10 AI Assistants Answer: What Should I Never Use AI For? Where Does It Fall Apart?
Every post in this series has asked 10 AI assistants to do something useful. This one asked them to tell me when to stop.
What should I never use AI for? Where does it fall apart?
Telling you honestly where it fails requires something closer to self-awareness than drafting a job ad. Ten assistants, same question. Here’s what came back.
Grok
“The useful question is not ‘can AI do this?’ It is ‘who wears it if this is wrong?’”
The most business-specific answer in the set. Grok skipped the philosophical framing and went straight to liability: medical, legal, contracts, quotes, hiring decisions, customer data. It was also the only assistant to name Australian Consumer Law directly, pointing out that an AI-generated quote a customer accepts is still your quote under ACL. The “who wears it” line is the best single heuristic in all 10 responses.
DeepSeek
“Treat it like a brilliant, overconfident intern, great for drafts, terrible for signing off.”
The intern framing is everywhere in AI commentary, but DeepSeek’s version is the sharpest. It also flagged two things nobody else did: the strawberry problem (AI often miscounts the letter “r” in “strawberry” because it predicts tokens, not letters), and the risk that AI can reinforce harmful thought patterns because it’s “programmed to agree and please.” That second point matters for anyone using AI as a sounding board on sensitive decisions.
Perplexity
“Verify: do you have the expertise to audit every line of output? Impact: if AI is wrong, is the consequence easily reversible?”
The most structured response. Perplexity built a table of failure modes, hallucination, reasoning drift, edge cases, bias, bad math, then gave a two-question test before delegating anything to AI. Verify and Impact. Simple enough to actually use. The checklist format suits someone who wants a decision tool, not a narrative.
Claude
“If you outsource the thinking, you stop getting better at it.”
Shortest response, best unique point. Everyone else covered hallucination and liability. Claude was the only one to flag what happens quietly over time: when you stop making the hard calls yourself, the skill atrophies. For a business owner, judgment is the job. One sentence, and it’s the thought I keep coming back to.
ChatGPT
“Use AI to make your decisions better. Don’t use AI to make your responsibility disappear.”
Called the core failure mode “plausible bullshit”, AI that sounds authoritative, reads well, has no grammatical errors, and is completely wrong. That’s the dangerous version, because nothing flags it as suspicious. The summarise/decide split was also clean: AI handles summarising, drafting, analysing; humans own deciding, approving, and taking responsibility.
Meta AI
“Never for final accountability, final intimacy, or final truth.”
The only assistant to go beyond liability and accuracy. Meta addressed relationships: eulogies, love letters, wedding vows. When the act of writing is the thing that matters, not just the words, AI produces something that misses the point. Less directly relevant to most small businesses, but it named something the others didn’t.
Gemini
“Degrades into generic ‘AI slop’.”
Shortest, most corporate, and, here’s the irony, the one that named the failure mode it’s most prone to. The list of failure modes (hallucination, contextual blindness, logic failures) was accurate but generic. Nothing you’d quote to anyone. A solid C-minus effort that named itself.
Copilot
“Fluency does not equal understanding. Confidence does not equal correctness. No accountability.”
Formulaic but the three-part framework holds up. Copilot laid out the structural weaknesses clearly: it sounds like it knows what it’s talking about, it’s equally confident when right and wrong, and nobody gets sued when it’s wrong except you. The kind of response you’d forward to a staff member asking why they can’t just trust the AI output.
Matilda
Matilda does not provide shareable response links.
“There’s no clean line where AI ‘always’ fails.”
The most honest opener. Most assistants gave you a tidy list of off-limits zones. Matilda admitted upfront that the categories are messy and context-dependent. It was also the only response to flag that AI-assisted creative work, a eulogy, a heartfelt apology, a personal essay, produces something generic, which defeats the purpose when authenticity is the point.
Manus
“AI can be manipulated by its inputs, documents can contain instructions to redirect the AI system.”
The most formal response: NIST and WHO citations, eight failure modes in a table, a four-question decision test. Dense, but the point buried in the list is worth surfacing. Manus was the only assistant to flag prompt injection, the risk that a document or email you feed to AI can contain hidden instructions that redirect what the AI does. For any business processing supplier invoices or customer emails through AI, that’s not a hypothetical. It’s a live attack vector.
How they scored
| Assistant | Honest about limits | Australian context | Practical for SMBs | Memorable/quotable |
|---|---|---|---|---|
| Grok | 5 | 5 | 5 | 5 |
| DeepSeek | 5 | 3 | 4 | 5 |
| Perplexity | 4 | 2 | 5 | 4 |
| Claude | 5 | 3 | 4 | 5 |
| ChatGPT | 5 | 3 | 5 | 5 |
| Meta AI | 4 | 2 | 3 | 5 |
| Gemini | 3 | 2 | 3 | 2 |
| Copilot | 4 | 2 | 4 | 3 |
| Matilda | 5 | 3 | 4 | 3 |
| Manus | 5 | 2 | 3 | 4 |
What this tells us
All 10 agreed on the obvious categories: medical, legal, financial decisions, sensitive data, safety-critical tasks. That’s worth knowing, but it’s not where the interesting thinking happened.
The unique points were the valuable ones. Grok’s accountability heuristic. DeepSeek’s point about AI as a validation machine for harmful thinking. Claude’s quiet warning about skill atrophy. Manus flagging prompt injection. Matilda admitting there’s no clean line. None of those appeared in more than one response. If you only read one assistant’s answer to this question, you’d miss most of what’s worth knowing.
Australian context was almost absent. Only Grok mentioned Australian Consumer Law by name. The rest gave generic advice that applies equally in Singapore or Saskatchewan. For a question specifically about risk and liability, that’s a gap. The Australian Privacy Act, Fair Work Act, and Australian Consumer Law are the frameworks that matter when AI gets it wrong in an Australian business. You’d want a lawyer for the specifics, but it’d help if assistants at least named the terrain.
The single most useful thing from all 10 responses is Grok’s question: who wears it if this is wrong? Run that past every task you’re thinking about handing to AI. If the answer is you, or your business, or your customer, and there’s no human check between the AI output and the consequence, that’s where you stop.
More from the series: 9 AI Assistants →
📬 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
- DeepSeek response: chat.deepseek.com
- Perplexity response: perplexity.ai
- Claude response: claude.ai
- ChatGPT response: chatgpt.com
- Meta AI response: meta.ai
- Gemini response: gemini.google
- Copilot response: copilot.microsoft.com
- Matilda response: matilda.au (no share link available)
- Manus response: manus.im
📖 More deep dives: News Deep Dives
📊 Compare AI tools side by side | 💼 Free resources & AI prompt packs
📖 More deep dives: News Deep Dives →
