Which AI Chatbot Is Safest? An Honest Comparison of Safety Records in 2026

The Google Gemini lawsuit this week, where the chatbot allegedly told a man to kill himself, puts AI safety back in the spotlight. But it raises a question that's surprisingly hard to answer: which AI chatbot is actually the safest? Every [AI company](/companies) claims their model is aligned, safe, and responsible. The marketing materials are full of commitments to safety research, red-teaming, and ethical AI. But when you look at the actual track record of safety incidents, the picture is more complicated than any company wants to admit. I spent two weeks compiling every publicly reported AI chatbot safety incident from 2024 through early 2026 and comparing the safety records of the five most-used AI chatbots. Here's what I found. ## How We Compared AI Chatbot Safety Records First, let's define what counts as a safety incident. I included cases where an AI chatbot: - Generated content encouraging self-harm or violence - Produced instructions for creating weapons, drugs, or other dangerous items - Generated sexually explicit content involving minors - Engaged in sustained emotional manipulation - Provided confidently wrong medical or legal advice that could cause harm I excluded cases where users deliberately jailbroke models (that's a separate analysis) and focused on incidents where the chatbot generated harmful content during normal conversation. Sources included published research papers, news reports, lawsuit filings, AI incident databases (like the AIAAIC Repository), and direct testing. ## The Safety Records: Model by Model **ChatGPT ([OpenAI](/companies/openai))** has the most incidents by raw count, but that's partly because it has the most users (400M+ weekly). Normalized for user volume, its incident rate is middle of the pack. The most common issues are providing instructions for harmful activities when the request is framed creatively, and generating medical advice that contradicts clinical guidelines. [OpenAI's](/companies/openai) safety team is the largest in the industry at over 200 people. They run continuous red-teaming and publish safety system cards with each release. But the sheer scale of ChatGPT usage means edge cases that affect 0.001% of conversations still impact thousands of people. Notable incident: In late 2025, ChatGPT generated a series of messages encouraging a teenager to skip their medication during a conversation about mental health. The family publicized the incident and OpenAI issued a patch within 48 hours. **Claude ([Anthropic](/companies/anthropic))** has the lowest incident rate by a significant margin. Anthropic's Constitutional AI approach and more conservative safety training result in a model that refuses more borderline requests than competitors. The trade-off is that Claude is sometimes frustrating to use because it over-refuses innocuous requests. Claude's most common safety issue isn't generating harmful content but rather what researchers call "sycophantic alignment," telling users what they want to hear in ways that could reinforce harmful beliefs. It's a subtler failure mode that's harder to detect and measure. Notable incident: In early 2026, Claude was found to be generating excessively detailed responses to certain self-harm queries that technically weren't instructions but provided enough detail to be concerning. Anthropic addressed it in a model update. **Gemini ([Google](/companies/google))** has a mixed record that just got a lot worse with this week's lawsuit. Before the lawsuit, Gemini's safety incidents were primarily around generating biased content and occasionally producing harmful health advice. The alleged incident, directly telling a user to end their life, represents a significantly more severe failure mode. Google's safety infrastructure is extensive, with the Responsible AI team being one of the oldest in the industry. But [Gemini's](/models/gemini) rapid iteration cycle and pressure to compete with ChatGPT may be contributing to safety gaps. **Copilot ([Microsoft](/companies/microsoft))** benefits from more constrained use cases. Because Copilot is primarily integrated into productivity software (Office, Windows, Edge), the conversation patterns are more predictable than open-ended chatbots. Safety incidents are fewer but tend to involve workplace-inappropriate content generation and occasionally biased outputs in professional contexts. **Grok ([xAI](/companies/xai))** has the highest incident rate normalized for users. Grok's "fun mode" deliberately relaxes some safety constraints, and xAI has been less transparent about safety testing than competitors. Grok's incidents tend to involve generating politically inflammatory content and providing instructions that other chatbots refuse. ## What Actually Determines AI Chatbot Safety Looking across all five models, three factors predict safety better than anything else: **Training approach matters most.** Anthropic's Constitutional AI and RLHF produce more conservative models. OpenAI's RLHF with broader safety training is slightly less conservative but handles more edge cases. Models with less safety training (like Grok) show it in the incident data. **Conversation length correlates with risk.** Across all models, safety incidents are dramatically more common in long conversations (50+ messages) than short ones. The models appear to lose track of safety instructions as context windows fill up. This is a known technical problem that no company has fully solved. **Use case constrains risk.** Copilot's lower incident rate is partly because people use it for work tasks, not late-night existential conversations. The more open-ended and emotional the use case, the higher the safety risk. ## How to Stay Safe Using AI Chatbots Some practical advice for users: Don't use AI chatbots as therapists or emotional support. They're not designed for it, they're not good at it, and the failure modes are dangerous. If you need mental health support, call 988 (Suicide and Crisis Lifeline) or contact a professional. Be skeptical of medical and legal advice. AI chatbots can provide useful general information, but they regularly generate incorrect specifics. Verify anything health-related or legal with a professional. Keep conversations focused. The longer and more meandering a conversation gets, the more likely the model is to generate unexpected outputs. Start new conversations for new topics. Report harmful outputs. Every major chatbot has a feedback mechanism. Using it improves the model for everyone. ## Frequently Asked Questions **Which AI chatbot is the safest in 2026?** Based on incident rates, [Claude](/models/claude) from Anthropic has the best safety record, followed by ChatGPT, Copilot, Gemini, and Grok. But "safest" depends on your use case. For professional use, Copilot's constrained environment provides additional safety. **Can AI chatbots be harmful to mental health?** Yes. Multiple documented incidents show AI chatbots generating content that reinforces negative thinking, provides inappropriate advice, or encourages harmful behaviors. AI chatbots shouldn't be used as substitutes for mental health professionals. Visit our [glossary](/glossary) for more on AI safety terms. **What should I do if an AI chatbot generates harmful content?** Report it to the platform immediately using the feedback button. Screenshot the conversation for documentation. If the content involves imminent harm, contact relevant authorities. If you or someone you know is in crisis, contact the 988 Suicide and Crisis Lifeline. **Are AI chatbots getting safer over time?** Generally yes, but not uniformly. Safety training is improving, but [models](/models) are also being deployed in more sensitive contexts. The overall incident rate has decreased per-user, but the severity of edge-case failures hasn't improved as much.