More and more of the content I’m handed on live bids started life in an AI tool. Familiar patterns catch my attention with the page. Then the task is to work out what the draft actually says.
Some sentences need untangling. Others sound impressive but say very little. A proposed reporting arrangement appears halfway through a paragraph, although it’s not clear whether anybody has agreed to it or even knows it’s there. I find myself worrying whether a claim came from the supplied evidence or was added to make the answer sound complete.
For a bid writer, this creates two kinds of work. One is correcting and polishing the language, including adapting it to the document and its reader. The other is establishing whether the content is accurate and whether any proposed commitments have been authorised.
Both take effort and may need different people.
The writing habits go beyond punctuation
Em dashes attract plenty of attention in discussions about AI writing. But my concern is broader than a punctuation mark.
I notice long sentences carrying several ideas, vocabulary that makes a simple point harder to understand and paragraphs that keep adding qualifications. A colon introduces a string of separate points, joined by semicolons. Repeated constructions such as “not just A, but B” give ordinary statements an inflated and oftentimes unneeded sense of importance.
These are observations about drafts, rather than proof that a particular passage was written by AI. Human writers use these constructions too. I in fact like an em dash – in the right place! A long sentence can be useful when its parts belong together. The difficulty comes when the same habits recur across a substantial response and obscure what the evaluator needs to know.
My working theory was that the way large language models are trained explained these habits. I asked an AI assistant to investigate, using original sources and checking its claims. After much fact checking and iteration, the research gives a more qualified answer than I expected.
In their 2025 paper, Alex Reinhart and colleagues compared human writing with output from GPT-4o, GPT-4o Mini and four variants of Llama 3. They found systematic stylistic differences, both between human and model-generated writing and between the models. Models further trained to follow instructions tended towards a noun-heavy style that packed information densely, and struggled to match the different writing styles humans use for different kinds of writing. [1]
That helps explain why a draft can be grammatically correct but still be difficult to read. However, the study did not test tender responses specifically or establish that every model consistently produces long sentences, semicolon chains or formulaic contrasts. Those specific concerns need to be assessed in the drafts we use.
Singhal and colleagues’ research, accepted to the Conference on Language Modeling in 2024, found that AI systems trained to score answers could favour longer responses in three experimental settings.[2] This finding relates to the length of whole answers, but doesn’t tell us why individual sentences are long or whether other models have the same writing habits.
E. M. Freeburg’s 2026 paper, published as a preprint, tested 12 models from five providers and found considerable differences in their use of em dashes.[3] He suggests the habit comes partly from training on text written in markdown, a format used for headings, lists and other document structure. That explanation remains a hypothesis.
Some models in Freeburg’s tests used em dashes much more frequently than the human sample average. Others used fewer or none. Frequent em dashes may be a familiar feature of some AI drafts, but they are neither universal across models nor reliable proof of AI authorship.
Different bid documents need different writing
Good tender writing depends on the question being answered, context, and the relevance to the person reading it. “Writing professionally” is something we want, but leaves too much unspecified.
A response to a strategically important question needs to help an evaluator understand why our proposed approach meets the client’s needs. It needs to empathise with the client concerns behind the question, and provide a clear answer tied to the scope and any evaluation criteria and relevant evidence. The structure should make the reasoning easy to follow. Where we claim an advantage, the reader should be able to see what supports it.
A method statement has a different job. It needs to explain how an activity will be carried out, with enough detail to assess whether the sequence and controls are workable. Precise descriptions matter more than promotional language.
A management plan needs to make responsibilities and controls usable. Readers may include evaluators, contract managers and the people who will implement it. They need to understand who does what, when action is required and how problems are escalated.
These documents can overlap. A method statement still needs to give the evaluator confidence. The writing should follow each document’s requirements and intended use.
Every response needs to fit this project
Whatever the document type, the response needs to reflect the project scope, location and duration. Specification requirements and contractual obligations shape what we can propose and how we explain it.
An approach that worked on another project may need changing for this one. Access restrictions can affect sequencing. A remote location can change the staffing arrangements. A contractual reporting requirement can make a generic statement about “regular updates” inadequate.
This is part of the bid writer’s job. An expert asked to explain one activity may not have found time to have read every relevant requirement elsewhere in the tender documents. In any individual drafting task, we need to bring more understanding of the tender context than perhaps our experts may reasonably expect us to.
A key role for the pro bid writer is to assist in connecting expert contribution to those requirements, identify omissions and ask informed questions. They remain responsible for the technical judgement. We need enough understanding of the wider bid to recognise when their draft needs another conversation.
Give the AI brief that same context, including relevant requirements and any amendments. Otherwise, even accurate company information and an approved methodology can produce a response that does not fit the job.
Polishing a sizeable draft can mean rebuilding passages
Replacing an awkward word is straightforward. Correcting the same writing habits across thousands of words is more demanding.
Take this invented example:
“Our integrated maintenance methodology facilitates the proactive identification and timely rectification of asset deficiencies, ensuring continuity of service through coordinated intervention and stakeholder engagement.”
“Identification” and “rectification” describe actions without saying who performs them. The final clause bundles several ideas together. “Ensuring continuity” may promise more than the process can deliver.
A useful revision needs the actual process, including who identifies and assesses the fault, what happens next and how service continuity is protected.
Even when those answers are available and accurate, the editor may need to rebuild the paragraph around them. Across a long response, that means moving information, removing repetition and deciding which details belong in another section. Splitting every long sentence would leave the larger problems untouched.
This is why I am wary of treating a complete-looking draft as nearly finished. Much of the editorial work concerns the organisation of meaning. The measure I care about is how much professional work remains before an evaluator can use the response.
Clearer language can conceal an unapproved commitment
The distinction I want to preserve is this:
“I can fix the style. The slow part is working out what the draft actually says.”
Fixing the style can itself be substantial work. Establishing the truth and authority behind a statement adds another problem.
Adam Tauman Kalai and colleagues’ September 2025 preprint examines why language models produce plausible false statements. They argue that training and evaluation can reward guessing rather than acknowledging uncertainty.[4] The paper concerns general model behaviour. It does not measure errors in bids or the effort required to review them.
In bidding, watch out for proposals presented as authorised decisions. An invented performance figure is a factual problem. A draft promising weekly site inspections may describe a perfectly plausible approach, but it still needs approval if the delivery team has not agreed to it.
Editing “regular inspections will be undertaken” into “our site manager will inspect the works every Friday” makes the sentence clearer. It also introduces an owner and a schedule. Unless both are approved, the editor has made the commitment more specific without establishing its authority.
Evidence of past capability and approval of future delivery need separate attention. A previous project can support a claim about experience. It does not establish that the same staffing or service arrangement has been agreed for this bid.
Resolving these gaps may require a technical specialist or the person authorised to approve the proposed approach. Being clear and confident goes a long way to influencing evaluators. But clarity comes with a commitment and treading this line well is what sets a good bid writer apart.
Separate existing evidence from decisions about this bid
Here are some things you can do to speed up the review of text in the AI writing era. Start by separating content in two ways.
Solution-agnostic content draws on information that exists before we design this bid’s solution, such as company policies, accreditations and insurance details. Start with approved material. Check that it is current, applies to the bidding entity and supports the claim being made. Check project summaries and CVs for relevance to this job too.
Solution-dependent content describes what we propose to deliver here. Methodologies, programs and workforce arrangements require decisions from the team. Until those decisions are made, AI can help organise the response or identify options for discussion. It should leave unresolved commitments visible.
A plausible approach created by AI is worth considering but becomes part of the proposed solution only when the right people agree to it.
A citation still needs checking
A reference can look convincing and still fail to support the sentence beside it. Finding a real document is only the first check. The cited passage must support the actual claim, including its conditions and scope.
The way citations are produced depends on the tool. In the retrieval-based systems studied by Gao and colleagues, relevant passages are retrieved and supplied to the model, which generates an answer with references. Their 2023 study assessed whether statements were supported and whether citations were relevant. It found substantial gaps in citation support in the tested systems.[5] That is evidence of a problem in those tests, rather than an error rate for today’s bid tools.
Some products provide more precise source links. Anthropic’s citation feature, for example, returns references to locations in supplied documents and extracts the cited text directly.[6] That helps a reviewer find the passage. Still check whether it supports the wording used and whether the source is current and applicable to this bid.
For an important response, number the paragraphs in the review draft and keep a simple evidence table alongside it. Ask AI to identify the source and supporting passage for each material claim, then check those references against the originals. Record proposed commitments separately, with the person responsible for approving them.
A single reference at the end of a paragraph can hide partial support. Suppose a paragraph says we completed a comparable project, achieved a particular result and will use the same delivery team here. A project report might support the first two statements. The third needs a staffing decision for this bid.
Ask AI to prepare the evidence record from the supplied material, then check the entries against the originals. Require it to flag conflicting documents for resolution. Any automated comparison should identify the passage and explain its assessment. A model saying “verified” gives you little to work with on its own.
Keep this record with the review draft, even if the internal tags are removed from the submission. When a paragraph changes, check its evidence again. Clearer wording must not quietly broaden what the source supports.
Brief the tool like a new bid writer
Give an AI drafting tool the question, evaluation criteria and intended reader, together with the relevant project requirements, approved evidence and delivery decisions. Specify what the document needs to achieve and show it an appropriate example where one is available.
Tell the tool how to handle gaps. Use [EVIDENCE NEEDED] for an unsupported claim and [DECISION REQUIRED] for an unresolved delivery arrangement. Ask it to produce the paragraph evidence record alongside the draft.
For writing, give specific directions about paragraph focus, vocabulary and sentence structure. Ask it to separate ideas that have been strung together and remove formulaic contrasts. Short sentences are useful, but a rigid word ceiling can make a technical explanation harder to follow.
Then review the result through two distinct checks:
- Writing: Does it answer the question, reflect this project, use an appropriate structure and make the evidence easy to assess? Where do language and repetition make the reader work unnecessarily?
- Accuracy and authorisation: What supports each material claim? Which future commitments have been approved, and which still need a decision?
Resolve material uncertainties before giving a passage its final polish. Otherwise, time goes into refining text that may need to be removed or changed.
A useful AI draft should leave the bid team with less work to do. Judge that by the effort needed to reach a clear, evidenced and approved response, including the work returned to other people.
Next time you’re handed an AI draft, ask the author one question before you touch it: what did you give it to work with?
Book a call and find out how Synthrics solves these problems.
Sources
[1] Reinhart et al. (2025). Do LLMs write like humans? Variation in grammatical and rhetorical styles. PNAS. Full text: https://arxiv.org/html/2410.16107v2
[2] Singhal, Goyal, Xu and Durrett (2024). A Long Way to Go: Investigating Length Correlations in RLHF. Accepted to COLM 2024. https://arxiv.org/html/2310.03716v2
[3] E. M. Freeburg (2026). The Last Fingerprint: How Markdown Training Shapes LLM Prose. Preprint. https://arxiv.org/html/2603.27006v1
[4] Kalai, Nachum, Vempala and Zhang (September 2025). Why Language Models Hallucinate. Preprint. https://arxiv.org/html/2509.04664v1
[5] Gao, Yen, Yu and Chen (2023). Enabling Large Language Models to Generate Text with Citations. EMNLP 2023. https://aclanthology.org/2023.emnlp-main.398.pdf
[6] Anthropic. Citations. Product documentation, accessed 13 October 2026. https://platform.claude.com/docs/en/build-with-claude/citations
