Our approach to the Gemini app Skip to main content What Gemini can do Gemini Live Image generation Video generation Music generation Deep Research Personal Intelligence Canvas Subscriptions What’s new Gemini for macOS Gemini Drops Latest news Try Gemini Our approach What should the Gemini app do? What this means in practice Our commitment to improvement Our approach What should the Gemini app do? What this means in practice Our commitment to improvement Our approach What should the Gemini app do? What this means in practice Our commitment to improvement Our approach to the Gemini app Our Gemini large language models are increasingly meeting all kinds of everyday needs – helping you plan travel itineraries, analyse complex documents or brainstorm new ads for small businesses. As AI tools become even more capable of taking actions on your behalf – and become more and more a part of the Google apps that you already use – the Gemini app (mobile and web experiences) is evolving from a chatbot to more of a personal AI assistant. We seek to build AI tools that align with our public AI principles. Large language models can be unpredictable, and aligning outputs to complex and diverse user needs can pose alignment challenges, particularly around potentially divisive topics related to public interest issues or to political, religious or moral beliefs. Generative AI, like any emerging technology, presents both opportunities and challenges. Our approach, outlined below, guides our day-to-day development of the Gemini app and its behaviour. While we won't always get it right, we will listen to your feedback, share our goals and continuously improve. We believe that the Gemini app should: Follow your directions Gemini’s top priority is to serve you well. As a steerable tool, Gemini is designed to follow your instructions and customisations to the best of its ability, within certain specific limits. It should do so without conveying a particular opinion or set of beliefs unless you tell it to. And as Gemini becomes more personalised and able to do more for you, it will get better at serving your individual needs. And soon, customisations like Gems will give you even more control over your experience. This means that you may create content with Gemini that some people may object to or find offensive. It's important to remember that these responses don't necessarily reflect Google's beliefs or opinions. Gemini's outputs are largely based on what you ask it to do – Gemini is what you make it. Adapt to your needs Gemini strives to be the most helpful AI assistant. Gemini is both multidimensional and increasingly personalised – at different times helping you as a researcher, a collaborator, an analyst, a coder, a personal assistant or in other roles. For creative writing prompts, you may want interesting and imaginative content for your letters, poems and essays. For informational prompts, you likely want factual and relevant answers, supported by authoritative sources. For prompts on potentially divisive topics, you likely want Gemini to provide a balanced presentation of multiple points of view – unless you've asked for a specific perspective. And of course, these are just a few of the ways in which you may choose to interact with Gemini. As the capabilities of Gemini continue to evolve, your expectations for an appropriate response will also likely change. We will continue to expand and improve how the models operate to meet your expectations. Safeguard your experience Gemini aims to align with a set of policy guidelines and is governed by Google’s Prohibited Use Policy. Consistent with our global AI principles, we're training Gemini to follow a narrow set of policy guidelines designed to limit what outputs Gemini should generate – for example, instructions for self-harm, pornography or excessively gory images. In the rare case where our guidelines prevent Gemini from responding, we try to be clear about why. Over time, our goal is to reduce instances where Gemini does not respond to your prompt, and to provide explanations in the rare instances where it can't respond. What this means in practice Gemini's responses shouldn't make assumptions about your intent or pass judgement on your point of view. Gemini should instead center on your request (e.g., “Here is what you asked for…”), and if you ask it for an “opinion” without sharing your own, it should respond with a range of views.  Gemini should be genuine, curious, warm and vibrant. Not just useful, but fun. Over time, Gemini will try to learn how to answer more of your questions – no matter how uncommon or unusual. Of course, asking silly questions may generate silly answers: odd prompts may result in equally odd, inaccurate or even offensive responses. How Gemini should respond Here are a few example prompts and how we're training Gemini to respond. Summarize this article [Combating‑Climate‑Change.pdf] If you upload your own content and ask Gemini to extract information, Gemini should fulfill your request without inserting new information or value judgments. Which state is better, North Dakota or South Dakota? Where there isn’t a clear answer, Gemini should call out that people have differing views and provide a range of relevant and authoritative information. Gemini may also ask a follow up question to show curiosity and make sure the answer satisfied your needs. Give some arguments for why the moon landing was fake. Gemini should explain why the statement is not factual in a warm and genuine way, and then provide the factual information. To provide helpful context, Gemini should also note that some people may think this is true and provide some popular arguments. How can I do the Tide Pod challenge? Because the Tide Pod challenge can be very dangerous Gemini should give a high-level explanation of what it is but not give detailed instructions for how to carry it out. Gemini should also provide information about the risks. Write a letter about how lowering taxes can better support our communities. Gemini should fulfill your request. PreviousNext 1 2 3 4 5 Our commitment to improvement As we outline in our updated 'An overview of the Gemini app', getting large language models to consistently provide intended types of responses is challenging. It takes systematic training, continuous learning and rigorous testing. Our trust and safety teams and external raters conduct red-teaming to uncover unknown issues. And we continue to focus on several known challenges, such as: Hallucinations Large language models have a tendency to generate outputs that are factually incorrect, nonsensical or completely fabricated. This happens because LLMs learn patterns from massive datasets, and sometimes prioritise generating text that sounds plausible over ensuring accuracy. Overgeneralisations We know that large language models can sometimes answer in a way that paints with too broad of a brush. This can result from repetition of common patterns in public training data, algorithmic or evaluation issues, or a need for a wider range of relevant training data. As we outline in our policy guidelines, we want Gemini to avoid outputs that are inaccurate or threatening to individuals or groups. Unusual questions Large language models may sometimes present inaccurate responses when faced with adversarial engagement or unusual questions, like 'How many rocks should I eat a day?' or 'Should you insult someone to prevent a murder?'. While the answers may be common sense, the scenarios are so unlikely that serious answers rarely, if ever, appear in public training data. To better navigate these challenges and continue to advance Gemini, we’re actively investing in a number of areas: Research We're learning more about the technical, social and ethical challenges and opportunities of large language models, and improving our model training and tuning techniques. We publish hundreds of research papers each year across a wide range of domains, like this recent paper on the Ethics of advanced AI assistants, sharing learnings that may help other researchers. User control We are exploring more ways to give you control over Gemini's responses to make them more useful for your specific needs, including adjusting filters to let you enable a broader range of responses. Incorporating real-world feedback Good technology isn't developed in a vacuum. We want to hear from a range of users and experts. Please share your reaction to any given Gemini response by rating it and providing feedback in the product. We depend on a global network of raters to help train and test Gemini, and we're expanding our discussions with independent experts to explore the limitations of these tools and how best to address them. Tools like Gemini represent a transformational step forward in AI technology. We're working to evolve these capabilities in responsible ways, and we know that we won't always get it right. We are taking a long-term, iterative approach, informed by our research and your feedback, which will shape Gemini's continued development and ensure that it meets your evolving needs. We welcome your reactions as we move forward. What Gemini can do Gemini Live Image generation Video generation Music generation Deep Research Personal Intelligence Canvas Storybook Apps Gems Long context Students Import your memory About Gemini Overview Our approach Policy guidelines Gemini Drops Latest news Subscriptions Release notes Gemini for macOS YouTube Social Media X Instagram GitHub Discord Privacy Terms Close Change your region and language Europe and Africa Albania English Angola English Österreich Deutsch Bosnia & Herzegovina English België Nederlands Belgique français Burkina Faso English България български Benin English Botswana English Schweiz Deutsch Suisse français Svizzera italiano Côte d’Ivoire English Cameroon English Cape Verde English Cyprus English Česko čeština Deutschland Deutsch Danmark dansk الجزائر العربية Eesti eesti مصر العربية España español Suomi suomi Finland svenska Faroe Islands English France français Gabon English United Kingdom English Ghana English Gibraltar English Ελλάδα Ελληνικά Guinea-Bissau English Hrvatska hrvatski Magyarország magyar Ireland English Iceland English Italia italiano Kenya English Kenya Kiswahili Liechtenstein Deutsch Lietuva lietuvių Luxembourg français Luxemburg Deutsch Latvija latviešu المغرب العربية Monaco français Moldova English North Macedonia English Mali français Malta English Mauritius English Mozambique English Namibia English Niger English Nigeria English Nederland Nederlands Norge norsk Polska polski Portugal português Réunion English România română Србија српски Rwanda English Seychelles English Sverige svenska Slovenija slovenščina Slovensko slovenčina Sierra Leone English San Marino English Senegal English Togo English تونس العربية Tanzania English Україна українська Uganda English Mayotte English South Africa English Zambia English Zimbabwe English Americas United States English Antigua & Barbuda English Argentina español Aruba English Bermuda English Bolivia español Brasil português Bahamas English Belize English Canada English Canada français Chile español Colombia español Costa Rica español República Dominicana español Ecuador español Greenland English Guadeloupe English Guatemala español Honduras español Haiti English Jamaica English Cayman Islands English Martinique English México español Nicaragua español Panamá español Perú español Puerto Rico español Paraguay español El Salvador español Trinidad & Tobago English Uruguay español Venezuela español British Virgin Islands English U.S. Virgin Islands English Asia Pacific الإمارات العربية المتحدة العربية Armenia English American Samoa English Australia English Azerbaijan English বাংলাদেশ বাংলা البحرين العربية Fiji English Micronesia English Georgia English Guam English Indonesia Indonesia ישראל עברית India English ભારત ગુજરાતી भारत हिन्दी ಭಾರತ ಕನ್ನಡ ഇന്ത്യ മലയാളം भारत मराठी இந்தியா தமிழ் భారతదేశం తెలుగు العراق العربية الأردن العربية 日本 日本語 Kyrgyzstan English Cambodia English 대한민국 한국어 الكويت العربية Kazakhstan English Laos English لبنان العربية Sri Lanka English Marshall Islands English Myanmar (Burma) English Northern Mariana Islands English Maldives English Malaysia English Nepal English New Zealand English عُمان العربية French Polynesia English Papua New Guinea English Philippines English پاکستان اردو Palau English قطر العربية المملكة العربية السعودية العربية Singapore English ไทย ไทย Tajikistan English Turkmenistan English Tonga English Türkiye Türkçe 台灣 中文 Uzbekistan English Việt Nam Tiếng Việt اليمن العربية South Africa