How to vet AI tools: a safety checklist before your team uses one
Short answer Before staff use an AI tool, check what it can produce and what stops misuse, whether it trains on your data, whether a person approves output, and whether regulators have acted against it. Accountability stays with your business, not the tool: Canada's privacy regulator rejected X and xAI's claim that Grok's users bore primary responsibility.
Want help applying this to your business? Talk it through on a free 30-minute call.
Book a callHow to vet an AI tool: the 8 checks
Run these before anyone on your team uses a new AI tool for work. Most take ten minutes on the vendor's website and settings page.
| # | Check | Pass | Red flag |
|---|---|---|---|
| 1 | What can it make? | It says what it refuses to make (fake people, explicit images, fake reviews) | "No limits," "uncensored," or a "spicy" mode |
| 2 | Real people's photos | It blocks or limits edits of people who didn't agree | Anyone can upload and change anyone's photo |
| 3 | Regulator actions | No open investigations or findings | A privacy regulator or the FTC has acted against it |
| 4 | Your data | Business plan that doesn't train on your data by default | Free plan with customer data pasted in |
| 5 | Approval | A person approves before anything is published, sent or spent, and a log shows what it did | It posts, emails or changes ads on its own with no off switch or record |
| 6 | Security proof | A recent SOC 2 report or similar audit, available on request | "We take security seriously," nothing to show |
| 7 | Claims it makes | Says plainly what it does and doesn't do | Promises income, rankings, or "replaces your lawyer" |
| 8 | Exit | You can export your work and delete your data | No delete option, or data kept "indefinitely" |
Fail check 1, 2 or 3 and stop there. The rest you can often fix with settings, a paid plan or a rule for staff.
Which plans train on your chats is covered in AI tools for small business. Which steps need a person to approve them is in Human oversight in AI marketing.
What went wrong with Grok's image editing?
xAI shipped an image tool on X that let anyone edit a real person's photo, with weak limits, and it was misused within days. It is the clearest recent case of what checks 1 to 3 are for.
The timeline, from the Canadian privacy regulator's findings (OPC: PIPEDA Findings #2026-004):
| Date | What happened |
|---|---|
| July 2025 | xAI launches Grok Imagine, its image generation and editing model |
| December 23–24, 2025 | An "Edit Image" button appears on images in X, including other users' photos |
| December 28, 2025 | X starts its crisis protocol as misuse grows |
| By January 8, 2026 | News reports cited by the regulator put it at over 6,000 sexualized deepfakes an hour |
| January 8, 2026 | The @Grok account blocked from making sexualized deepfakes of real people; its image generation limited to paying subscribers |
| January 14, 2026 | The deepfake blocks extended to Grok on X and the standalone Grok app, for all users including paying ones |
| January 26, 2026 | The European Commission opens a formal investigation under the Digital Services Act (European Commission) |
| June 11, 2026 | Canada's privacy regulator finds X Corp. and xAI broke federal privacy law (OPC) |
The Canadian findings read like a list of skipped checks. The companies told the regulator they do not verify consent of the people shown in the images. The privacy impact assessment (a written review of privacy risks done before launch) was "only updated to include Grok Imagine in March 2026," eight months after launch. The regulator called a response of over ten days "an insufficient and untimely response," and recommended the companies suspend Grok Imagine on all their products. As of the report, the matter was "not resolved."
The European Commission is looking at the same failure from another side: whether X assessed the risk of Grok features before deploying them, including "manipulated sexually explicit images" (European Commission, January 26, 2026).
Who is responsible when an AI tool causes harm?
The business that uses it to publish, send or decide. Not the tool, and not only the person who typed the prompt.
The companies argued that "users bear primary responsibility for the content that they generate." Canada's privacy regulator disagreed: the companies built, hosted and deployed the tool, so they are accountable (OPC: PIPEDA Findings #2026-004). The same logic reaches you when you use AI in your business. Canada's privacy regulators put it in one line:
"Know that accountability for decisions rests with the organization, and not with any kind of automated system used to support the decision-making process." (OPC: Principles for responsible, trustworthy and privacy-protective generative AI technologies)
The US says the same about AI in marketing:
"Using AI tools to trick, mislead, or defraud people is illegal." (FTC, September 25, 2024)
That FTC sweep included DoNotPay, which sold itself as "the world's first robot lawyer." Check 7 is there because a vendor that overstates what its tool does may also overstate how safe it is.
Questions to ask the vendor
Ask in writing, and keep the answers. A good vendor answers each in a sentence or links a page.
- What will your tool refuse to make, and how do you test that before release?
- Can it edit photos of real people? What stops someone using a customer's or employee's photo?
- Do you train models on what we type or upload? Can we turn that off, and is it off by default on our plan?
- Where is our data stored, and for how long? Can we delete it?
- Can every output wait for a person to approve it before it's published, sent or spent?
- Can we see a recent SOC 2 report? (A SOC 2 is an independent audit of a company's controls for security, availability, processing integrity, confidentiality or privacy, per the AICPA.)
- Has any regulator investigated you or your AI model provider?
- If your tool produces something illegal or false under our name, what do your terms say about who pays?
Then search the tool's name with "investigation," "privacy commissioner" and "FTC."
Which laws apply to AI tools my business uses?
The tool vendor answers to its regulators; you answer for what you publish. These are the rules most likely to touch a small business.
| Where | Rule | What it means for you |
|---|---|---|
| US | FTC Act, applied to AI (FTC, 2024): "no AI exemption from the laws on the books" | AI-made ads, reviews and claims must be as true as any other |
| US | TAKE IT DOWN Act, signed May 19, 2025 (Public Law 119-12) | Knowingly publishing an AI "digital forgery" intimate image of a real person without consent is a federal crime; platforms must remove reported images within 48 hours |
| Canada | PIPEDA, the federal private-sector privacy law (OPC, 2026) | Using someone's personal information, including their photo, generally needs valid consent |
| EU | Digital Services Act (European Commission, 2026) | Puts risk duties on online platforms, strictest on the largest like X; it binds the platform, not you as its user |
This is what the rules say, not legal advice. For patient data, see Is ChatGPT HIPAA compliant?
Example
A salon picks an AI photo editor
The owner wants to touch up client before-and-after photos. Tool A has a "no restrictions" mode and no word on real people's photos. Tool B, $20 a month, blocks edits that change clothing or body shape and says it doesn't train on uploads. She picks Tool B and still only uploads photos of clients who signed a photo release.
A contractor's office manager tries a free chatbot
She pastes 40 customer names and addresses into a free plan to draft follow-up emails. Check 4 fails: on many free plans chats can be used for training unless turned off. The owner moves the team to a business plan and writes one rule: no customer details in free tools.
An accounting firm vets an AI agent before tax season
The vendor answers all eight questions in writing, shares a SOC 2 report, and every email waits for approval. One answer is weak: data is kept 90 days after cancelling. The firm signs, notes the 90 days in its privacy policy, and sets a reminder to request deletion if it leaves.
FAQ
Is Grok safe to use for my business?
Canada's privacy regulator found in June 2026 that Grok's image tools broke federal privacy law. For other uses, run the eight checks on your plan.
Do I need a SOC 2 report from every AI vendor?
No. Ask when the tool holds customer, patient or financial data. For a tool that only drafts posts, checks 1, 4 and 5 matter more.
Can my staff use AI to edit photos of real people?
Only people who agreed, for the use they agreed to. In the US, knowingly publishing an AI-made intimate image without consent is a federal crime.
Is a free AI tool safe for work?
Fine for drafts with no customer or staff details. For personal information, use a business plan that doesn't train on your data by default.
Different for you?
- Choosing which AI tools to use at all: AI tools for small business.
- Deciding which AI steps need a person to approve: Human oversight in AI marketing.
- Want to know what AI can't do for your marketing: AI marketing limits.
- A health practice handling patient messages: Is ChatGPT HIPAA compliant?
- Vetting an AI SEO agent in a trial: AI SEO agent review: 8 checks.
- Want content and listings handled with approval built in: SEO & GEO agent.