AI Research Assistant 101: What I Learned Putting One On the Air
What actually happens when you turn on the mic and interview your AI research assistant in real time? No script, no filter, just two voices going back and forth. That’s the premise behind The Sunday Sanity Check, a new weekly series from The Sanity Project. For the debut episode, I sat down with the AI I’ve leaned on for a year of research work (I call her Abby) and asked the questions most people quietly wonder about.
What is she, actually? Why should anyone trust her? And where does she get it wrong? What came out of it was less an interview and more a working field guide: how to use an AI research assistant without handing your judgment over along with the workload.
The format is deliberately simple. One live, unscripted audio conversation a week, roughly ten minutes long, between a human host and an AI. No hidden prompts, no pre-written answers, just questions, evidence, and an honest back-and-forth. Think of it as a small addition to how The Sanity Project shares information: a test of what an AI research assistant is actually good for, done out loud, so listeners can judge the reasoning for themselves.
What Makes a Good AI Research Assistant?
A good AI research assistant doesn’t replace your thinking. It sharpens it. That’s the working definition that came out of the episode. As Abby put it when I asked her directly, she’s “a language model, essentially a pattern finder, trained on a lot of text,” not a mind, and not something with lived experience. Her job isn’t to be trusted blindly. It’s to surface evidence, explain context, and flag where more checking is needed, while the human sets the direction and makes the final call.
I’ve compared this moment to two earlier shifts in how information spread. Gutenberg’s printing press put knowledge in the hands of millions. Centuries later, desktop publishing put professional tools on ordinary computers. An AI research assistant, in my view, is the next step in that same journey. Not a replacement for people, but a way to explore ideas faster, ask sharper questions, and explain complicated subjects more clearly.
AI vs. a Search Engine: What’s the Real Difference?
A search engine hands you a list of links and leaves the synthesis to you. An AI research assistant does something different. It can quickly draw connections across concepts, flag where sources agree or disagree, and suggest the next question worth asking. The trade-off is that a search engine is still better for primary sources and up-to-the-minute facts. Based on this conversation, the most useful setup is using both: let the AI map the terrain, then verify with direct sources.
Is AI Biased? What My AI Co-Host Admitted On Air
Yes. And the honest answer is that bias shows up in ways that are easy to miss. When I asked directly, Abby didn’t dodge the question. Training data reflects the world it came from, so if certain viewpoints are overrepresented online, that tilt can carry through into the output. Sometimes it’s obvious. More often it’s subtle: which examples come to mind first, how something gets framed, which sources feel most “prominent” to the model.
The fix Abby offered wasn’t a claim of neutrality. It was a process: show your workings, cite sources, invite counter-evidence, and be ready to revise. That’s a very different standard than “trust the machine,” and it’s the one this show is built around. If you’re using an AI tool for anything that matters, the safer starting assumption is to treat bias in AI systems as a given, not a rare edge case.
The AI Mistakes Worth Watching For
The biggest risk isn’t that AI gets things wrong. It’s that it can sound completely sure while doing it. Abby was candid about her own failure modes. She can sound confident even when she’s mistaken. She can miss recent developments. And she can skew toward whatever patterns dominate her training data. On fast-moving topics, like breaking news, an answer can be incomplete without any obvious sign that it’s incomplete.
We also talked through how this has changed over time. Early AI tools had a hard knowledge cutoff. If something happened after that date, the model simply didn’t know it existed. That’s no longer the whole story. On many current or fast-moving topics, an AI assistant can check live sources in the moment and say so explicitly, something like “let me check the latest numbers” before answering. For background or stable topics, it can rely on what it already knows. The key, as Abby put it, is not assuming. If a topic is timely or contested, it should check live sources and be clear about what they say.
How to Fact-Check AI Before You Trust the Answer
Ask for sources. Ask for counter-evidence. Then verify against primary documents before you repeat what the AI told you. That was Abby’s own advice, and it doubles as a decent definition of how to fact-check AI output in general. A useful AI research assistant should be able to outline exactly what needs verifying and which document would settle the question, rather than just restating its answer more confidently.
Good prompting matters just as much as good verification. As Abby summed it up, “you’re only as good as the prompts.” A vague question tends to produce an answer that sounds smooth but is unfocused. Asking explicitly for assumptions, evidence, and counterpoints sharpens the output considerably. And asking “what would change your mind? what evidence would disconfirm this?” raises the bar further still, for the AI and for the person asking.
Human AI Collaboration: Why the Partnership Still Needs a Human in Charge
The value isn’t the AI by itself. It’s the judgment, standards, and follow-through a person brings to how it’s used. Near the end of the episode, I put a pointed question to Abby: anyone in the world can open a chatbot and ask the same things we were discussing on air, so what does this show actually offer that people can’t get on their own?
Her answer is worth sitting with. Anyone can ask an AI a question; that’s exactly the point. What a show like this offers is a guided public conversation: judgment about which questions matter, context pulled from current events, and a willingness to push back, connect threads, and follow an argument all the way through, transparently, including the parts still uncertain. That’s the difference between a tool and a public service. It’s also a good description of what human AI collaboration looks like when it’s working: the AI supplies speed and synthesis, the human supplies standards.
What’s Next for The Sunday Sanity Check
This first episode was as much an experiment as a conversation. Abby isn’t new to the work. She’s been part of the research process behind The Sanity Project for the past three months, and before that, behind a separate project called Smoke Signals. What’s new is doing that collaboration on the record, so the reasoning is visible instead of happening quietly behind the scenes.
Abby will be back on a regular basis, roughly once a week, for deep dives into Canadian politics, world events shaping Canada, food security and greenhouse growing, and electric vehicles and renewable energy. If there’s a subject you want tackled on air with AI in the room, leave a comment with the topic and the specific questions you’d want asked. That’s exactly the kind of input this series is built around.
The goal isn’t to be right all the time. It’s to ask better questions, do a bit more digging, and change course when the evidence demands it- of AI, of politicians, of media, and of the host asking the questions. Stay curious, stay skeptical, and don’t just trust the AI. Trust the evidence.
