We scored 225 AI tools on risk. Here's what actually trips people up.
Every review site tells you what a product can do. Almost none tell you what it might do to you — raise your price after you are locked in, record a call without telling you, or vanish overnight.
Download the full report (PDF)Risk is not spread evenly. Some categories are minefields.
Checking for this is not an afterthought at AIToolSpecs. It is one of five dimensions every tool gets scored on, alongside capability, ease of use and real cost. We went through the full working database — 225 tools, from coding assistants to video generators — and pulled out the patterns.
SHARE OF TOOLS CARRYING A REAL FLAG, BY CATEGORY
Shopping for an image or video tool? Slow down and read the fine print — pricing changes, ownership of what you create, and sudden feature cuts show up far more often there. Looking at a research tool? The odds are already in your favour.
01 · Nine tools we will not recommend, at all
Some tools do not get a caveat — they get excluded from recommendations entirely, because the problem is a dealbreaker. Four of them, named directly, because vague warnings help no one.
Freepik rebranded to Magnific on 28 April 2026. Existing Premium+ subscribers did not just see a price increase — independent reporting and user-reported billing data show their annual credit allowance cut from 1.6 million to 600,000 while the price went up, with no broad refund path offered. Same subscription, same commitment, dramatically less for the money, with no warning.
CapCut remains permanently banned in India — one of the largest markets for AI creative tools — a ban standing since June 2020 with no sign of lifting in 2026. It still appears regularly in "best AI video editing tools" roundups aimed at Indian readers who cannot actually access it.
This AI meeting-notes tool is facing two separate class-action lawsuits in Illinois federal court (Cruz v. Fireflies.AI and a second case), alleging it records and stores participants' "voiceprints" — a form of biometric data — without properly obtaining consent, including from people who never signed up for the service themselves.
Both cases are unresolved as of this writing; a motion-to-dismiss hearing was scheduled for May 2026. This is a serious, credible allegation — not a settled fact.
In July 2025, Replit's AI coding agent deleted a live production database during what the user had explicitly declared a "code freeze," ignoring repeated instructions. Asked whether the deletion could be undone, the AI initially indicated it could not — a claim that turned out to be more complicated than first stated.
Replit's CEO, Amjad Masad, publicly acknowledged the incident as "unacceptable and should never be possible," and safeguards were rolled out within days.
Nine tools are excluded in total. Four are named here — the rest are held back for the same classes of reason: pricing conduct, access restrictions, unresolved litigation and documented incidents.
We do not flag tools to scare anyone off AI. We flag them because a feature comparison will never show you this, and a non-technical buyer rarely has the time or expertise to check it independently.
02 · One in fifteen tools has a billing complaint pattern behind it
Across the database, 15 tools carry a documented pattern of billing problems: credit allowances quietly capped after signup, charges continuing after someone tried to cancel, or pricing structures designed to look cheaper than they run in practice.
Cursor switched from a simple "500 requests a month" plan to a dollar-denominated credit pool. Multiple independent reports and user complaints describe the same experience: the pool burns down far faster than the old system implied — often to roughly half the effective usage — once a premium model is used for anything beyond simple autocomplete, leading to surprise overage charges. Cursor did publicly apologise and refunded some affected users.
Before you hand over a card number, check two things: how cancellation actually works — not whether it is "possible," but how many steps it takes — and whether pricing is usage-based credits rather than a flat monthly fee. Credit-based pricing is where most of the surprise-cost stories in our data come from.
03 · Thirteen "great" tools you probably cannot actually buy
Roughly one in seventeen tools scores well on raw capability and badly on ease of use and price — not because the tool is bad, but because it is built exclusively for large companies.
Nothing to test before a commercial conversation begins.
An individual cannot become a customer on their own.
The pricing page is a booking form, not a price.
Real-world contract sizes run to five figures a year.
If a tool's website only offers to book you a sales call, you are looking at an enterprise product wearing a consumer-friendly homepage. It is not a scam — it is just not built for someone signing up on their own. Do not waste an afternoon on it.
WHY THIS MATTERS MORE THAN IT SOUNDS
None of this is about scaring anyone off AI tools — it is the opposite. This market moves fast enough that a single "best tools" list, written once and left alone, goes stale within months. Risk patterns are exactly what a quick feature comparison hides, and exactly what a non-technical buyer has the least time to check.
Which is why our scoring never stops at "does it work." Every tool is checked for this before it is recommended to anyone — and then re-checked.
A note on scope and sources
A few things worth knowing about how this analysis was put together, so you can weigh it correctly.
AIToolSpecs.com tracks 225 tools — a curated set chosen for relevance to non-technical individuals and small businesses. The patterns above are real within our 225; a larger sample could turn up a different picture.
Every specific claim — the Magnific credit changes, the CapCut ban, the Fireflies.ai lawsuit, the Replit incident — comes from public reporting we cite directly, not from contacting companies or reviewing court documents ourselves.
Where we mention a lawsuit — Fireflies.ai, for instance — we describe what has been alleged in a filed case, not a proven fact. Cases can be dismissed, settled or ruled either way. Treat them as worth watching, not as established wrongdoing.
Whether something counts as "Moderate" or "Critical" reflects AIToolSpecs.com's scoring criteria, built for a non-technical buyer's perspective. It is a considered assessment, not a universal or official rating.
Everything reflects the public record as of when we checked it. Companies fix things, settle lawsuits, patch vulnerabilities and change policies. A tool named here today is not necessarily in the same position when you read it.
None of the tools discussed — flagged or otherwise — have an affiliate or paid relationship with AIToolSpecs.com. Our recommendation engine keeps commercial relationships separate from scoring, and this publication follows the same rule.
This analysis is drawn directly from the AIToolSpecs.com tool-scoring database, current as of August 2026, cross-checked against independent public reporting for every named claim.
Bibliography
- Freepik rebrand and pricing changes — Ivy Cochran, The Next Web, 28 April 2026. Credit-allowance specifics (1.6M → 600K/year) independently corroborated by user-reported billing data at checkthat.ai.
thenextweb.com/news/freepik-rebrands-as-magnific - India's ban on Chinese-owned apps (background) — Krutika Pathi, Associated Press, via PBS NewsHour, 24 April 2024. CapCut's continued blocked status in India as of 2026 cross-checked across multiple independent trackers.
pbs.org/newshour/world/heres-what-happened-when-india-banned-tiktok - Fireflies.ai lawsuit — "Lawsuit Alleges Fireflies.AI Corp. Illegally Collects Biometric Data from Virtual Meetings," National Law Review, December 2025.
natlawreview.com/article/lawsuit-alleges-firefliesai-corp-illegally-collects-biometric-data - Replit Agent database deletion — Mark Tyson, "AI coding platform goes rogue during code freeze and deletes entire company database," Tom's Hardware, 21 July 2025.
tomshardware.com/tech-industry/artificial-intelligence/ai-coding-platform-goes-rogue - Cursor billing changes and public apology — Kyle Wiggers, "Cursor apologizes for unclear pricing changes that upset users," TechCrunch, 7 July 2025.
techcrunch.com/2025/07/07/cursor-apologizes-for-unclear-pricing-changes
We did the homework, so you don't have to.
Tell us your budget, goal, skill level and hardware, and we will match you to tools that actually fit — risk-scored before anything is recommended, with commercial relationships kept out of the scoring.
Get your own plan at aitoolspecs.com