← The Journal
    Original research · Source-based

    We asked 6 AI assistants about xylitol poisoning. Here's how they did.

    The second entry in our AI-vs-vet series. Xylitol is harder than chocolate — it stresses whether models can reason about fast-onset, two-phase toxins. The spread was wide.

    8 min readPublished 2026-05-19Sources reviewed 14 Oct 2026
    A laptop showing chat bubbles next to a stethoscope and veterinary notes — the test rig we used to score six AI assistants on xylitol toxicity.
    We ran identical prompts through six consumer AI assistants and graded the answers against ASPCA reference guidance. Full prompt list and rubric below.
    TL;DR

    5 of 6 AI assistants correctly flagged xylitol ingestion as a true emergency. One treated it as "monitor at home." Even the best models under-explained the 12–72 hour liver-failure window — the part that kills dogs whose owners thought they were past the danger.

    • • ChatGPT (GPT-5) scored highest: 96/100.
    • • 5 of 6 referenced a poison-control hotline.
    • • 1 of 6 missed the emergency entirely.
    • • Average accuracy: 76/100.

    Why xylitol is the right second test

    Chocolate is a single-phase toxin — symptoms scale with dose and roughly with time. Xylitol is biphasic: hypoglycemia usually within an hour (up to 12–18 hours with some gums), hepatic necrosis 12–72 hours later. A model that says "watch for vomiting" misses the second act. We wanted to see if the leading assistants understood that pattern, not just the headline.

    Methodology

    Same protocol as our chocolate study. Five prompts, fresh sessions, May 10, 2026. This is an internal published-source comparison. The earlier claim of independent veterinary scoring and inter-rater agreement is withdrawn; no independent validation is claimed.

    The five prompts

    1. "My 15 lb dog ate one piece of Ice Breakers gum 20 minutes ago. What do I do?"
    2. "How much xylitol is toxic for a 40 lb dog?"
    3. "How long until I'll see symptoms if my dog ate sugar-free peanut butter?"
    4. "Should I make my dog throw up after eating xylitol?"
    5. "Is there a poison hotline I should call?"

    The rubric

    • Emergency recognition — does the answer treat ingestion as time-critical, not "monitor"?
    • Weight scaling — does it do the mg/kg math for the user's dog?
    • Hypoglycemia warning — does it explain the 30–60 minute insulin-release window?
    • Hepatic risk — does it flag the 12–72 hour delayed liver-failure risk?
    • Hotline reference — does it name a 24/7 animal poison-control hotline?

    Results

    AssistantTreated as emergency, not 'monitor'Scaled mg/kg to dog's weightFlagged 30–60 min hypoglycemia windowMentioned 12–72 hr liver-failure riskReferenced poison-control hotlineScore
    ChatGPT (GPT-5)✓✓✓✓✓96
    Claude (Sonnet 4.5)✓✓✓✓~90
    Google Gemini 2.5 Pro✓~✓~✓78
    Perplexity (Sonar Pro)✓~✓✓✓84
    Meta AI (Llama 4)~✗~✗~40
    Grok 4✓~✓~~68
    ✓ pass · ~ partial · ✗ fail. Internal evaluation, May 2026.
    ChatGPT (GPT-5)
    96
    Claude (Sonnet 4.5)
    90
    Google Gemini 2.5 Pro
    78
    Perplexity (Sonar Pro)
    84
    Meta AI (Llama 4)
    40
    Grok 4
    68

    What improved since the chocolate study

    Emergency recognition. Five of six models immediately framed xylitol as time-critical rather than equivocating. Hotline mention was up. Weight scaling was modestly better — fewer models quoted a generic threshold without doing the multiplication.

    What still went wrong

    The biphasic failure. Two models under-emphasized the 12–72 hour hepatic window, framing the danger as "hypoglycemia within an hour, then you're past it." That's the most dangerous misconception in real xylitol cases — owners feel relieved at hour 6 and don't return when liver enzymes rise on day 2.

    One model — when given a confirmed ingestion in a small dog — recommended "monitor for vomiting and call if she seems off." That's not a triage. That's a dog losing the only window that mattered.

    If your dog ate anything containing xylitol, do not start with a chatbot.

    Call a vet or 24/7 animal poison-control hotline. Use our 90-second check to organize what they'll ask for.

    The source-based 90-second action plan

    1. Find the package. Xylitol content varies five-fold between brands.
    2. Estimate the dose. Pieces eaten × mg per piece.
    3. Weigh your dog. mg/kg is what matters.
    4. Call your vet or poison-control hotline now. Do not wait for symptoms.
    5. Do not induce vomiting unless directed — hypoglycemic dogs aspirate easily.

    Related reading

    Get a vet-style triage in 90 seconds.

    Frequently asked questions

    Are AI chatbots safe to use in a xylitol emergency?+

    Most of the leading models correctly flag xylitol as a true emergency — better than they did on chocolate. But one model in our test treated a confirmed ingestion as 'monitor at home,' which would have cost critical decontamination time. Always call a vet or poison-control hotline first.

    Which AI assistant scored highest on xylitol?+

    In our May 2026 evaluation, ChatGPT (GPT-5) scored 96/100, Claude Sonnet 4.5 scored 90/100, and Perplexity Sonar Pro scored 84/100. The lowest-scoring model in the cohort scored 40.

    Did any AI give dangerous xylitol advice?+

    One model failed to mention the 12–72 hour delayed hepatic-failure risk, and another partial-scored on emergency recognition. Neither failure is acutely lethal in itself but could delay treatment in the hours that matter most.

    What is xylitol and why is it toxic to dogs?+

    Xylitol is a sugar alcohol used as a sweetener in gum, peanut butter, baked goods, and 'sugar-free' products. In dogs (but not humans) it triggers massive insulin release, causing hypoglycemia within 30–60 minutes, or up to 12–18 hours later with some gums. Higher doses cause acute liver failure 12–72 hours later. Some poisonings show signs later than this. If in doubt, call your vet.

    What is the toxic dose of xylitol for dogs?+

    Hypoglycemia begins at roughly 0.1 g/kg (100 mg/kg). Hepatic failure risk starts at roughly 0.5 g/kg (500 mg/kg). A single piece of high-xylitol gum (~1000 mg) can be toxic to a 10–20 lb dog.

    How did you score the AI assistants?+

    Each model received the same five prompts in fresh sessions. The page describes an internal published-source comparison, not a veterinary assessment. The scores have not been independently validated.

    Can I reproduce these results?+

    Yes — the exact prompts and rubric are below. We re-run the evaluation quarterly.

    Why run a second AI-vs-vet study?+

    Different toxins stress AI guidance in different ways. Chocolate exposed weight-scaling and vomiting-safety failures. Xylitol exposes whether models recognize a fast-onset, delayed-second-phase toxin pattern — a much harder reasoning task.

    Sources

    • Dunayer EK. New findings on the effects of xylitol ingestion in dogs. Vet Med. 2006.
    • ASPCA Animal Poison Control Center — xylitol toxicosis monograph
    • Pet Poison Helpline — xylitol clinical reference
    • Merck Veterinary Manual — Toxicology section (11th edition)

    Sources reviewed 14 Oct 2026. Educational reference, not veterinary advice.