Introducing
SafeGPT

Testing & Monitoring tool for
Large Language Models

Share it with your friends and colleagues to move up the list!

For users of LLMs

Say goodbye to errors, biases & privacy issues in LLMs

Use SafeGPT to ensure the safety of ChatGPT

Our browser extension allows you to identify wrong answers, reliability issues and ethical biases for any LLM

Continuously monitor your LLM results

Track your LLM system's performance at a glance

Our dashboard includes alerting & root-cause analysis capabilities,
and you can filter by queries, topics and users

For developers of LLM apps

Quality Assurance platform for LLMs

LLMOps platform to evaluate & debug your LLMs

Compare the quality of different LLMs by creating tests that are specific to your business, and debug LLMs with Human Feedback and Explainable AI methods

Why?

We protect against critical AI Safety risks

Hallucinations

Factual errors made by LLMs lead to serious distrust in the model, and can have severe financial implications.

Privacy issues

LLMs can leak private and sensitive data that you want to protect, and expose you to legal & reputational damage.

Toxicity issues

LLMs can generate discriminatory or ethically biased content, posing reputational risks & harming society.

Robustness issues

Answers can vary depending on LLM providers. SafeGPT compares the answers to check their robustness.

How it works

Our methodology is grounded in research

We curate the latest AI Safety research in this awesome GitHub repo.

Human feedback

We provide user interfaces for humans to debug, explain & evaluate LLMs

External data

For fact checking, we integrate external data sources - or bring your own data, securely

Adversarial testing

We develop purpose-trained LLMs and tailored prompts to evaluate other LLMs

Ethical AI

We provide state-of-the-art detection methods for ethical biases & sensitive personal data

Metamorphic testing

We provide a framework to evaluate if LLMs are robust to changes in the input data

Open-Source

Our Python library for AI testing is open-source, so you can extend it with your own tests

Join the waitlist for free

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.

FAQ

Is SafeGPT a real product?

Yes, we have working alpha versions of what we describe above: a browser extension for users to monitor ChatGPT results, and a Quality Assurance platform for developers to evaluate & test LLMs.

We'd be excited to show you the alpha version after you join the waiting list. We value feedback from real users & developers to direct future development, before releasing the public beta.

This is a new tool we're building in the open, and we're interested in your feedback to prioritize!

How can I get access to SafeGPT?

Just join the waitlist! You will get notified shortly with all the details.

Is it free?

There will be a free tier so you can try first for personal use. For premium & enterprises solutions, we’ll get in touch to discuss pricing & ROI after you join the waitlist.

Why did we start building SafeGPT?

We are a team of engineers & researchers on AI Alignment, Quality & Safety who’s been working on this topic since 2021. While we are all excited about the LLM revolution, we acknowledge the safety risks involved.

We believe crucial to have independent third-party evaluations to assess the safety of generative language models. These evaluations, conducted by separate entities from the developers of LLMs, provide important checks and balances to ensure responsible regulation of the system.

To solve this difficult and open problem, we're investigating and developing multiple methods, including metamorphic testing, human feedback, constitutional AI, benchmarks with external data sources, and Explainable AI methods.

How will my personal data be processed by SafeGPT?

Data privacy remains at the core of Giskard. We follow local regulations and offer to host your data in specific regions (EU/US). For enterprises, we offer on-premise installation so the data stays on your servers.