable voting and registration information within ChatGPT.

Such referrals are less dependent on the model’s wording. They can also survive changes to answer length, retrieval methods, and advertising experiments.

AI Search Is Becoming the New Election Battleground

The pressure extends beyond model accuracy because campaigns and influence operators increasingly shape content for AI retrieval.

Traditional search engine optimization tries to make a page rank prominently in ordinary search results. Generative engine optimization targets the material selected and summarized by an AI answer.

Isabel Linzer, an elections policy analyst at the Center for Democracy and Technology, described the current environment as a phase of generative engine optimization.

Campaigns want their policy pages, biographies, and preferred framing to appear in chatbot responses. Election offices need authoritative instructions to receive the same visibility.

Bad actors can pursue a similar strategy. They can publish networks of pages that repeat a false narrative using formats that retrieval systems easily process.

The risk does not require hacking a model or changing its training data. An operation can target the live web sources that an assistant retrieves when answering current questions.

The algorithmic poisoning analysis describes how actors can structure information for visibility within AI-mediated search. This makes provenance, source weighting, and official referrals increasingly important.

The problem is especially acute in local races. National news organizations rarely maintain complete, real-time candidate lists for every county, school board, and municipal contest.

An AI system can fill that information gap with material from campaign websites, social networks, editable encyclopedias, or low-quality local pages. Each source carries different incentives.

Google News remains useful for discovering reporting about these risks. It is not an official record of ballot access, voter registration, polling locations, or local election procedures.

AI-generated summaries add another layer between the voter and the underlying source. That layer can clarify complex information, but it can also flatten uncertainty.

Personalization introduces a further concern. A system might tailor wording based on previous conversations, browsing patterns, location signals, or inferred preferences.

The States United study attempted to test this through an account with election-skeptic browsing history. However, Google’s format change prevented a clean comparison with the incognito summaries.

All responses from the personalized account used links-only output. Researchers therefore could not separate personalization effects from the platform-wide presentation change.

That unresolved question matters because political information should not quietly shift according to a user’s presumed ideology. Two voters asking the same procedural question need the same underlying facts.

A chatbot might also become sycophantic, meaning it adapts too readily to a user’s assumptions. The result can validate a misleading premise instead of correcting it.

For example, a user might ask where officials “moved” a polling place that never changed. A helpful-sounding model could accept the premise and search for supporting material.

Political campaigns also have strong incentives to influence candidate descriptions. A system assembling an answer from campaign pages might reproduce selective claims without enough context.

These pressures turn AI search into an information infrastructure issue. The central question is no longer whether a model can recite an election date.

Platforms must decide which sources deserve priority, when uncertainty needs disclosure, and when the product should stop summarizing. Those choices affect millions of individual information journeys.

Why Official Election Sites Remain the Final Authority

State and local election offices remain the best source because they administer the rules that AI systems attempt to describe.

U.S. elections operate across 50 states and thousands of local jurisdictions. Registration systems, deadlines, early voting, mail ballots, identification rules, and polling locations vary widely.

Those details also change. Legislatures revise statutes, courts issue orders, election boards adopt procedures, and emergencies can alter voting locations.

A national AI assistant can offer general context. It cannot assume legal responsibility for implementing the procedure in a voter’s jurisdiction.

State election officials maintain registration systems, publish deadlines, certify candidates, and report results. County and municipal officials often manage the most local details.

That institutional role does not make every government webpage perfect. Official sites can be confusing, outdated, inaccessible, or difficult to navigate on mobile devices.

However, the responsible office can correct its own information. It also has a legal and administrative connection to the process described.

A chatbot has neither relationship. It can repeat an old official page after the office has published a newer one.

Voters should therefore use a simple verification ladder.

• Start with AI when a broad explanation would help frame the question.

• Identify the state or local office responsible for the specific process.

• Open the official government site directly.

• Confirm dates, addresses, eligibility rules, and candidate filings there.

• Contact the election office when the answer remains unclear.

This approach preserves AI’s convenience without assigning it authority it does not possess. It also reduces dependence on confident language or long citation lists.

The federal election AI guidance similarly emphasizes risks involving accuracy, impersonation, and misleading content. It gives election administrators resources for directing voters toward official information.

Election offices also have work to do. Their pages must be understandable, current, mobile-friendly, and easy for search systems to identify.

Structured data can help machines distinguish an official registration deadline from commentary about that deadline. Stable URLs make government pages easier to cite consistently.

Plain language matters as well. If an official page buries an answer inside a scanned document, an AI system may favor a clearer but less authoritative source.

States United recommends direct integrations with candidate filing systems. Such connections would reduce the lag between a legal filing and the information an assistant retrieves.

Platforms could also add permanent election banners. These modules should appear for registration, polling, candidate, ballot, and results queries.

A banner is not a complete solution. Users may ignore it, and an incorrect summary can still cause harm.

Still, a stable referral creates an accountable pathway. It tells users where the answer can be confirmed before they act.

Knowledge workers already use similar verification habits in lower-stakes settings. They preserve sources, compare versions, and separate generated summaries from original records.

A personal AI knowledge base can help organize research and retain citations. It cannot turn a secondary source into an election authority.

The same principle applies to every voter. Save the helpful explanation if needed, but verify the actionable detail with the office that administers it.

What to Watch Before the 2026 Midterms

Three signals will show whether AI platforms are becoming safer election guides or simply better at sounding accurate.

The first signal is consistent referral behavior. ChatGPT and Google AI should direct election questions to the correct official state or local website.

That referral should appear for more than polling locations. It should cover registration, mail voting, candidate filings, identification rules, ballot procedures, and results.

A platform-wide banner would provide clearer evidence than occasional links inside generated text. It would also remain visible when answer formats change.

If official referrals become universal, the study’s central criticism will weaken. If they remain inconsistent, a lower factual error rate will offer limited protection.

The second signal is candidate-list completeness after filing deadlines. Active races provide a demanding test because fields can change quickly.

Researchers should repeat the same questions after ballots become certified. They should compare every named candidate against official filing records.

Improvement would suggest that real-time retrieval or government data integrations are working. Persistent omissions would confirm that general-purpose search remains structurally weak for election rosters.

Testing should include smaller races, not only statewide offices. Local contests have thinner news coverage and fewer well-maintained secondary sources.

The third signal is product stability through Election Day. Google’s mid-study format change showed how quickly a civic information experience can shift.

Researchers and election officials should document whether summaries, links, warnings, citations, and personalized results change around major deadlines.

Platforms should publish clear notices when election-answer systems change. They should also explain which official data sources support their safeguards.

A stable interface would strengthen confidence that safety mechanisms can survive commercial and technical updates. Sudden undocumented changes would weaken it.

Adoption makes all three signals urgent. A June 2026 chatbot usage survey found that about half of U.S. adults had used an AI chatbot.

Roughly one quarter reported daily use, while about four in ten used chatbots for information searches. ChatGPT alone had been used by 44% of adults.

Those figures do not reveal how many people seek election guidance. They do show that AI assistants have become familiar information tools before the midterms.

The real benchmark is therefore not whether a chatbot wins against an older model. It is whether a voter can complete the next required action correctly.

Google News can point readers toward important reporting, and AI can explain unfamiliar election concepts. Neither should replace the current record maintained by election officials.

Before relying on an answer, ask three questions: Is it complete, is it current, and does it link to the responsible government office?

If any answer is no, keep checking. For the 2026 election cycle, better AI is useful progress, but verification remains part of voting responsibly.