Using ChatGPT, Claude, or Gemini to draft a federal proposal is legal, common, and — unless the solicitation says otherwise — undetectable as AI. What gets you eliminated is what unsupervised AI produces: fabricated FAR citations, boilerplate that never touches the evaluation criteria, and past performance claims nobody can verify. The tool is not the risk. The unedited output is.
I spent eighteen years in federal acquisition as a Contracting Specialist and Contracting Officer at GSA, IRS, DoD, and DOI, and I evaluated more technical proposals than I can count. Since AI drafting went mainstream, contractors keep asking me the same question: can the government tell? Wrong question. Evaluators do not run detection software — they read for substance. The right question is what AI-drafted text does that a tired human evaluator notices in ninety seconds. Here is the answer, and the workflow that keeps AI on your side of the desk.
Can a Contracting Officer actually detect AI-written proposals?
Not directly. No FAR provision bans AI-assisted writing, and evaluators do not run AI detectors, which are unreliable anyway. What a CO detects are the failure modes of unedited AI output — and those failures are disqualifying on their own merits, no AI accusation required.
From the CO seat, the tells are not stylistic. They are substantive:
- Citations that do not exist. A FAR clause number that resolves to nothing, a GSAR reference with the wrong subject, a made-up agency policy memo. One fabricated citation destroys your credibility for the remaining 40 pages.
- Requirements restated instead of answered. AI is fluent at paraphrasing the PWS back at the government. Evaluators score against Section M criteria, and a paraphrase earns exactly zero.
- Specifics that are conspicuously absent. No named tools, no staffing numbers, no timeline tied to the agency's actual milestones. Human proposal writers over-anchor on specifics; AI floats above them.
- Internal contradictions. Section 2 promises a five-person team; Section 4 budgets three. AI drafts sections in isolation and never reconciles them unless you force it to.
What has the government actually said about AI-generated filings?
GSA officials have publicly warned about AI-generated protest and claim filings containing fabricated case citations — the same hallucination problem that has gotten attorneys sanctioned in federal court. The message to contractors is direct: submit fabricated authority in anything the government adjudicates, and you own the fabrication.
Apply that warning to proposals. When I reviewed contractor submissions as a Contracting Officer, a wrong clause citation read as carelessness. Today, a fabricated one reads as either dishonesty or an unreviewed AI draft — and the evaluator does not have to decide which, because both destroy your rating. Worse, statements in your proposal become representations. A past performance narrative describing work your firm never did is not a hallucination problem when it is submitted under your signature; depending on the facts, it is False Claims Act exposure. FAR 52.203-13 obligations do not carve out an exception for "the model wrote it."
Where does AI genuinely help a proposal team?
In the structured, verifiable work around the writing: compliance matrices, annotated outlines, requirement decomposition, red-team review, and consistency checking. These tasks are where AI is faster than any human and where every output can be checked against the solicitation before it matters.
| Task | AI value | Why it is safe |
|---|---|---|
| Compliance matrix from Sections L and M | High — extracts every "shall" and instruction in minutes | Fully verifiable against the RFP text you feed it |
| Annotated outline mapped to evaluation criteria | High — forces criteria-first structure | Structure is reviewed before any prose exists |
| Red-team review of your draft | High — simulates a skeptical evaluator scoring against Section M | Critique output; nothing enters the document |
| Consistency and cross-reference check | High — catches staffing and timeline contradictions | Flags issues for human resolution |
| First-pass boilerplate (quality management, transition-in) | Moderate — drafts to be rewritten with your specifics | Safe only if a human injects real process detail |
| Technical approach specifics | Low — it does not know your solution | Not safe unedited; it will invent plausible methodology |
| Past performance narratives | Lowest — it does not know what you actually did | Highest fabrication risk in the entire proposal |
Where does AI kill your proposal?
In the two volumes that win or lose the evaluation: technical approach and past performance. Both are scored on specific, verifiable claims about your solution and your history — the two things a language model cannot know and will confidently invent.
As a Contracting Specialist, I saw the same pattern on nearly every weak proposal, years before AI: generic technical approaches that could have been submitted by any offeror on the shortlist. AI industrializes that weakness. Ask a model for a "technical approach for agency IT modernization" and you get competent-sounding text with no architecture decisions, no named personnel, no risk register tied to this agency's environment. Evaluators call that "parroting the PWS," and it scores Unacceptable or Marginal every time.
Past performance is worse. The model does not know your CPARS history, your contract numbers, or what your team delivered — so it fills the gaps with plausible fiction. A CO can verify past performance claims against SAM.gov, FPDS, and CPARS in minutes. Fabricated relevance in a past performance volume is not a quality problem. It is a responsibility and integrity problem.
What is a safe AI proposal workflow?
Use AI on the inputs and the reviews, and keep humans on every claim of fact. This is the workflow we run, and it maps to how evaluation actually works.
- Feed the model the actual solicitation. Upload the RFP, attachments, and amendments. Never let it answer from general knowledge — general knowledge is where hallucinated clauses come from.
- Generate the compliance matrix first. Every Section L instruction and Section M criterion, in a table, with an owner. Verify it line-by-line against the RFP before anything else happens.
- Build an annotated outline mapped to Section M. Each heading gets the criterion it answers and the proof points it must contain. Humans supply the proof points.
- Draft with human facts in, AI prose out. Writers give the model the real solution: named staff, actual tools, real metrics, true contract history. The model organizes and tightens. It never originates a fact.
- Verify every citation by hand. Every FAR, GSAR, and policy reference gets checked against acquisition.gov or eCFR. No exceptions, no batch trust.
- Run an AI red team against Section M. Have the model score your draft as a skeptical evaluator: where is this generic, unsupported, or contradictory? This is AI's highest-value pass.
- Human sign-off on every factual claim. Someone with personal knowledge certifies the past performance and technical claims before submission. Your signature on the offer makes every statement yours.
- Check the solicitation for AI disclosure language. Some agencies have begun asking about AI use in submissions. If Section L asks, answer honestly — a truthful disclosure costs nothing; a false one is a misrepresentation.
Should you disclose that AI helped write your proposal?
Only if the solicitation asks — and then, honestly. Absent a disclosure requirement, AI assistance is no more reportable than using a proposal consultant or a grammar checker. What is always your responsibility, disclosed or not, is the accuracy of every statement in the submission.
In eighteen years of federal acquisition — as both a Contracting Specialist and a Contracting Officer — I never once cared what word processor, consultant, or template produced the text in front of me. I cared whether the offeror understood the requirement and could prove it. That standard has not moved. Across our 70+ proven GSA contract awards, we use AI daily — on compliance matrices, outlines, and red teams — and a human who knows the facts still writes every claim that gets scored. If you want that discipline applied to your next Schedule offer or proposal, start at our GSA Schedule services page.
What Is the Bottom Line?
- AI use is not detectable or prohibited; unedited AI output is disqualifying. Evaluators score substance, and unsupervised AI fails on substance.
- Fabricated citations are the fastest way to lose. Verify every FAR and GSAR reference against acquisition.gov or eCFR by hand.
- Keep AI on structure and review — compliance matrices, outlines, red teams, consistency checks — and off unsupported facts.
- Never let AI originate technical approach or past performance content. Those volumes are scored on verifiable specifics only your team knows.
- Your signature owns every word. Hallucinations submitted under your offer are your representations, with CPARS and False Claims Act consequences.
Frequently Asked Questions
Is it against the FAR to use AI to write a proposal?
No. No FAR provision prohibits AI-assisted proposal writing. Individual solicitations may include AI disclosure or restriction language in Section L, so read each RFP — but absent such language, AI assistance is treated like any other writing support.
Do agencies run AI-detection software on proposals?
No standard evaluation process includes AI detection, and detection tools are too unreliable to support an evaluation finding. Evaluators judge whether your content answers the criteria with verifiable specifics — which is exactly where unedited AI output fails.
What happens if my proposal cites a FAR clause that does not exist?
At minimum, evaluators discount your credibility and may downgrade your rating for lack of understanding. If the fabricated authority supports a material representation, you risk integrity findings that follow your firm. Verify every citation against acquisition.gov or eCFR before submission.
Can I use AI on past performance narratives at all?
Use it to tighten and restructure text after a human writes the facts: contract numbers, scope, outcomes, relevance. Never let it draft from scratch, because it will invent details, and past performance claims are checkable against CPARS and FPDS.
Is it safe to paste a government RFP into ChatGPT or Claude?
For publicly posted solicitations, generally yes — the document is already public. Do not paste controlled or source-selection-sensitive information, and use an enterprise AI tier with data-training turned off for anything containing your own proprietary pricing or technical data.
What is the single highest-value AI task in a proposal effort?
The red-team review. Feeding the model your draft plus Section M and asking it to score you as a skeptical evaluator surfaces generic language, unsupported claims, and contradictions faster than most human color-team reviews — and nothing it writes enters the document.
Will AI-written text hurt me in a best-value tradeoff?
Only if it reads generic. Tradeoff decisions turn on discriminators — the specific strengths the Source Selection Authority can point to. AI cannot invent discriminators you do not have; it can only articulate the ones your team gives it.