AI Writing Tools User Reports: 13 Logged, Dated Sources

How this record was built

Every page of tool claims on this site is built under the Mimo Evidence Protocol (MEP v1.0): a documented query census, screening with recorded exclusion reasons, live re-fetching of every included source, and dual independent coding of every entry. The sweep ran July 10, 2026, off 21 documented queries plus 8 direct captures.

A source got in only if it was dated, publicly linkable, and one of three things: a first-person account, a test with a disclosed method, or a vendor primary page (always labeled self-interested). I ran this sweep myself, and the part worth telling is what fell out of it: two hallucination-rate figures that circulate widely, attributed to a benchmark paper, dissolved on contact — the live abstract contained no such numbers. They were a fetch away from being quoted here as fact. That near-miss is the whole argument for re-fetching before printing, and those figures appear nowhere on this page.

The 15 entries

Each entry carries an EV id — the same codes cited across every Mimo review — a source class, and a source date. The full JSON, with quotes and URLs, lives in the public evidence log.

What users report, tool by tool

Koala AI is the most-documented tool here: 4 entries, from two independent affiliate-disclosed testers plus vendor pricing. A long-term reviewer got a one-shot draft that self-structured and built a table (EV-04); a ten-plus article test logs drafts shorter than the length set (EV-05) and too thin on specialist subjects — “I can’t deliver this to my clients” (EV-06). The full Koala AI review turns these into a verdict.

SEOWriting.ai enters on two entries, neither an independent test — that absence is a logged gap. The vendor markets 1-click publish, 48 languages, and “1000+ articles in hours” (EV-13); Capterra showed 4.8/5 across 78 reviews at capture (EV-14). Full write-up: the SEOWriting.ai review. Jasper carries two: the one disclosed-method three-way test rated its long-form the strongest (EV-07); the vendor prices Pro at $69/seat with no money-back window stated (EV-12). Weighing a switch? Our Jasper alternatives guide.

Writesonic appears in that same test — best short-form copy of the three (EV-08), which says nothing about long-form — plus vendor pricing (EV-11). Copy.ai has zero tool-specific entries: the honest label is not yet evidenced, and the next sweep owes it a pass. Rytr’s single entry is its pricing — a free 10K-chars/mo tier and an “Unlimited” plan at $7.50/mo annual (EV-10); no independent Rytr quality evidence verified, so quality claims stay off this page.

Claude and ChatGPT: the no-affiliate baselines

Claude and ChatGPT sit in this record for a structural reason: a reader who subscribes to a reviewed drafting tool can earn this site a commission, while these two earn it nothing — so logging them under identical rules is the cheapest independence check we can offer. ChatGPT’s one entry is a peer-reviewed 2024 test: of the references it produced for medical case reports, 65% did not exist (EV-03) — flagged as a GPT-3.5/4-era, medical-domain result. Claude’s one entry is pricing: Pro at a vendor-stated $17/mo annual (EV-15). No Claude-specific quality evidence passed verification this sweep; the record prints that gap instead of borrowing an impression.

Documented failure modes

The strongest source here is a 2025 peer-reviewed Scientific Reports study that did at scale what this page does in miniature: it analyzed ~3 million user reviews of 90 LLM-based apps, manually annotated 1,000, and confirmed 350 as hallucination reports. Within those 350, factual incorrectness was the most frequently reported type, at 38% — a defined sample with its denominator printed, not a rate you may apply to any one tool (EV-02). A 2025 field survey pins the definition: fluent output that is factually inaccurate or unsupported (EV-01). The practical reading holds regardless of tool: budget a fact-check pass for every AI draft that ships under your name.

On credit caps, no complaint passed verification this sweep, so this works from captured quotas and arithmetic alone. Taking a 1,500-word article as the unit: Koala’s 15,000-word Essentials quota divides into ten drafts (~$0.90 each, EV-09); Writesonic’s 15 articles at $79 annual is ~$5.27 each (EV-11); Rytr’s free 10K chars covers ~one draft (EV-10); Jasper prices seats, so per-article math isn’t computable (EV-12). If your calendar says thirty posts a month, run this division before you subscribe.

Does Google penalize AI content? The dated reports

This record’s honest answer today: it holds zero verified entries on the question, and refuses to fake one. The evidence class that would answer it — dated before-and-after accounts from site operators, linkable, near a confirmed update — circulates constantly in secondhand summaries, but the platforms hosting the originals blocked re-fetching this sweep, and nothing enters this registry from a snippet. The largest exclusion bucket — 39 of 91 candidates — was affiliate material asserting conclusions with no method shown, including at least one roundup asserting a Google penalty without linking anything. The question sits first on the next re-sweep; verified operator reports will be logged as dated individual accounts with links — never converted into a rate.

The vendor-claim audit: "unlimited" vs fair use

Vendor pages are admissible here, but only as quotes with capture dates — and the audit value is often in what a page does not say. As of the July 10, 2026 captures:

A plan called “Unlimited” coexisting with a metered free tier tells you the word is a brand name, not a quota. A refund conditional on staying under the very usage the plan sells is a real constraint a sticker price hides. And “not stated” is not neutral: rollover terms and money-back windows are the clauses people hit at cancellation time.

What this record can't show

Complaint bias: people who hit problems write more, so negative accounts are over-represented. Platform skew, unusually sharp this run: review platforms and forums largely blocked verification, so the sweep tilts toward vendor pages, peer-reviewed studies, and independent blogs — all three named testers disclose affiliate links, recorded on their entries. A sample is not the population: 15 entries describe those sources, not the average subscriber. Staleness: every price was true of its page on its capture date; the ChatGPT accuracy entry is explicitly a 2024, GPT-3.5/4-era result. Perfect coder agreement on 15 items is expected at this sample size and should not impress you. Anecdotes here stay anecdotes — nothing has been turned into a percentage of anyone.

FAQ