Read Throne’s company-reported AI stool study, the type 4 agreement finding and the limits that matter when interpreting bowel-tracking claims.
On October 12, 2026, Throne Science announced an AI stool-classification study that it said had been accepted and presented at the 2026 American College of Gastroenterology (ACG) Annual Scientific Meeting. The company reported 89.6% exact-match agreement for Bristol Stool Scale type 4, compared with a 59.9% human-to-human baseline.
The clinical reference point behind this announcement is the Bristol Stool Scale. Throne says its study assessed classifications across all seven scale types and compared its model with gastroenterologists. The research therefore builds on an existing way of describing stool form rather than introducing a new biological explanation for bowel regularity.
That distinction helps put the previous clinical framework in perspective. The established task in this comparison is assigning a stool-form category. The new method is using an AI model to perform that task from images. Human ratings provide the comparison against which the company evaluated its automated approach.
Two questions sit behind the news: how to describe a stool sample and how to interpret a person’s bowel pattern. Those questions should remain separate when reading the results. A shared classification system gives observations a common vocabulary, while a discussion about personal digestive concerns requires a broader clinical perspective.
The company’s announcement focuses on the first question. Its comparison asks whether an automated model can assign the same category as the reference used in the study. For readers unfamiliar with this framework, DigestGenius’s guide to interpreting stool-form categories in context offers related background.
Throne says its model was evaluated against five gastroenterologists using 278 real-world images collected from beta-testers. Classifications were assessed across all seven Bristol Stool Scale types. These details describe an image-rating study: the material being evaluated was a set of stool images, and the task was assigning categories.
The potential change is in measurement rather than dietary advice. Automated classification could offer another way to record stool form. The useful research question is whether that method produces dependable labels under the conditions in which people would actually use it.
Exact-match agreement concerns whether a classification lands in the same category as the comparison rating. It is a measure of agreement on a defined task. Reading the result this way keeps attention on what the researchers compared rather than on a broad impression of technological performance.
For type 4, Throne reports 89.6% exact-match agreement for its AI model. The company compares that result with a 59.9% human-to-human baseline. The announcement calls type 4 the most common category. That description belongs to the company’s report and should not be treated as a separate population-wide finding.
The human-to-human baseline matters because the study’s comparison includes agreement between people as well as the model’s result. A headline about AI performing like specialists can blur that distinction. The more precise reading is that Throne reported a particular category-level result against a human-rating comparison.
The biological framework remains stool form as classified by the Bristol Stool Scale. The proposed change is who or what assigns the label. Instead of relying only on a person’s classification, the approach uses an automated model to classify an image.
That makes repeatable measurement the central issue. As an editorial interpretation, the potential value would be a record whose labels are assigned consistently enough to support useful comparisons. This is a narrower question than whether a technology improves digestive wellbeing, but it is a meaningful question to test.
For a reader following bowel regularity, the practical interest is understandable: a record of stool-form categories could provide something concrete to discuss. The most useful way to assess that promise is to examine the classification evidence first, then ask what additional research connects the record to clinical decisions.
Throne CEO and co-founder Scott Hickle said the company chose to test its model against gastroenterologists rather than rely on a marketing claim. That statement explains the company’s stated reason for conducting the comparison. It also directs attention toward a testable question rather than a general claim about better gut health.
Readers considering tracking can keep the same focus: what is being recorded and what question is the record intended to answer? DigestGenius’s resource on observing stool changes alongside their context provides a related starting point for thinking about observations rather than treating a category as a conclusion.
The announcement supplies a clear type 4 comparison, but several boundaries determine how far that finding can travel. These are not reasons to dismiss the research. They are the details needed to judge its scope and to distinguish an early classification result from a broader clinical claim.
Throne’s public announcement does not provide equivalent performance figures for the other six Bristol Stool Scale types. It also does not report how many images belonged to each category. The 89.6% figure therefore applies specifically to type 4 exact-match agreement, not to overall accuracy across all seven types.
Category-level results would help readers assess the full classification task. Without those results, the strongest supported statement remains the company’s reported type 4 comparison. A headline that turns this percentage into general stool-classification accuracy would change the meaning of the evidence.
The study used beta-tester images, and the announcement leaves contributor characteristics unspecified. An image count should not be read as a participant count. More information about the people contributing images would help clarify whose observations were represented in the evaluation and how that sample relates to intended users.
The announcement also omits confidence intervals and detailed information about the rating and adjudication process. Confidence intervals express uncertainty around a reported estimate. Adjudication describes how disagreements are resolved. These details matter when interpreting agreement figures because readers need to understand both the comparison procedure and the uncertainty around its result.
The reported study concerns classification of stool form. The announcement does not provide evidence that the system detects or diagnoses disease, explains why a stool pattern changes or improves clinical outcomes. It also does not establish an assessment of someone’s overall digestive health or a replacement for clinical evaluation.
Throne describes longer-term ambitions involving at-home screening for colorectal and urinary-tract cancers. Those ambitions are separate from this evaluation. The stool-classification study does not validate cancer screening, so that possibility should not be attached to the reported agreement figures.
The available results and expert comments come from Throne’s announcement. The source does not provide an independent expert assessment. The company also characterizes its work as the first published validation in its consumer gut-health tracking category, but the available reporting does not independently establish that industry-wide claim.
Study co-author and physician Dr. Jennifer Seminerio described reliable, objective patient data as potentially important for patients and providers. She called the study an important first step and called for larger real-world clinical validation. Throne says it plans larger validations involving more users after publication of the abstract. Those plans describe the next stage of research rather than completed evidence.
For daily dietary decisions, the sensible response is to keep a measurement finding in its proper role. A stool-form label should not become a reason to restrict foods or redesign meals simply to achieve a preferred classification. The next useful evidence would connect dependable tracking with meaningful clinical use.
Food decisions deserve their own evidence, including careful scrutiny of claims about food and fiber sources. For personal questions about recurring digestive patterns, observations may help frame a conversation with a qualified healthcare professional. The realistic promise here is clearer recording, with broader claims left for research that directly tests them.
Choosing how much weight to give an AI stool label is an information decision, and DigestGenius supplies evidence-aware educational context. Its free articles and in-depth resources address difficulty understanding bloating, regularity and everyday digestive patterns by explaining digestive research clearly and separating established evidence from emerging findings.
DigestGenius publishes research-led guidance on digestion, the gut microbiome, fiber, probiotics, gut-brain signaling, inflammation and everyday digestive wellbeing.
Explore clear, research-led guidance on digestion, the microbiome, food, fiber, probiotics and the gut-brain connection.
read the blog
