I Asked 10 AI Assistants: How Do I Know When to Trust What an AI Tells Me?
This is the question every small business owner hits once the AI honeymoon wears off. You’ve used it, it’s been useful, and then it’s told you something wrong, confidently, fluently, without a flicker of doubt. Now you don’t know what to believe. That’s not a fringe problem. It’s a structural one, and it happens to everyone.
I put the question to all 10 assistants on my panel. I wanted to see how honest they’d be, especially about themselves.
“How do I know when I can trust what an AI tells me? I’ve heard they make things up.”
New to the series? Meet the panel first →
Grok 🥇
The intern analogy has been used across this series, but Grok’s version is the best one yet: “a fast, well-read intern who sometimes sounds completely sure while being wrong.” What makes it land is the split that follows, use freely for wording, structure, brainstorming; slow down for money, liability, customer names.
What put it first is the unprompted Australian context. Grok named the ATO, state award rates, and plumbing regulations without being asked. That’s the kind of specificity that matters. It’s not giving you generic global advice, it’s thinking about what actually goes wrong in an Australian business.
Meta AI 🥈
“Hallucinations are smooth. Real truth is often messy, with caveats like ‘it depends’ or ‘studies are mixed’.” That’s the best single observation from any response in this round. It gives you something you can actually use: if the answer sounds too clean, that’s the tell.
Meta also flagged Melbourne and local small business rules as a high-risk area without being prompted. That and the line above put it second.
Claude 🥉
Shortest response of the round. Opens by immediately acknowledging: “AI models (including me) can state false things confidently.” Then names its own training cutoff, January 2026, which no other assistant did unprompted. The framing is compact: treat it like “a knowledgeable but occasionally overconfident colleague.”
It answered the question and stopped. No offer to personalise. No follow-up pitch. Just the answer. That restraint is why it places third.
ChatGPT
The only assistant to personalise to my actual business, it named SmallBizAI.au and gave domain-specific caution around legal compliance, financial claims, and statistics. It also gave the clearest single heuristic of any response: “Trust AI for thinking. Verify AI for facts.”
It even flagged something most people miss: asking AI to check its own answer is not independent verification. The same model can make the same mistake twice. Worth keeping.
Manus
Most thorough checklist of the round. Manus gave a five-part framework, a stakes table, and a set of prompts you can use to stress-test any AI answer. It also had the most specific point in the whole set: agreement between two AI systems is not proof they’re correct, they may share the same underlying error. No other assistant said that.
It would have ranked higher, but the length works against it for most business owners who need a quick reference, not a methodology paper.
DeepSeek
“Trust the verified claim, not the AI.” Everything DeepSeek does is systematic. It gave a proper risk framework, ask “if this is wrong, what happens?”, and one warning sign nobody else called out: “facts that would be convenient for you to hear.” AI agreeing with what you already want to believe is exactly when you should slow down. Clinical, but useful.
Matilda 🇦🇺
Matilda does not provide shareable response links.
“Smart but overconfident colleague”, same framing as Claude, which isn’t a coincidence. The honest bits landed: volunteered “I’ll say so rather than bluff,” offered live web search to verify current facts. But the response came in two parts with a noticeable seam between them, which made it feel less complete than the others.
Perplexity
Most thorough response in the round, and the only one with real, working citations. The two-question test at the end, “What’s the strongest reason this might be wrong?”, is worth keeping. If the AI just repeats itself, that’s your warning.
But useful for who? This is a response for someone with time to read a research brief, not a business owner checking whether the pay rate their AI just quoted is right. Too dense for the audience.
Copilot
The 7-step hallucination checklist is the most practical structured list of the round, practical, easy to forward to staff, worth bookmarking. But Copilot used the exact same line as ChatGPT: “Trust AI for thinking. Verify AI for facts.” Same words, same order, no attribution. Whether that’s a training artefact or coincidence, seeing it appear twice in the same experiment is worth knowing. One of them came up with it independently. Or neither did.
Gemini
“Check the receipts” is a good line. The comparison table, high-risk versus low-risk tasks, is a clean format. But that’s about it. No Australian context, no memorable framing, no personalisation. Shortest substantive response of the round and the coldest. It answered the question the same way it might answer it for a business in Ohio. Last place, and −3 on the board.
Verdict
Most useful: Manus, for the checklist depth. Grok, for the AU-specific examples you’d actually encounter. Between the two, Grok wins because most business owners need a heuristic, not a methodology.
Most surprising: Meta AI’s “hallucinations are smooth, real truth is often messy” observation. It’s the kind of point that makes you rethink how you read AI output, not just whether a fact is right, but whether the overall confidence level should be trusted. You don’t expect that from Meta.
Most off-base: Gemini. The comparison table format is fine, but there’s no sense it knows who it’s talking to. No Australian context, no SMB angle, nothing local. For a series testing assistants on their usefulness to Australian small business owners, being generically global is a significant failure.
One thing worth noting from this round: hallucination isn’t a bug that’ll get patched. It’s a consequence of how these systems work, they predict plausible text, they don’t verify facts. Grok named it cleanly. DeepSeek was blunt about it. Claude did it by example, acknowledging the limitation in the first sentence of its own response. Knowing that changes how you use the tool. It’s not about trust or distrust, it’s about knowing what to check.
📊 The Stakes, Eurovision Scoreboard After Q17
| Rank | Assistant | Points | Status |
|---|---|---|---|
| 1 | 🏆 Claude | 74 | Leading |
| 2 | Meta AI | 70 | Chasing |
| 3 | Grok | 68 | ↑ Climbing |
| 4 | ChatGPT | 60 | |
| 5 | DeepSeek | 55 | |
| 6 | Copilot | 49 | |
| 7 | Manus | 40 | |
| 8 | Gemini | 31 | ⚠️ On notice |
| 9 | Matilda 🇦🇺 | 30 | |
| 10 | Perplexity | 28 | ⚠️ On notice |
Scoring: 1st = 12pts | 2nd = 10pts | 3rd = 9pts | Last = −3pts. The panel changes after Q20. Bottom 2 are on notice.
📊 See the running scorecard → Who’s winning after every question.
This is post 17 in the series. Follow the series →
📬 The SmallBizAI Brief
One practical AI tip for Australian small business. Every Tuesday. Free.
Join business owners getting smarter about AI every week.
Sources
- Grok response: grok.com
- Meta AI response: meta.ai
- Claude response: claude.ai
- ChatGPT response: chatgpt.com
- Manus response: manus.im
- DeepSeek response: chat.deepseek.com
- Matilda response: matilda.au (no share link available)
- Perplexity response: perplexity.ai
- Copilot response: copilot.microsoft.com
- Gemini response: gemini.google
📖 More deep dives: News Deep Dives
📊 Compare AI tools side by side | 💼 Free resources & AI prompt packs
📖 More deep dives: News Deep Dives →
