I tried teaching ChatGPT the Canadian Income Tax Act. It failed miserably.

Last week, I was experimenting with ChatGPT for tax calculations and interpretations. Even with detailed prompts and guidance from someone who actually works in this space, the AI consistently stumbled on common Income Tax Act (Canada) (the “Income Tax Act”) provisions.

If a machine trained on the entire Internet can't properly interpret our Canadian tax laws, what does that say about expecting ordinary Canadians to comply with them?

Here’s my take on the current state of AI for lawyers: 

➡️ Great for: First iteration clause drafting and general summaries of documents for client communication. 

➡️ Decent for: Synthesis of a large number of documents and general contract analysis with human oversight (i.e., line-by-line review), and general research (again, with direct oversight and review of findings).

➡️ Terrible for: Contract interpretation, statutory interpretation (especially Canadian tax law), understanding how various acts work together.

Don't get me wrong, AI is going to transform the legal practice in incredible ways. It can do a good job interpreting a basic act, like the Alberta Business Corporations Act (Alberta), but the Income Tax Act has achieved something remarkable: it's so unnecessarily complex that even AI throws up its hands (repeatedly, no matter how many reminders I give it).

Maybe the problem in this case isn't AI's limitations. Maybe it's that we've created laws so absurdly complex that neither humans nor machines can navigate them without a PhD in bureaucratic nonsense. 

We expect Canadian taxpayers to comply with legislation that stumps machine learning. If AI trained on billions of data points can't interpret our tax laws correctly, maybe the problem isn't the technology—it's the laws themselves.

If we're embracing AI in legal practice (and we should), we need to be honest about where it excels and where it falls flat. More importantly, we need to ask why our legal system is so complex that advanced technology can't help ordinary people understand their obligations.

At Outsiders Law, we use AI extensively to improve our service delivery, but always with tight human oversight. The key is knowing when to trust the technology, when and how to prod it for better answers, and when to step in.

Where have you seen AI hit the wall in legal applications? Should we be building smarter AI, or writing simpler laws?