Google’s AI Overviews were designed to give fast, synthesized answers at the top of Search. But when the topic is health, speed and confidence can be a liability, especially if the summary is wrong, incomplete, or too generalized for a person’s situation.
In January 2026, Google began pulling AI Overviews for certain medical queries after reporting highlighted examples that experts described as “dangerous” and “alarming.” The move underscores a central tension in AI-powered search: users want clarity, while medicine often demands nuance, context, and careful triage.
What changed: Google pulls AI Overviews for specific health searches
On January 11, 2026, multiple outlets reported that Google “appears to have removed the AI Overviews” for specific medical searches, including queries such as “what is the normal range for liver blood tests” and “what is the normal range for liver function tests.” The reporting followed a Guardian investigation into the accuracy and safety of AI-generated health summaries.
These removals were not framed as a broad shutdown of health-related AI Overviews. Instead, they looked targeted, certain query formulations stopped showing the feature, while other health topics continued to trigger AI Overviews, according to subsequent summaries in the press.
The Verge characterized the change as Google “pulling” AI Overviews for some medical searches after coverage of dangerous inaccuracies. TechCrunch similarly noted the apparent removals and pointed to liver-test “normal range” queries as clear examples of the feature being switched off.
The investigations that triggered the response
The Guardian’s January 2, 2026 investigation cataloged cases where AI Overviews allegedly produced misleading or harmful health guidance. The article highlighted examples that experts said could lead users toward inappropriate self-management or false confidence about serious conditions.
Among the cited issues were problematic summaries involving cancer-related information and test interpretation. The reporting emphasized that even small errors in medical context, what to eat, what a lab value means, when to seek urgent care, can have outsized consequences.
On January 11, 2026, The Guardian reported that the liver test “normal range” AI Overviews had been removed after experts warned they could mislead people with serious liver disease. The article also observed that slightly altered queries could still produce AI Overviews at the time of testing, raising concerns about how consistently safety changes were applied.
Why “normal ranges” can be a trap in liver function tests
One of the key flashpoints was the way AI summaries treated “normal range” questions for liver blood tests and liver function tests (LFTs). In clinical practice, interpreting LFTs is rarely as simple as comparing a number to a reference interval, context matters, trends matter, and symptoms matter.
Vanessa Hebditch of the British Liver Trust warned that interpretation is “complex” and that AI Overviews may fail to communicate that someone can have “normal” results and still have serious liver disease. She added: “This false reassurance could be very harmful.” That critique reflects a classic patient-safety risk: an overly neat answer that discourages follow-up.
The Guardian’s reporting suggested that experts were particularly concerned about the potential for an AI-generated summary to imply that “normal” equals “healthy,” without prominently stating limitations, red flags, or the need for clinical interpretation, especially for people already at risk.
Pancreatic cancer diet advice: when AI confidence meets clinical reality
Another widely cited example involved dietary guidance for pancreatic cancer. The Guardian reported that an AI Overview included advice that experts challenged as unsafe or misleading for patients navigating complex treatment pathways and nutritional needs.
A Pancreatic Cancer UK representative described the guidance as “completely incorrect” and warned it “could be really dangerous and jeopardise a person’s chances of being well enough to have treatment.” In oncology, nutrition is often individualized and can be integral to a patient’s ability to tolerate therapy.
This episode illustrates a broader AI risk: a summary can sound authoritative while compressing complex, patient-specific guidance into simplistic rules. Even when the underlying sources are mixed-quality, or when the model generalizes beyond the evidence, the final output can appear decisive.
Query variations: why small wording changes still matter
One of the most concerning technical details in the coverage is that removals did not necessarily apply cleanly across similar searches. TechCrunch and other reports noted that slight variations, such as “lft reference range” or “lft test reference range”, could still surface AI Overviews (at least at the time of reporting).
This matters because real users do not search in standardized clinical language. They try different phrasings, abbreviations, and fragments, often iterating until they see something that looks like an answer, meaning gaps in enforcement can undermine the intended safeguard.
If safety protections depend heavily on exact phrasing, a system may reduce harm for one query while leaving near-identical queries exposed. That dynamic is particularly risky in health, where the same underlying question can be asked in dozens of everyday ways.
Google’s public stance and the broader policy context
Across coverage, Google’s spokesperson declined to address specific examples, saying: “We do not comment on individual removals within Search… we work to make broad improvements…” That position leaves outside observers to infer what changed from live testing and reported behavior rather than from a detailed changelog.
Google has previously emphasized protections for AI Overviews. In a May 2024 update, the company said it added restrictions for queries where AI Overviews weren’t helpful and that “in the case of health, we launched additional triggering refinements to enhance our quality protections.” Google also stated that “less than one in every 7 million unique queries” had a content policy violation, according to its internal measurement.
The January 2026 removals suggest that, despite stated safeguards, edge cases (and sometimes not-so-edge cases) can still slip through, especially when the question invites oversimplified medical interpretation. The continuing presence of AI Overviews on other medical topics, as noted in business reporting, also indicates that Google is iterating rather than retreating from AI summaries in health altogether.
Google pulls AI Overviews for health queries in moments like this not only because the examples are alarming, but because health search is uniquely high-stakes: users may act on what they read immediately, without a second opinion. When an AI summary compresses uncertain or context-dependent medicine into a neat paragraph, the harm is not only misinformation, it’s misplaced confidence.
The practical takeaway is that the battle is less about whether AI should summarize health information and more about how consistently it can recognize when it should not. The January 2026 episode, spanning liver test ranges, pancreatic cancer diet advice, and query-variation loopholes, shows that accuracy is necessary but not sufficient; safe presentation, robust triggering rules, and conservative defaults may matter just as much.
