Accuracy

Surfboard is powered by large language models (LLMs) - the same category of AI that powers tools like ChatGPT and Claude and Gemini. These models are remarkably capable, but they are not infallible. Understanding their limitations helps you get the most out of Surfboard and know when to apply extra scrutiny.

What is a "hallucination"?

The term hallucination refers to when an AI model generates information that sounds plausible but is factually incorrect, outdated, or entirely fabricated. This is not a bug in the traditional software sense - it is a known characteristic of how language models work.

LLMs generate responses by predicting what text is most likely to follow a given prompt, based on patterns learned from vast amounts of training data. This process is powerful, but it does not involve "knowing" facts the way a database does. As a result, a model can produce a confident-sounding answer that is simply wrong.

How Surfboard reduces this risk

Surfboard is designed to ground its responses in real data from your connected sources - your email, calendar, Slack, GitHub, and other tools. Rather than relying purely on the model's training data to answer questions about your work, Surfboard actively queries your sources and cites what it finds.

This approach significantly reduces the risk of hallucination for work-related queries. When Surfboard says "you have a meeting with Peter at 11:30am," it retrieved that from your calendar - it did not guess.

However, this does not make Surfboard immune to errors. A few things to keep in mind:

• Synthesis can introduce inaccuracies. When Surfboard combines information from multiple sources, it may draw incorrect inferences or mischaracterize a relationship between two facts.
• Source data can be incomplete. If a connected source has missing or stale data, Surfboard will reason from that incomplete picture.
• General knowledge questions are higher risk. When you ask Surfboard something that requires drawing on its underlying model's knowledge (not your sources), it is more likely to be imprecise.
• Citations are not a guarantee. Surfboard will cite its sources, but you should still verify important claims - particularly for decisions with real-world consequences.

What to do if you see an incorrect response

Correct it in the conversation. Simply tell Surfboard what is wrong. For example: "That's not right - the meeting was canceled." Surfboard will incorporate the correction and update its response. This is the fastest path forward in most cases.

Correct it directly in the document. If Surfboard has written an incorrect result into a document, you can edit the document content directly. You do not need to tell Surfboard to make the change - you can just open the document and fix the inaccuracy in place.

Check the citations. Surfboard attaches citations to the facts it pulls from your sources. If something looks off, click through to the original source to verify.

Ask Surfboard to show its work. If you are unsure how Surfboard arrived at an answer, ask it to explain its reasoning or identify which sources it used. This can help you spot where an error crept in.

Do not rely on Surfboard alone for decisions. Surfboard is a powerful tool for synthesis, summarization, and staying organized - but for consequential decisions (legal, financial, medical, or otherwise), always verify with authoritative sources or qualified experts.

Ongoing improvements

We take accuracy seriously and are continuously working to improve Surfboard's reliability. If you encounter a response that seems consistently wrong or misleading in a way that is not explained above, please use the feedback button to let us know. Your reports directly inform how we improve the product.