An AI cover letter generator that doesn’t lie is one that can only write sentences traceable back to something you actually did. Most tools can’t do that, because they generate the letter from the job description and a vague sense of what a strong candidate sounds like. SiviGen drafts the letter from a structured record of your real experience, then extracts every claim in the draft and verifies each one against that record before you see it.
Why do AI cover letters make things up?
An AI cover letter generator invents things because of what you asked it to do. The instruction is “write a persuasive letter for this job,” and the model has exactly one document that describes the job and, at best, a copy of your resume it is free to interpret loosely. Persuasiveness has a shape — a track record, a metric, an initiative led, an enthusiasm for the mission. The model reaches for that shape. It has no mechanism for checking which parts of the shape are true about you, because nothing in the request asked it to check.
So it fills the gaps. Not maliciously, and usually not with anything outrageous. It writes that you have a proven track record of delivering at scale, that you led cross-functional teams, that you consistently exceeded targets. Every one of those is a factual claim about your career. None of them came from your career. They came from the statistical middle of every cover letter ever written.
The tell is a letter you’d struggle to defend
The symptom is easy to spot once you know what to look for: read your draft and ask, for each sentence, whether you could talk about it for two minutes with specifics — a project name, a number, a person you worked with, what went wrong. A fabricated claim gives you nothing to say. It has no detail underneath it because there was never anything underneath it.
The cost is not embarrassment. Embarrassment would be survivable. The cost is that the letter is read by the person who will interview you, and it sets the agenda for that interview. If your letter promised cross-functional leadership, the first question is about a time you led across functions. You now have a choice between a long pause and a story you’re partly inventing on the spot, and the interviewer — who has done this many times — watches you make it. A letter that overpromises doesn’t get you a better interview. It gets you a worse one, later, after you’ve already spent the travel and the nerves.
Before and after: the same paragraph, fabricated and verified
Here is a paragraph of the kind a general-purpose model produces. It’s illustrative, not from a real user.
I am a results-driven senior engineer with a proven track record of delivering scalable, high-impact solutions in fast-paced environments. I have consistently exceeded targets, led cross-functional teams to successful outcomes, and driven significant improvements in both system performance and customer satisfaction. I am passionate about your mission and confident I would be a valuable addition.
Now assume the resume behind it says exactly this and nothing more:
Backend Engineer, Payments — 2021–2025. Built and operated the Stripe payments integration. Reduced checkout API p95 latency from 800ms to 300ms. Mentored two junior engineers.
Every claim in the paragraph, checked against that. The verdicts are the three used by the truthfulness audit — supported covers both a claim your record states outright and one that follows safely from what it states:
| Claim in the letter | What the resume supports | Verdict |
|---|---|---|
| Proven track record of delivering scalable solutions | Nothing about scale is stated | Unsupported |
| Consistently exceeded targets | No targets appear anywhere | Unsupported |
| Led cross-functional teams | Mentoring two juniors is not cross-functional leadership | Unsupported — a scope upgrade |
| Drove improvements in system performance | The p95 latency bullet | Supported |
| Drove improvements in customer satisfaction | Nothing measures satisfaction | Unsupported |
| Passionate about your mission | Not a factual claim, but also not evidence | Not a claim |
Of six statements, one is supported, one isn’t a claim at all, and four have nothing behind them. Note that none of them is contradicted — nothing here conflicts with the resume, it simply isn’t in it, and those are different problems. Here is the same paragraph written so that every assertion traces to a fact:
I spent four years on the payments team at Acme, where I built and operated our Stripe integration. The work I’d point to first is a latency problem: our checkout API sat at 800ms p95, and I brought it down to 300ms. I also mentored the two junior engineers on the team. Your posting leads with payment reliability and with bringing newer engineers up to speed, which is most of what I did there.
Four claims, four sources: the dates and team from the role line, the integration from the first bullet, the latency numbers verbatim from the second, the mentoring from the third. The final sentence is not a claim about the candidate at all — it’s an observation about the job posting, which is always safe to make.
Notice what the honest version gained. It’s shorter, it’s more specific, and it is harder to write, which is the whole reason models don’t produce it by default.
How to audit your own cover letter draft
The method is the same one you’d run on a CV — pull out every factual assertion, name the evidence, fix what conflicts, then add or soften the rest — and it is written out step by step, with a worked example, in the full method, with a worked example. Three things about it are specific to a letter.
Apply the two-minute rule, not the ten-second one. A letter sentence is an invitation to a follow-up question, so the bar is whether you could talk about it for two minutes with concrete detail — a project name, a person, what went wrong — not whether it sounds true. Don’t soften; delete. “Helped drive significant improvements” is the same claim as “drove significant improvements,” just wearing a hat; hedged fabrication is still fabrication, and it reads worse. Then reread what survived. If cutting the unsupported claims left a generic letter, that’s the finding: the letter was being carried by the fabrication, and the real work is now finding the specific things you did that the first draft skipped over — which is what the free achievement finder is for.
For a fast first pass on the wording itself, the free resume claim checker flags the sentences most likely to need evidence — no signup, and it runs entirely in your browser. It can’t see your resume, so naming the evidence is still your job.
How SiviGen generates a cover letter it can defend
SiviGen’s cover letter runs through the same verification machinery as the resume, not a separate, gentler version of it. That reuse is the design.
First, the letter is drafted from your fact graph — the structured record built when you upload your resume, where every role, bullet and skill becomes an atomic, individually addressable fact. The drafting step is explicitly instructed never to introduce an employer, title, metric, date or skill that isn’t in that graph. The same principle drives veridical CV tailoring on the resume side.
Then the interesting part. The letter’s prose is wrapped in a minimal resume structure and handed to the same claim extractor and claim verifier the resume pipeline uses. The system doesn’t have a special, more forgiving path for letters. It converts the letter into something the existing verifier already knows how to read, then mines every factual assertion out of it and labels each one against the fact graph: supported — whether your record says it outright or the inference follows safely from what it says — unsupported, or contradicted. If you’ve read how the truthfulness audit works on a resume, it is that, pointed at prose.
Finally, a gate. If any claim comes back contradicted, or unsupported at high severity,
the letter is redrafted once — with the specific failing claims passed back in and
marked for removal or rephrasing. Once. The result is labelled clean or revised,
and either way you see it. There is no loop that grinds until the verifier gives up, and
no silent retry that hides how the sentence got there.
The letter can use facts your resume had to cut
This is the part most people don’t expect, and it’s the strongest argument for generating the letter properly rather than asking a chatbot to summarize your resume.
A tailored resume is a document under pressure. It has a length limit, a target job, and a ranking problem: some of your genuinely strong material is less relevant to this posting than the material that displaced it, so it gets trimmed. That trimming is correct for the resume. It’s a real loss of information anywhere else.
Most cover letter generators take the finished resume as their input. That means they work from the trimmed version of you — everything that lost the ranking contest is invisible to them — which is exactly why so many generated letters read as a restatement of the resume in longer sentences. They have nothing else to say.
SiviGen’s letter draws on the full fact graph, not the tailored resume. The side project that didn’t fit, the earlier role that got compressed to one line, the tool you used at a job three positions back — all still there, all still verified against the same graph. So the letter can do the job a cover letter is actually for: saying the thing the resume couldn’t fit. And because verification runs against the full graph too, using a trimmed fact is not a loophole. It’s still a claim, still extracted, still checked.
What this doesn’t do
Verification is against your resume, not against reality. If your uploaded resume overstates something, the fact graph inherits it and the letter will treat it as supported. The system guarantees that the letter doesn’t invent anything beyond what you gave it — it can’t audit your source material. It also revises at most once, and requires a tailored resume before it will generate a letter at all, because a letter built without the job match is just prose.
That’s the honest boundary, and it’s worth comparing against what other vendors promise. The comparison takes ten minutes to run yourself: give any of them — this one included — a job description asking for something your history doesn’t contain, and read the draft it hands back. Inside the boundary, every sentence in the letter is one you can defend in the room. You can see the rest of the product, or generate a letter and read the claim list yourself at app.sivigen.com.
FAQ
Do AI cover letter generators actually fabricate, or is that overstated?
They fabricate routinely, and it’s structural rather than occasional. A model asked to write persuasively from a job description will produce the shape of a strong claim — track record, scale, leadership — whether or not the underlying fact exists, because nothing in the request requires it to check. The fabrications are usually mild and generic, which is what makes them easy to send without noticing.
How do I check an AI cover letter for made-up claims?
Take each factual assertion in the letter and ask whether you could talk about it for two minutes with concrete detail — a project, a person, what went wrong. A fabricated claim gives you nothing to say, because there was never anything underneath it. Delete those rather than hedging them; a softened invention is the same claim. The full claim-by-claim method, with a worked example, is in the truthfulness audit.
Is a hedged claim safer than a direct one?
No. “Helped contribute to significant improvements” makes the same assertion as “drove significant improvements” — an interviewer will follow up on it identically — while reading as evasive rather than modest. Hedging changes the tone, not the truth value. If you can’t support the claim, remove it; if you can, state it plainly with the specifics that make it credible.
How is SiviGen’s cover letter different from asking ChatGPT to write one?
The letter is drafted only from a structured record of your real experience, then every claim in the draft is extracted and verified against that record by the same machinery that audits the tailored resume. A contradicted or high-severity unsupported claim forces a redraft. A general-purpose chatbot has no verification step, so nothing stops a plausible-sounding invention from reaching the page.
Can the cover letter mention things that aren’t on my tailored resume?
Yes, deliberately. The letter draws on your full fact graph rather than the trimmed, job-targeted resume, so genuinely strong material that lost the ranking contest for resume space is still available to it. Those facts are verified the same way as any other. It’s the reason the letter can add something instead of restating the resume at greater length.
Should I still write the letter myself?
If you can write a good one, yes — you know things about your own work that no record of it contains. What generation is genuinely useful for is the version of the problem most people actually have: a blank page at eleven at night on the fourth application of the day. The test either way is the same. Every sentence has to trace back to something you did, and you have to be able to talk about it when asked. SiviGen is free to use today; paid tiers are planned but don’t exist yet.