GitHub co-pilot CLI as a tipping point in hallucination-sensitive legal work

The speaker, Dan, frames legal work as "a very hallucination-sensitive context" where the cost of error is high. He describes encountering a clear turning point when using GitHub co-pilot CLI: the tool repeatedly handled tasks he expected to fail, so he pushed it with increasingly complex prompts until it produced an analysis that referenced source documents but contained false claims about timings. To regain verifiable trust he built a skill that 'embeds inline screenshots' of the relevant document passages; that feature he calls 'eyeball'.

The embedded screenshots let Dan "verify the AI outputs with my eyeball": when the model flagged an 'indemnity clause' he could see the exact passage driving the claim and the highlighted text that underpinned the conclusion. That visual verification converted suspicion into actionable confidence because "we act on what we're getting out of these tools" and therefore need a "solid basis and truth" behind outputs.

Adoption followed quickly: "Co-pilot CLI bred like wildfire over a month or two, and now we have every lawyer in GitHub working in the terminal," demonstrating unexpected, rapid uptake. His concrete recommendation is practical and direct: 'Don't hold back on what you asked the tool to do, because you'll probably be surprised.'

[inferred from description of screenshots] The visual feature likely shows inline screenshots of source documents with highlighted excerpts synchronized to the model's assertions, enabling fast human verification.

Final identity line: "I'm Dan, a lawyer at GitHub, and now I'm an open source developer."