TrueStandard
AI fact-checker for writers
AI writes confident.
Not correct.
Check every claim before your name is on it. Paste a draft and four models, from OpenAI, Anthropic, Google and xAI, flag the claims they disagree on and the citations that lead nowhere, in 60 seconds.
For newsletter writers, journalists and bloggers who publish AI-assisted drafts under their own name.
Free check runs one model on one claim, no signup. Four-model runs need an account.
You read it twice. 24 factual claims in the sample draft below, 3 of them wrong — and nothing on the page tells you which.
Every claim carries a label. 3 challenged, 2 partial, the rest verified, with the fixes ordered by what gets screenshotted first.
01 AI cites studies that don't exist. We catch them before you publish.
Ask a chatbot for a source and it will write one, with plausible authors, a journal that sounds real and a DOI in the right format. TrueStandard looks each DOI up in Crossref, the registry that issues them, and flags the ones that point nowhere.
Of 178 references ChatGPT wrote into research proposals, 28 had no DOI on record and no Google result. — Athaluri et al., Cureus, April 2023
Ellery, J., & Okafor, N. (2024). Sixteen-hour fasting windows trigger autophagy in healthy adults. Journal of Metabolic Longevity, 12(3), 201–214. https://doi.org/10.4821/jml.2024.0318
That DOI isn't registered with Crossref, so the paper it points to doesn't exist.
02 A model can't grade its own answer.
Ask ChatGPT to double-check its own draft and it tends to side with itself. In our own 80-prompt test, GPT picked its own answer about 70% of the time as judge, against a roughly 33% impartial baseline. So what if all four models agree and are wrong? Models from different companies rarely invent the same source. From our August 2026 test on the AI fact checker page: "Those 1,200 answers held 946 different fake sources. Only 2 came from more than one model." That figure covers made-up sources only. On other claims, agreement among all four still leaves room for a shared mistake.
Four companies, not four prompts.
Disagreement is the finding.
Ship a receipt, not a promise.
03 The check, end to end.
Paste a draft and four models from four labs read it separately. Nothing ships unlabeled — every claim comes back Verified, Partially Verified or Challenged, with the fixes ordered by what gets screenshotted first.
Gate A · four labs, no self-grading Gate B · no claim ships unlabeled
New research confirms intermittent fasting reduces inflammation by 70%.
It goes out to 50,000 subscribers. One reader checks the number. The screenshot outlives the issue.
Trials of intermittent fasting found a modest drop in some inflammation markers, such as CRP.
17 minutes of fixes. No correction email — and a report that shows exactly what you checked.
'Extend Lifespan by Up to 20 Years'
No human study supports this claim. The 10-15% lifespan extension it echoes comes from rodent work. Extrapolating that to '20 years' in humans is unsupported.
'A Harvard Study'
Multiple Harvard-affiliated studies on fasting exist with varying conclusions. No specific paper, authors, or year cited.
17 minutes of fixes prevents a correction email to 50,000 subscribers
Sample draft · 1,200 words · checked in 52 seconds
04 Proof you fact-checked, not just a promise.
Every check ends in a report that lists each claim, its label, how each of the four models voted and why. Send your editor the link and they see what you checked.
05 They Published Without Checking
These aren't hypothetical risks. These are real outcomes from the last 18 months.
$67.4B in global business losses from AI hallucinations in 2024. Employees spend 4.3 hours/week verifying AI output. — McKinsey, Suprmind 2026
More than one in three marketers (36.5%) say hallucinated AI content has already gone out publicly. — NP Digital, AI Hallucinations and Accuracy Report, Feb 2026, n=565
Published 77 AI-written articles. 41 needed corrections.
53% error rate. AI publishing program paused.
AI bot auto-published summaries with errors for months.
36+ corrections in 3 months. Staffers couldn't stop it.
Published article with AI-hallucinated quotes.
Reporter fired, March 2026
Submitted AI-fabricated case citations to federal court.
$5,000 fine. Career-defining embarrassment.
One verification step would have caught every one of these.
Written for solo newsletter publishers, independent journalists, and content teams.
06 What it costs
One prevented correction pays for months of service.
Checking by hand takes you 9 hours a month. TrueStandard checks the same 12 drafts in about 12 minutes. You still make the fixes yourself. Pro is $20 a month for 250 drafts, which comes to 8¢ a draft.
These are example inputs, so put in your own.
For writers who want to check drafts before publishing
- ✓ Council check. 4 AI models, every claim
- ✓ Verified / Partially Verified / Challenged labels
- ✓ Prioritized fix list per draft
- ✓ Credit estimate before every run
For writers and teams who publish frequently
- ✓ Everything in Pro
- ✓ Deep check. The 4 models argue it out, then a mediator rules
- ✓ Verification reports you can share
- ✓ Export for editors and clients
- ✓ Email support
Best for daily publishers and content teams
For agencies running verification at volume
- ✓ Everything in Max
- ✓ Deep check on every draft, not just the risky ones
- ✓ Shared credit pool across the team
- ✓ Advanced model routing
- ✓ Priority model access
- ✓ Priority email support
Is AI Getting More Accurate?
No. OpenAI's o3 hallucinates 33% of the time. o4-mini hits 48%. Newer 'reasoning' models hallucinate more, not less. Two independent research teams have shown that hallucination elimination is mathematically impossible with current LLM architecture.
This isn't a bug being fixed. Multi-model verification is structurally necessary.
Every plan includes claim-by-claim verification, prioritized fix lists, and full transparency. We never train on your data, on any plan. Cancel anytime.
07 Who it's for
- Newsletter writers, journalists and content teams who publish AI-assisted drafts under their own name.
- Drafts full of numbers, studies and quotes, the claims readers screenshot.
- Writers whose editor wants to see what was checked.
- Rewriting: TrueStandard flags a claim and says why, and the rewrite stays with you.
- Opinion pieces with no factual claims, because there is nothing for it to check.
- Certainty: four models can agree and still share a mistake on claims that aren't sources, so the report shows each vote for you to judge.
Common Questions
Paste your draft or a specific claim. TrueStandard sends it to four independent AI models. Each one fact-checks your claims separately. You get a report showing which claims are Verified, Partially Verified, or Challenged, with a prioritized list of what to fix before publishing.
We didn't build it for researchers with a reference manager. We built it for the writer the night before a send, with 50,000 subscribers waiting and one number they aren't sure of. That's why a check takes about a minute and the fixes come ordered by what gets screenshotted first.
Asking ChatGPT to check work it helped you write is circular. It tends to agree with its own output. TrueStandard uses four independent models that cross-check each other, so disagreements surface errors a single model would miss.
Overclaimed statistics ('reduces by 70%' when the research found a modest drop in one marker), unsourced claims ('a Harvard study' with no specific paper), outdated facts, and confidently wrong assertions. The claims most likely to be screenshotted if wrong.
About 60 seconds for a typical article. Googling each claim yourself and cross-referencing the sources takes the better part of an hour.
Yes, it can happen. Models from different companies rarely make up the same thing. In our August 2026 test, four models gave us 1,200 answers. Those 1,200 answers held 946 different fake sources. Only 2 came from more than one model. That result covers invented sources only. Sources never rest on a vote anyway: we look each DOI up in Crossref. For every other claim, the report shows each model's reasoning, so you can check the ones that matter most yourself. How we ran the test
No. TrueStandard flags risky claims and explains why they're problematic. You see exactly what's wrong and what the evidence actually says, but the rewriting stays in your hands.
Never. We use enterprise API agreements with all model providers. Your drafts are never used to train models.
Yes. Every check produces a verification report with claim-by-claim results, confidence levels, and timestamps. Share it with editors, clients, or attach it to internal docs.
08 Check Your Draft Before Your Readers Do
60 seconds. Four models. Every claim checked. Paste your draft and see what TrueStandard catches.
An AI fact checker for everything you publish with AI
TrueStandard is an AI fact checker that verifies the claims in your AI-written drafts across four models at once. Instead of trusting a single model to check its own work, it flags every claim the models disagree on, from fabricated citations to invented statistics and wrong dates, so you can verify AI content before it reaches your readers, not after.
See how the AI fact checker works →