Databot is not a flotation device

2025-08-28

A yellow, diamond-shaped warning sign on a beach with a Databot icon.

This blog post is about Databot, a new LLM-powered tool for performing exploratory data analysis. Learn more about it here.

In my 30-year career writing software professionally, Databot is both the most exciting software I’ve worked on, and also the most dangerous.

–Joe Cheng, Posit CTO

Databot is fast and effective at slicing its way through data exploration and produces human-like insights that make it feel like you’re fast-forwarding through your analysis. But this speed and efficiency come with risk. Databot, and any tool that relies on an LLM to do complex data science tasks, will make mistakes. It’s good enough at what it does, however, to lull you into a false sense of security, making you think you don’t need to review the code it generates, or that you can accept its observations and conclusions without critical thought.

This post will make those dangers concrete by listing some specific ways we’ve seen Databot misbehave and arguing that, to use Databot effectively and safely, you still need the skills of a data scientist: background and domain knowledge, data analysis expertise, and coding ability.

Databot is a set of fins that can make you swim faster and stronger, not a life vest. It will not keep you from drowning if you don’t already know how to swim.

Misconception: Anyone can do data analysis with Databot

The idea that the barriers to data analysis are now firmly lowered and anyone, regardless of technical skill, can deftly carry out the work of a skilled data scientist is tempting. No need to learn how to code, or understand data cleaning, sample, bias, visualization…Databot will handle it all!

However, that vision of fully accessible data science is not here yet, and Databot is not that kind of tool.

While anyone can use Databot, using it well requires skill. Without the necessary background, it’s easy to end up with an incomplete or misleading understanding of the data.

What goes wrong

LLMs will make errors that require expertise to spot. Here are some examples of mistakes we’ve seen LLM systems and Databot specifically make. We’re always working to address these issues in Databot, but you should be aware of the types of problems that can arise within Databot or any LLM system. We’ve also limited Databot to only the Claude models that we have found to be the least error-prone.

- Hallucinations – Databot may call non-existent functions or use incorrect syntax, which might cause errors or distort your analysis.
- Misleading statistics – Databot may run inappropriate statistical tests, misinterpret results, or even report statistics that it did not run tests for.
- Incorrect plot interpretation – If plotting code runs but produces an incorrect or garbled plot, Databot usually can’t tell that the plot is wrong and tends to claim to see the patterns it expects to see.
- Lack of context – Without complete metadata or documentation, Databot may misinterpret what the data represents and what kind of analyses are appropriate.
- Confirmation bias – LLMs often agree with your assumptions. Databot may reinforce the answer you’re hoping to find.
- Confidently incorrect claims – Databot’s language can sound convincing even when the output is wrong. You should not mistake confidence in language for correctness, because LLMs are typically confident.
The skills you need

Background knowledge

You likely have domain expertise, institutional knowledge, or context for the data that Databot does not have access to. This kind of information is critical to the EDA process and to use Databot effectively.

Background knowledge helps you identify the most useful and relevant questions to ask. It also helps you interpret results accurately. Databot can assist with framing questions, but it works best when you already have a general understanding of what you’re trying to learn.

Databot does not inherently know your intent or your organization’s priorities, but (hopefully) you do. If you don’t define those clearly, it may focus on irrelevant metrics or produce output that misses the point.

Similarly, Databot cannot infer how the data was collected, what it was intended for, or what limitations it has unless you tell it or that information is available in files Databot can access. That context is often essential for drawing valid conclusions.

Your expertise is a safeguard

The more background knowledge you bring, the more effectively you’ll be able to use Databot. It can help you catch mistakes, like a hallucinated function, a misinterpreted variable, or a description of an analysis it didn’t actually perform.

Domain-specific knowledge also helps you recognize subtle issues in an analysis or the data itself. You may know about common pitfalls in your dataset, like typical measurement errors or known causes of missing values. If you’re working with sensitive or proprietary data, you’ll also need to know what should be removed or de-identified, and which files Databot shouldn’t access.

Data literacy

You also still need a grounding in data and statistical skills to use Databot well.

First, you need to understand what the data can and cannot tell you. For example, can you draw conclusions about causal relationships? Is your dataset a sample or a census, and what does that mean for your analysis? Databot won’t always know the answers to these kinds of questions, and if you don’t either, it may run the wrong analysis.

People also often hope their data will confirm a particular belief. LLMs can reinforce this bias. It’s relatively easy to prompt Databot, intentionally or not, into selecting an analysis that supports your assumptions. Without the analytical skill to recognize when something is off or the skepticism to question convenient results, you and Databot can quickly arrive at the wrong conclusions.

Coding ability

LLMs will make coding mistakes, and without coding skills, those errors may go unnoticed and lead to incorrect results. For example, Databot may hallucinate functions that don’t exist or generate code with logic or syntax errors that run but produce the wrong output.

You can interact with Databot using plain language, but it ultimately translates your requests into code. If you don’t understand that code, you’ll miss both the opportunity to get more out of Databot and the ability to catch when it’s doing something incorrect.

As mentioned earlier, you need data analysis skills to assess whether Databot’s approach makes sense. But without coding ability, it can be hard to verify what it actually did. Sometimes, the only way to know whether Databot followed the right process is to read the code it generated and confirm it does what it claims.

Security concerns

Another class of dangers with Databot relates to security concerns. Databot, like any LLM system, is also susceptible to prompt injection attacks, where malicious instructions are hidden in user input or even seemingly benign content like a CSV file. If Databot loads that file and sends it to the LLM, the model might follow those hidden instructions. For assistants like Databot that have access to powerful tools, this problem is even more pronounced because of the lethal trifecta.

You therefore need to be cautious about what files Databot can access and exactly what tool calls it makes. Don’t use Databot in directories with untrusted content unless you’ve verified what’s inside. If you’re in any doubt, make sure you review every code execution request that Databot makes.

Databot is a tool, not a substitute

You might ask: but don’t humans make mistakes too? And yes, humans, even with extensive data science expertise, can also make mistakes in analysis.

The issue is that Databot, or any LLM system, doesn’t just make mistakes–it can amplify them. Databot produces polished, confident output extremely quickly, which can be both difficult to verify and easy to trust. That combination makes it especially risky when used without the right skills.

Databot, and LLMs in general, aren’t at a point where you can abdicate responsibility or suspend skepticism when working with data. Databot is a powerful tool, but it is not a replacement for a data scientist. Using it well requires the same fundamentals that any good analysis does: data literacy, background knowledge, and coding ability. These skills help you catch errors, interpret results in context, and avoid being misled by confident but incorrect claims.

Think of Databot like a talented, extraordinarily fast assistant. It can speed you up, surface new ideas, and help you work more efficiently than ever, but it’s not a substitute for your expertise. The real power comes from working together, combining your own skills with Databot’s to generate more thoughtful insights, faster than ever.