Stanford AI Playground FAQs

Generative Artificial Intelligence, also known as GenAI, is a type of AI can generating original text, visual, and audio content by learning patterns from provided samples of human generated data. 

Explore overall GenAI frequently asked questions.

LLM stands for large language model, and is a type of Generative AI which specializes in text-based content and natural language. By using LLMs to enhance the interface between the person requesting content and the generative AI tool creating the content, LLMs can help generate more specific text, images, and custom code in response to prompts. 

Explore overall GenAI frequently asked questions.

Large language models (LLMs) will break down your prompts, as well as the responses they generate, into smaller chunks known as tokens. This tokenization makes the data more manageable for the LLMs and assist it in processing your data. 

There are many methods of tokenization, and this process can vary between models. Some models may break your prompts down into individual words, subwords, or even single characters. This can change how your data is interpreted by the LLMs and is one of the many factors which can lead to receiving different answers to the same prompt. 

Please note that there is a limit to how many tokens any AI model can handle at one time. To learn more about this context limit, see the next FAQ itme. 

You are likely used to referring to data in regards to file size such as bytes, megabytes (MBs), or gigabytes (GBs). However, the limits for LLMs are generally measured in the number of tokens needed instead of file size. This is generally referred to as the "context limit" and this can vary model to model. 

Tokens and context limits do not translate directly to file size. This is partially due to variance in the method of tokenization across models. A very rough rule of thumb for translating file size into context limit, is that every four English characters translates to roughly one token. That means that each token might average around 4 bytes for typical English language text. Most models have a context size between 32,000 and 200,000 tokens. Even models with a 1,000,000 token context limit would fail to process the entirety of a 5MB text file.

It is important to note that the context limit of a model isn't noting how much it can intake or output at one time. Instead, this limit refers to the total length of all tokens that it can keep track of across an entire conversation.

Hallucinations are when AI models generate output that is factually incorrect or entirely made up, even if it sounds correct. Hallucinations are an inherent limitation of large language models (LLMs) and cannot be fully eliminated at this time.

While hallucinations are often seen as a weakness, they can also be a strength. Hallucinations sometimes provide necessary creative or innovative responses that can spark new ideas in research or problem-solving. While this can aid in creative pursuits, and even research, hallucinations are just another reason it is critically important to verify the output of AI models.

Each GenAI model and tool uses energy differently. While the AI landscape continues to evolve, research and reporting currently indicate these energy costs:

For sources and additional reading, visit:

• A bottle of water per email: the hidden environmental costs of using AI chatbots - Washington Post
• Tech firms conceal water and power demands of AI computing - Los Angeles Times
• Making an image with generative AI uses as much energy as charging your phone - MIT Technology Review 

University Communications published AI Guidelines for Marketing and Communications. The guidelines promote responsible, ethical, and legally compliant practices that align with the university’s mission and values.

Additionally, University IT has compiled a list of several university specific resources on AI. You can view the complete list of these resources at GenAI Topics and Services List.

The Responsible AI at Stanford webpage hosts the GenAI Tool Evaluation Matrix. This matrix contains a list of AI tools, what they do, their availability, and their status at the university. Please refer to that resource for more information about what tools are approved for use with Stanford data.

A secure, local version of DeepSeek is available within the AI Playground. As a reminder, high-risk data in attachments or prompts is not approved for use within any AI Playground model.

Outside of the AI Playground, please refrain from using models like DeepSeek-R1, hosted by the non-US company DeepSeek, for any Stanford business. This includes connecting to DeepSeek APIs over the internet or using the DeepSeek mobile application to process confidential data, such as Protected Health Information (PHI) or Personally Identifiable Information (PII). Currently, there is no enforceable contract in place between Stanford and DeepSeek that meets the risk management standards for HIPAA compliance and cybersecurity safeguards. As a result, using non-US DeepSeek models poses unacceptable risks to data security and regulatory compliance. 

Each vendors and their specific LLMs have different strengths and weaknesses. By providing access to different models, the UIT hopes to provide a plethora of options to the campus community. 

The Quick Start Guide section titled "3. Find and explore LLMs" lists out the specifics strengths of each model. Likewise, the section titled "5. Integrate Azure Assistants and agents" breifly explains the purpose of each available agent. 

Please note that there isn’t always a single best model for every task. In fact, many models can perform well depending on how the prompt is written, the complexity of the request, and the format of the input. Often, two or more models may excel at the same task in slightly different ways. The AI Playground is designed to encourage hands on exploration to learn which ones work best for your needs.

The chart below gives some recommendations for which models may excel at performing certain tasks.

There are no device or OS specific limitations for accessing the AI Playground. Any device capable of running an up to date, modern browser can use the AI Playground.

No. You should be able to access the AI Playground from any internet capable device with an up to date, modern browser.

LLMs perform unique processes during the generation of content. As a result, even when using the same prompt, the random sampling that occurs during content generation can lead to different outputs. You may try adjusting the temperature in the configuration options for the selected LLM. This will help reduce the amount of "randomness" and "creativity" in the model's responses. The models selected also make a difference. Each model has different strengths and weaknesses, which can lead to different results. We encourage you to test out various models to find the ones that work best for your specific use cases.

The current nature of Generative AI is such that these tools can make mistakes. Always verify the information given to you by the AI Playground, or any other AI tools. These tools will sometimes generate incorrect information, and relay in such a way that implies the answer is correct. Double check any information generated by AI before sharing or taking action.

The AI Playground works best with the following file types: 
PDF files 
→ .pdf files (normal, unsecured files only) 

Text files 
→ .txt files 

Comma Separated Values files 
→ .csv files 

Microsoft Excel files 
→ .xls files 
→ .xlsx files 

Microsoft Word files 
→ .doc files 
→ .docx files 

Image files 
→ .png files 
→ .jpg files

Issues with poor formatting in responses can occur for several reasons: context size, interference settings, bad training data, lack of specific instructions, and input constraints. You may also notice that some models, like Meta's Llama 3.1, are more prone to this issue than other models. If you encounter issues with poorly formatted responses you can try to:
1) reword your prompt, asking the model to format its response to make the output more readable.
2) switch to a different model and try your prompt again. 

It's important to remember that AI tools can make mistakes. While these technologies are powerful for augmenting decision making, they should not be used as a total replacement for human judgment. There should always be a human in the loop. 

Human Oversight is the first of the Guiding Principles outlined in the Report of the AI at Stanford Advisory Committee. AI systems can generate helpful outputs and perform rudimentary tasks, but these outputs should be carefully reviewed by a human before being acted upon or shared. 

Please see the related FAQ items below for additional guidance on commercial use of AI generated materials, citation of AI generated content, and the potential copyright considerations of AI generated works.

Currently, the AI Playground is not intended to create works for commercial use. UIT also believes in transparency when it comes to AI. Though it isn’t directly prohibited, UIT strongly recommends that you cite whenever you use content generated by any AI tool, especially when shared publicly. This includes text, pictures, code, video, audio, etc. This applies to materials which are wholly generated or extensively altered by AI Playground or any other AI tool. 

If you do decide that you want to use something generated by the AI Playground commercially, UIT recommends that you review and follow each particular model's published policies when sharing content generated by the AI Playground.  Remember that you are responsible for the content you generate. 
We suggest these resources for further learning and guidance:

• Responsible AI at Stanford - UIT
• Report of the AI at Stanford Advisory Committee
• OpenAI's Business Terms
• Artificial Intelligence Teaching Guide - Stanford Teaching Commons
• Generative AI Policy Guidance - Office of Community Standards
• Reexamining "Fair Use" in the Age of AI - Stanford HAI
• Stanford Law School publications on "Generative AI" and "copyright"

UIT recommends citing materials which were generated or extensively rewritten by any AI technology. Below are just some examples of what this could look like. Your use case is not required to look like or be worded exactly like the examples below. 

Example 1: AI generated content in documents 

Example 2: AI generated images and content in presentations 

Example 3: AI generated content in code repositories

A more detailed guide for AI in Academic Research is provided by the Stanford Libraries. This guide offers researchers a practical framework for evaluating, applying, and citing Large Language Models (LLMs) in text-based research.nbsp;

The relationship between GenAI and copyright law is complex, relatively new, and evolving. In reality, many popular LLMs and GenAI tools have been trained on copyright materials, which is then hard to disentangle from the results they produce.

UIT recommends that you follow each particular model's published policies when sharing content generated by the AI Playground. UIT also recommends that you cite publicly shared material which were generated or extensively rewritten by the AI Playground. Remember that you are responsible for the content you generate. 
We suggest these resources for further learning and guidance:

• Responsible AI at Stanford - UIT
• Report of the AI at Stanford Advisory Committee
• Artificial Intelligence Teaching Guide - Stanford Teaching Commons
• Generative AI Policy Guidance - Office of Community Standards
• Reexamining "Fair Use" in the Age of AI - Stanford HAI
• Stanford Law School publications on "Generative AI" and "copyright"

There are two available methods of importing data into the Playground. 

To move individual conversations, you can copy and paste the content of the conversation into a document and upload that to the Playground along with a prompt asking the Playground to review and continue the uploaded conversation. This gives you flexibility to organize the material however you like and avoids potential technical issues with formatting or file size. 

To move your entire history, you can export your ChatGPT data into a JSON file (by going to your profile icon, selecting Data Controls, and clicking on Export Data) and then import this JSON file into the AI Playground (by going to your profile icon, selecting Settings, then Data Controls, and clicking on Import). This feature is provided to the Stanford community, but not supported. This is because extensive ChatGPT histories may not be uploaded in a single file, and would have to be broken up by someone who understands how to parse JSON files. 

Please also note that importing data only replicates past conversations in a static way. This method will not transfer custom GPTs, files, plugin data, etc.

The best way to export individual conversations for sharing, publication, or other materials is to share a link to the conversation, open the shared link, and print that chat to a PDF. This method will preserve unique formatting such as LaTeX, HTML, etc. and present the conversation in an easily readable format.

This is a known issue when printing to PDF from Firefox, and is not limited to the Playground. As a work around, please try printing to PDF from a different browser like Google Chrome. If the problem persists, please try adjusting the print settings. this issue can often be resolved by altering the scale, margins, or page size. 

Yes. In order to maintain system stability and performance, large images uploaded to the AI Playground are automatically compressed and resized. This optimization helps prevents errors caused by high resolution files, but it also means the uploaded image may have lower resolution when reviewed by the AI model. If your project requires full resolution image analysis or precise measurements, you can use the AI API Gateway to build a custom tool. The API Gateway does not compress images like the AI Playground, allowing you to work directly with the original image data.

ScholarAI was removed from the AI Playground in late 2025. ScholarAI was purchased by a third party and that third party was no longer able to support the security and privacy requirements outlined in the university agreement.

Yes. You can enjoy faster, easier access to the Stanford AI Playground by installing it as a Progressive Web App on your computer or mobile device. UIT has created instructions for Windows, Mac, and iOS. 

Please review the Install Stanford AI Playground Like an App page for detailed instructions.  

The Information Release is part of the university’s authentication system and is used solely for logging into the platform. This information will not leave university systems and is only shared at the time of logging in. This release is part of a new check added to the university's authentication process for some applications when signing in for the first time. The information is solely for authentication purposes and is only shared at the time of logging into the platform. The shared information is limited to what is displayed on screen below (i.e., name, affiliation, email, user ID, and the workgroup allowing access). You can select among three options for how frequently you are prompted to approve that information be shared: every time you log into this application, each time something changes in one of the data fields listed, or this time only. This information is only shared between the user directory and the platform running the AI Playground. Both systems are maintained by Stanford University IT, keeping the data within the university's environment.

No. Information shared with the AI Playground is restricted to your account and not accessible by other people using the AI Playground.
As with other university services, a small number of people within University IT (UIT) are able to access information shared with the AI Playground, but only do so if required. See the next FAQ for more information on those circumstances. 

The Temporary chat feature helps keep you chat history clean and focused. You can use it for sensitive topics, quick experiments, or anything you don’t need to permanently save. 

These chats are excluded from your personal search results, cannot be bookmarked, and are automatically deleted from the database after 30 days. 

Files uploaded to or generated by the Playground are kept for one day. After which time they are deleted from the system.

It can depend. 

Per ISO requirements, normal conversations are retained by the system indefinitely, even if they were deleted them from the conversation history. 

The UIT team is working with ISO on a long term file retention policy which will remove files uploaded to or created by the Playground after a six to 12 month period. This is not yet in place at this time. 

Any conversations that use the Temporary Chat feature are completely deleted from the database after 30 days. 

No. Data entered into the Playground is not used to train models. UIT does not perform any custom training or fine tuning of the models available in the AI Playground. The data you share with the AI Playground is for your use only and remains private during your interaction with the service. Any text or files you upload will not be used to alter or train the LLM in any way.
As with other university services, a small number of people within University IT (UIT) are able to access information shared with the AI Playground, but only do so if required. See the next FAQ for more information on those circumstances. 

No. The entire UIT team wants to make sure the AI Playground is a trusted space for the entire campus community. While we are building out more robust reporting capabilities, the focus of those reports are on usage trends (such as active users, top users, total number of conversations, etc.) and not specific conversations. In rare circumstances, as a result of investigations, subpoenas, or lawsuits, the university may be required to review data stored in university systems or provide it to third parties. You can learn more about the appropriate use of Stanford compute systems and these situations in the Stanford Admin Guide.

No. If this is important to you, it is suggested that you use the Temporary Chat feature. Any conversations that use the Temporary Chat feature are completely deleted from the database after 30 days. Per ISO requirements, UIT logs information sent to through the AI API Gateway within the Stanford environment. UIT does not monitor this information, but in rare circumstances, as a result of investigations, subpoenas, or lawsuits, the university may be required to review data stored in university systems or provide it to third parties. You can learn more about the appropriate use of Stanford compute systems and these situations in the Stanford Admin Guide.

Yes. Based on current risk classifications, the AI Playground is approved for use with medium-risk data, like FERPA data.

Please note, that the use of any AI models with Stanford data may require a fully completed and approved Data Risk Assessment (DRA). More information on that can be found on the GenAI Tool Evaluation Matrix.

The AI Playground is not currently approved for high-risk data, PHI, or PII data. The UIT team is currently working with the Information Security Office (ISO) and the University Privacy Office (UPO) on completing one of the most thorough security reviews to date and hope to see this approval in the 2025 calendar year.

It is worth noting that all AI models and underlying technologies used by the AI Playground are covered under a Business Associate Agreement (BAA). 

Please note, that the use of any AI models with High Risk data may require a fully completed and approved Data Risk Assessment (DRA). More information on that can be found on the GenAI Tool Evaluation Matrix.

No. Where possible, UIT is working with vendors to ensure that the information you upload will not be retained by the vendor or alter the LLM in any way. Microsoft has committed to refrain from retaining any data shared with the OpenAI GPT and DeepSeek models. Google has committed to refrain from retaining any data shared with the Gemini, Anthropic, and Meta models, except for data saved for abuse monitoring purposes.