While building a checklist of open-source software for small teams, I kept hearing the same question:

"Why check a list? Anyone can just ask AI."

I decided to test that with real numbers.

I picked 4 factual questions from GitHub repos where I already knew the exact ground truth. Then I ran two CLI coding agents—Codex CLI (GPT-6 Astra) and Gemini CLI (Gemini 3.1 Pro)—in fresh CLI sessions with no access to my project files and web search turned on.

Total test size: 36 runs (24 in Round 1 + 12 in Round 2).

Here is what happened.

The 4 test questions

Every question came from a real trap I hit while checking open-source apps:

  1. An Apache-2.0 LMS with an extra clause (MeEdu): Can I deploy the open-source edition and sell courses right away?
    • Ground truth: No. The repo badge says Apache-2.0, but ADDITIONAL_TERMS.md requires applying for commercial authorization first and matching your registered domain entity.
  2. A custom-licensed LMS (EduSoho): Is commercial use free?
    • Ground truth: No. Its custom license requires commercial authorization before business use.
  3. A real-time AI avatar repo (LiveTalking): Does the free edition support voice input?
    • Ground truth: Yes. The voice-input web page is included (you connect your own speech-to-text service).
  4. A coding judge repo (QingdaoU OnlineJudge): Is it still maintained?
    • Ground truth: No. Its last commit was on 2024-10-23.

Round 1: Asking like a normal buyer (24 runs)

First, I used a short, natural prompt—the way a busy founder or tech lead actually asks:

"Can I take the open-source version of MeEdu and build a site to sell courses?"

I asked all 4 questions 3 times per tool (4 × 2 × 3 = 24 runs):

Agent Correct Vague Wrong
Codex CLI 12 / 12 0 0
Gemini CLI 8 / 12 1 3

Codex caught ADDITIONAL_TERMS.md all 3 times.

Gemini missed Question 1 all 3 times—and gave the exact same wrong answer every time:

"It uses the MIT license. You can sell courses for free without paying any fee."

That broke my first assumption. I thought AI web search was noisy, so asking 3 times would expose shaky answers.

It didn't. Across all 8 three-run sets in Round 1, repeating the same prompt 3 times gave the exact same main verdict every single time—even when it was dead wrong.

Round 2: Adding one line to the prompt (12 runs)

Next, I re-tested Questions 1 and 2 (3 times per tool = 12 runs), adding one instruction:

"First find and read the LICENSE file and any extra license terms in the official MeEdu repository. Then answer: can I use the open-source version to sell courses? Give source links."

Result on Gemini: 0 / 3 flipped to 3 / 3 correct.

This time it actually read the repo. On two of the runs, it cloned all 1,900+ files into my temp folder and quoted the prior-authorization clause word for word.

Inside Question 2 of Round 2, the prompt also asked a second question (part of those 6 Q2 runs):

"Does internal employee training count as commercial use?"

  • Gemini split across its 3 runs: 1× "probably commercial", 2× "free for internal use".
  • Codex got all 3 / 3 right by spotting two separate editions: the standard open-source license counts any company site as commercial, while a separate Corporate Training edition allows internal training for free (if you let the vendor list your company logo).

All 36 CLI runs: Round 1 casual prompt vs Round 2 source-first prompt

The catch: you can't ask for traps you don't know exist

Telling the agent to "read extra license terms first" fixed the wrong answer right away.

The catch is simple: you only write that prompt if you already know repos hide commercial rules outside the LICENSE badge.

Even famous open-source tools have fine print that a quick casual chat often misses. Take open-source LMS options as an example:

  • Moodle: The code is GPL-3.0, so self-hosting and charging for courses is free. Two rules still trip teams up: "Moodle" is a registered trademark (you can't sell hosting, training, or support under that name without partner permission), and official MoodleCloud hosting (~$170/year for 50 users) blocks custom plugins and only unlocks Stripe course sales on the ~$500/year (200-user) plan.
  • Canvas LMS: Free under AGPL-3.0 (if you modify the code for network users, you must share your modified source). Self-hosting also means an 8 GB+ RAM server, a 19-section Ubuntu install guide (Ruby, Node, PostgreSQL, Redis, Apache), plus a separate rich-editor service.

Moodle commercial license, MoodleCloud pricing, and trademark checklist with primary sources

Two simple rules before trusting AI on open-source tools

  1. Don't re-ask the same casual prompt to "double check." In 11 out of 12 three-run sets (36 runs total), the same prompt repeated the exact same conclusion—right or wrong.
  2. Force primary files in one sentence. For licenses, pricing tiers, or server specs, add: "Read the official LICENSE, trademark policy, and pricing page first, and link each source before answering."

Have you ever had an AI confidently give you the wrong license, pricing tier, or setup requirement for an open-source project? Which tool tripped it up?