Ask ChatGPT for small-business grants in your province and you'll get a fast, tidy list. Some of it will be real. The problem is you can't tell which part.
What AI assistants are genuinely good at
This isn't an argument that AI is useless. It's a fast, capable research assistant. In seconds it produces a plausible starting list of program names, funders and dollar ranges, and it's excellent for orientation, thinking through eligibility angles, and drafting the language of an application. With real-time web browsing, a program announced last week can even show up. For getting oriented and writing, reach for it.
Why it can't tell you what's open
An AI language model is trained up to a fixed point in time. As the reference definition puts it, "any information about events after this date is absent from the model's training data." A grant that closed, changed, or reopened after that cutoff is invisible to the model — and, in the same source's words, cutoffs cause the model to generate "confident but false statements" to fill the gap. So when you ask about a program, you may get a real name attached to a status that's a year or two out of date, delivered with total confidence.
The most common failure is a program that stopped taking applications a year or two ago, still described with total confidence and nothing in the answer's tone to warn you. An AI doesn't tell you when it's wrong.
"But it browses the web now"
Real-time browsing helps, but it doesn't fix the core problem. Browsing retrieves and faithfully re-cites the same expired listings that rank everywhere else: stale sources, faithfully cited. A citation to a dead page is not proof the program is open. If what the AI pulled is a 2024 deadline page, you get a 2024 deadline delivered in a 2026 voice.
This isn't hypothetical, and it's not only a grant issue. In Moffatt v. Air Canada, a British Columbia tribunal found a company's own AI chatbot gave a confident answer that contradicted the real policy, and held the company liable. The tribunal's point applies to any AI answer: "there is no reason why [the user] should know that one section is accurate, and another is not." You can't self-filter the wrong grant out of an AI's list, because nothing marks it as wrong.
How to check an AI's grant list
- Take the program name, not the status. Use the AI to learn what exists; confirm whether it's open somewhere that verifies.
- Find the official government page. A real program names its exact administrator and links its own page. If the AI can't, be skeptical.
- Watch for confident specifics with no source: exact dollar figures and deadlines stated in the present tense but with nothing to click.
- Check the date. "As of 2026" from a model trained earlier is not the same as verified this month.
Where BBN Labs fits
This is the exact gap we built for. Every program in our list is checked with its official government source, carries the date we last verified it, and a program whose deadline has passed is flipped to "Closed, monitoring for reopening" — so a closed program is never dressed up as an open one. Use ChatGPT to get oriented and to draft; then get live, dated program status from a source that actually verifies it. And when a program fits, we do the work the funding pays for. The two are complements: AI for speed, verification for truth.
The honest bottom line: ChatGPT is a great place to start and a dangerous place to stop. It can name the grant. It can't tell you the grant is still there.