An AI-assisted grant bid now needs its own evidence file

The dangerous part of an AI-assisted grant application is not the polished sentence. It is the missing account of where the idea came from, which facts were checked, what confidential material entered the tool and who accepted the final claim.

That record has become an immediate submission issue for UK space businesses. The UK Space Agency’s third National Space Innovation Programme call offers up to £20 million across an anticipated 15 to 30 grants. Its optional expression-of-interest window closed on 7 August, but outline proposals remain open until 23:59 UK time on Friday 21 August 2026.

The call permits artificial intelligence as a supporting tool. It also draws an unusually sharp boundary around that permission: the application must contain the applicant’s original ideas, AI should not substantially develop it, and the applicant must declare which tool was used and in what capacity. Responsibility for truth, accuracy, attribution and security stays with the applicant.

This is more than a sentence to add before submission. UKSA says plagiarism may lead to rejection and that substantive or undeclared AI use discovered after an award can result in withheld funds, grant suspension or termination. A team therefore needs evidence that survives beyond the person who happened to operate the chatbot.

The declaration is the end of the process

The guidance does not supply a universal template for the declaration or define “substantially develop” with a word count. That makes an internal boundary more important, not less.

A weak process asks the bid lead to remember every prompt on submission day. A stronger one decides in advance which activities are permitted and records them when they happen. The final declaration can then be short because the evidence behind it is specific.

Start with three zones:

• Mechanical support: formatting, spelling, reducing a passage to a stated limit or changing its structure without adding claims.
• Analytical support: suggesting alternatives, summarising material, comparing sources, generating an abstract or challenging an argument.
• Substantive development: originating the project concept, inventing the technical route, writing unsupported experience, constructing the commercial case or producing whole answers that the team merely approves.

The first zone still falls under the wording of the NSIP Call 3 instruction to declare which tool was used and in what capacity. The second needs a clear human source-and-review trail. The third is where the application risks ceasing to represent the applicant’s own work.

Do not treat these labels as a legal test invented by the funder. They are an internal control for keeping the proposal on the permitted side of the published rule.

Other funders draw a different line

UK Research and Innovation updated its general policy in July 2026. It expects transparency for substantive uses such as developing hypotheses, analysing data, comparing literature, generating code or drafting an abstract. It says minimal uses such as translation, improving English, formatting or reducing word count do not need disclosure.

That is useful context, but it is not permission to import UKRI’s minimal-use exception into a separate UKSA call. The specific opportunity governs the application being submitted. For NSIP Call 3, the safer reading is literal: if an AI tool was used, name it and state its capacity.

Document	Applicant boundary	Disclosure position	Assessor boundary
NSIP Call 3 guidance	AI may support the application but should not substantially develop it.	Declare the tool and the capacity in which it was used.	External assessors are not permitted to use AI to assess applications.
UKRI general policy	No entire application or sections generated without human involvement.	Substantive use should be transparent; listed minimal uses need not be disclosed.	AI is barred from assessment except limited language refinement with strict confidentiality controls.

Source: UK Space Agency NSIP Call 3 guidance and UKRI application and assessment policy. Last verified 2026-08-08.

The contrast matters because a business may reuse one bid process across several programmes. A declaration checklist built around the least demanding policy will fail when a call uses stricter wording.

Keep one row for each use

An AI-use file does not need prompt-by-prompt surveillance of every keystroke. It needs enough information for a reviewer to reconstruct the material contribution and the human decision.

Use one row per distinct task:

Field	What to record
Tool	Product, model if shown, account or workspace type, and date used
Capacity	Formatting, challenge, source comparison, outline critique, language edit or another precise purpose
Input boundary	Documents or data categories supplied, plus sensitive material deliberately excluded
Output used	The suggestion, structure or wording retained — or a note that nothing was retained
Human work	Material edits, rejected suggestions and the applicant’s own reasoning
Verification	Source links, calculations, evidence owner and checks performed
Approval	Named reviewer, decision and date
Declaration	The plain-language description that will appear in the application

Source: Flint Brief operational record derived from the disclosure, accuracy and confidentiality duties in the UKSA and UKRI guidance. Last verified 2026-08-08.

The row is deliberately centred on use, not conversation history. A raw transcript can be both excessive and incomplete: it may expose sensitive material while failing to show which output reached the bid. The useful record joins the retained contribution to its evidence and approver.

Protect the proposal before prompting

Space proposals can contain export-sensitive details, intellectual property, partner commitments, security assumptions, cost models and named staff information. The call itself warns applicants to use caution with commercially sensitive information and to follow their organisation’s AI policy.

The National Cyber Security Centre makes the practical risk clear. Queries to a public large-language-model service are visible to its operator, may be stored, and may later be exposed through a breach or accidental publication. NCSC recommends not placing sensitive information in public-model queries and not sending a query that would cause harm if made public.

Translate that into a prompt boundary before the drafting starts:

1. Mark the source pack as public, internal, confidential, personal or security-sensitive.
2. Decide which approved tool and account type may receive each category.
3. Remove partner names, personal data, unpublished performance figures, security details and proprietary diagrams unless an approved environment and purpose require them.
4. Give the tool a redacted placeholder, not a secret followed by an instruction to ignore it.
5. Record the category supplied in the AI-use row; do not copy the sensitive source into the log.

This boundary must include retrieved files and connected drives, not just text pasted into a chat box. UKSA explicitly notes the risk of models acting inconsistently with instructions or accessing files that were not requested. A workspace connector can widen the input without making the extra material visible in the final sentence.

Verify against the four scored questions

The outline proposal asks what the project is and how it supports strategic goals, whether it is innovative and technically feasible, who will deliver it, and what economic benefits and route to exploitation it offers. An elegant generic answer can fail all four.

For each answer, attach the claims to an owner and an artefact:

• strategic alignment to the published call objective and a named project outcome;
• technical novelty to prior work, a demonstrator, test result or defensible engineering argument;
• delivery capability to real people, roles, time commitments and partner letters;
• economic benefit to a stated market, evidence, assumptions and a route the company can execute.

AI can challenge whether an argument is complete. It cannot manufacture the underlying project history, customer evidence, partner commitment or technical result. A sentence without an artefact should be removed, qualified or assigned for verification.

The partner-letter check deserves particular attention. UKSA says missing letters of intent for named partners will cause rejection. No amount of fluent drafting compensates for a missing required document.

Run a clean-room review

Before submission, give a reviewer the draft, the call guidance, source pack and AI-use file — but not the original chat history. Ask that person to reconstruct the application from the controlled record.

The review passes only if the reviewer can answer five questions:

1. Which ideas and technical choices demonstrably came from the applicant team?
2. Where did AI materially influence analysis, structure or wording?
3. Can every quantitative, legal and market claim be traced to a current source or calculation?
4. Did any protected information reach a tool outside its approved boundary?
5. Does the proposed declaration match the actual use without euphemism or overstatement?

Then compare the final portal answers with the reviewed copy. A late shortening pass can introduce a new claim or remove the qualification that made an old one accurate. Freeze the submitted version and its declaration together.

Make the evidence smaller than the bid

The goal is not to create a second application about the first. A useful AI-use file for an outline bid may contain only a handful of rows, a source list, one approval note and the final disclosure wording.

Build it now, while the team is still drafting. Record the next AI-assisted task before opening the tool, then have someone other than the operator verify what reached the proposal. If that small file cannot explain the contribution, the final declaration will not repair the gap on 21 August.