Can I use ChatGPT or Claude to write a government proposal, and what are the risks?

What can ChatGPT or Claude do well on a government proposal?

Most RFP teams already use AI: 79% used it in their response work in 2025, up from 34% two years earlier, and 62% use it to generate specific answers^[1]. It does well at summarizing a long solicitation, drafting sections from material you supply, rewriting for length and tone, and turning notes into a first pass at a capability statement. ResponseRail uses AI models for the same work.

Prospects who already draft with AI have told us where it falls short. One said that with enough of the firm’s own material loaded in, the writing is now the easy part; what it needed was to know which bids it qualified for, after an RFP that called for a state disadvantaged-business certification it didn’t have. Another said that even with AI, knitting a response together takes a lot of back and forth, especially with teaming partners.

What are the risks of using ChatGPT or Claude on a government proposal?

The risks are specific:

• Invented facts: asked for past performance it doesn’t have, a model may write plausible contracts, dates, certifications or figures.
• Missed requirements: a chat reads only what you paste, so a required form, page limit or certification in an attachment or amendment you didn’t paste doesn’t exist to it.
• No eligibility check: unless you ask, it will draft a bid you can’t win, such as a set-aside you don’t qualify for.
• Data exposure: pasting pricing, customer details or controlled information into an account that isn’t approved for it can break an NDA, a customer contract or the rules on CUI.
• Someone else’s information: keep another offeror’s proposal and the government’s source selection information out of any AI tool; FAR 3.104-4 limits who either may be disclosed to^[2].
• False statements: the company that submits is responsible for every statement, whoever or whatever wrote it. Knowingly and willfully making a materially false statement to a federal agency is a crime^[3], and under the False Claims Act, knowingly submitting a false claim for payment, or using a false record or statement material to one, makes a company liable for three times the government’s damages plus a civil penalty^[4].
• Solicitation rules: a solicitation or contract can set its own rules on using AI tools and on disclosing that you did.

How do you use ChatGPT or Claude safely on a government proposal?

Treat it as a drafting assistant inside a process a person controls:

1. Read the solicitation and every amendment for rules on AI use, data handling and disclosure before you open a chat.
2. Use a business or enterprise account your company has approved, not a personal one, and check its training, retention and access settings.
3. Keep CUI, classified and export-controlled data out of any tool not authorized for it, and replace customer names, contract numbers and exact rates with placeholders or ranges.
4. Give it the whole package, every attachment and amendment included except anything marked CUI or export-controlled, and have it list each “shall” and “must” with where it appears; that list becomes your compliance matrix against the instructions and evaluation criteria (often Sections L and M in a federal RFP).
5. Check eligibility before drafting, including set-aside type, NAICS size standard, licenses, certifications and required registrations.
6. Draft only from material you’ve verified, then check every name, number, date, certification and citation against your own records.
7. Have an experienced proposal reviewer check the final against the instructions and evaluation criteria before anyone signs, and keep a record of sources and approvals.

Is it safe to share past performance and pricing with an AI tool like ChatGPT or Claude?

Sometimes, but not by default. Check your customer contracts and NDAs first, since they can limit what you may share, and use ranges or placeholders for rates and customer names where you can.

On personal ChatGPT accounts, conversations can be used to train OpenAI’s models unless you first turn off “Improve the model for everyone” in Data controls; OpenAI says it doesn’t train on content from Business, Enterprise or Edu workspaces by default^[5]. On Claude’s Free, Pro and Max plans, you choose whether chats are used for training in Privacy Settings, though chats flagged for safety review may still be used^[6]; Anthropic says it doesn’t train on Claude for Work (Team, Enterprise) or API content by default^[7]. Rating a response with thumbs up or down can send that conversation for training on a personal ChatGPT account or any Claude plan^[5][6][7].

Training is one question; also check how long chats are kept and who can read them. Never paste CUI or classified information into a public AI tool; federal guidance prohibits putting either into AI systems not cleared to handle it^[8].

Do government evaluators penalize proposals written with AI?

Usually not for using AI itself. Under FAR 15.305(a) (15.202(a)(1) under the FAR overhaul deviation), an agency evaluates competitive proposals solely on the factors and subfactors in the solicitation^[9]. Whether an evaluator can tell AI was used matters less than what it left behind: claims with no proof, the solicitation’s own words repeated back, invented details and missed instructions all cost points.

Some agencies now run AI compliance checks first. GSA’s CALI tool checks a proposal’s requirements, forms, and reps and certs against the solicitation, and Lohfeld Consulting, a proposal consultancy, expects that a proposal such a tool finds non-compliant, once the evaluation board confirms it, will likely be eliminated before a human evaluator reviews the approach^[10].

Grants differ: NIH won’t treat applications as original ideas if they, or sections of them, are substantially developed by AI^[11]. If a solicitation sets rules on AI use, follow them.

How is ResponseRail different from using ChatGPT yourself?

ResponseRail is a managed proposal service, not a chat window. It reads the whole solicitation package, checks eligibility and disqualifiers before writing, drafts from the facts your company approved and keeps on file, and an experienced proposal specialist reviews every result. Your files are stored encrypted, only the people working on your response open them, and they are never used to train any AI model. You submit.

Sources

1. 1Loopio, 2026 RFP Response Trends & Benchmarks Report, run with APMP; self-reported
2. 2FAR 3.104-4, Disclosure, protection, and marking of contractor bid or proposal information and source selection information
3. 318 U.S. Code § 1001, Statements or entries generally
4. 431 U.S. Code § 3729, False claims
5. 5OpenAI Help Center, Data controls in ChatGPT
6. 6Anthropic Privacy Center, Is my data used for model training? (Free, Pro and Max)
7. 7Anthropic Privacy Center, Is my data used for model training? (Claude for Work and API)
8. 8NARA ISOO Notice 2026-01, Responsible use of classified information and CUI with AI
9. 9FAR 15.305, Proposal evaluation
10. 10Federal News Network, commentary by Lohfeld Consulting: Federal agencies are using AI to evaluate proposals (Apr 13, 2026)
11. 11NIH, NOT-OD-25-132