--- site: agentic-rfp-software.ai slug: /une-ia-qui-n-apprend-que-de-vous langue: en title: Why two companies using the same generic AI produce similar-looking responses, and how a private instance pulls them apart. description: When several companies respond with the same AI model, their responses converge: this has been measured. A private instance, corrected only by your own experts, produces the opposite effect. What the studies say, and what it changes for a response competing against others. rubrique: Understand intention: Info requete_type: factual derniere_revue: 2026-09-05 statut: draft, pending review --- # Why two companies using the same AI produce the same response When two companies respond with the same generic AI, their responses converge: it is measured. AI on a private instance, corrected only by your experts, produces the opposite effect. ## What happens when everyone responds with the same AI? A tender response, a security questionnaire, a due diligence questionnaire (the DDQ a customer sends before awarding you a contract): these are documents put in competition with each other. They are read side by side, by the same evaluator, to tell apart companies answering the same questions. Anything that makes them look alike works against you. Consumer-grade generative AI, the kind your teams already use to move fast, has a documented effect on text produced by different people: it pulls them closer together. This is not an impression, it is a result measured across three separate fields of research. **Assisted writing.** In a controlled experiment on short stories, text written with AI-generated ideas was judged individually more creative and better written, but it resembled other such text more than text written without help: a gain for each author, a loss of diversity for the group as a whole (Anil R. Doshi and Oliver P. Hauser, *Science Advances*, 2024). A second experiment, on argumentative essays, measured that writing with a language model fine-tuned on human feedback increases similarity between different authors' texts and reduces content diversity, the effect coming from the narrow range of text the model brings to each person (Vishakh Padmakumar and He He, ICLR 2024). **Assisted decision-making.** When independent decision-makers rely on the same algorithm, their decisions align, and the quality of the group's decisions can drop even as each individual decision improves: researchers call this algorithmic monoculture (Jon Kleinberg and Manish Raghavan, *PNAS*, 2021). Sharing the same foundation model across different systems homogenizes the outcomes the same people get from one system to another (Rishi Bommasani et al., NeurIPS 2022). **Competitive responses.** What holds for stories, essays, and decisions holds, by construction, for responses that get compared. Several companies drafting their technical response with the same generator submit documents with the same structure, the same phrasing, the same generic commitments. The evaluator does not read distinct responses; it reads variants of a single one, and it decides on what is left: price, or the incumbent. ## Why does a private instance do the opposite? Convergence comes from a specific mechanism: the same model, trained on the same corpus, corrected by feedback from a very large number of users, pulls every text toward the same center. A private instance reverses each of these three terms. | What makes a shared model converge | What makes a private instance diverge | What that changes in the response | |---|---|---| | A corpus common to every user | Your reference base, and only yours: your policies, your procedures, your already-validated answers, indexed by a librarian agent | Every answer cites your document, your page, your version; it cannot resemble the neighbor's, which cites theirs | | Corrections diluted into a global model | Your corrections improve your AI instead of being diluted into a global model | When your expert rejects a recommendation, your AI does not make the same suggestion again; when they accept it, your reference base grows richer | | An average style, that of the largest group | Agents specialized in your role and your sector, writing in your reader's context | A compliance question in energy does not call for the same answer as a quality question in food and beverage; a private instance tells them apart, a shared model blurs them | On the Optivalue.ai platform, which publishes this site, this private instance mobilizes 85 specialized agents, 72 for your role, 12 for your sector, and a librarian that structures your knowledge. Over time, you do not have the same tool as your competitors. You have your own. The point that matters: this is not a promise of speed or accuracy. It is a property of structure. A model that learns only from one organization cannot pull it closer to others. ## The concession that clarifies everything For a draft, an internal note, a summary no one will sign, generic AI remains the right tool, and convergence costs nothing there: no one compares your internal notes to your competitors'. A draft commits no one. Your response does. The question of convergence only arises the moment the document is put in competition in front of a third party who decides. That is where, and only where, AI that learns only from you changes the outcome. ## Three questions that draw the line To find out whether the tool you use, or are being offered, pulls you apart from your competitors or closer to them: 1. **Where does the text come from?** If the answer is "the model," your competitors have access to the same one. If the answer is "your documents, cited to the page," no one else has those documents. 2. **Where do your corrections go?** If they improve a shared model, they benefit everyone, including those responding to the same tender. If they stay within your instance, they benefit only you. 3. **Would two companies in the same sector get the same answer to the same question?** With a shared model, yes, give or take a few words. With a private instance, no: each cites its own reference base. ## Frequently asked questions **Can't generic AI be "personalized" with our documents?** It can read your documents for the length of one conversation. But the model remains everyone's, its corrections go to everyone, and its style is the average of the largest group. Surface-level personalization does not change the convergence mechanism described in the studies cited. **Does "diverging" mean answering worse?** No. Diverging means answering from your own evidence, in your reader's context, with your own commitments. That is what the evaluator is looking for: a response specific to the opportunity, not a plausible-sounding response in general. **Do our competitors really use the same AI as us?** Consumer-grade generative models are commonly used by companies responding to tenders. Two bidders on the same tender can therefore end up writing, without realizing it, with the same generic model, whereas AI on a private instance learns only from your own documents. **What happens to our reference base if we switch tools?** A private instance worth the name provides for handing back your data in a reusable format, dated deletion of everything derived from it, and a certificate. Your reference base belongs to you. **Does this mechanism depend on machine learning trained on our data?** It rests first on two things that are simply available: your reference base as the only source for drafting, and your corrections staying within your instance. These two properties are what cause divergence. No speed or accuracy figure is attached to this argument, deliberately. --- ## See a private instance at work on your own documents Bring a real tender, a security questionnaire, or a DDQ. You will see the extraction coverage, the sources cited to the page, and the gap analysis on your document, not a prepared demo. [Discover the Optivalue.ai platform](https://www.optivalue.ai/en/?utm_source=agentic-rfp-software.ai&utm_medium=satellite&utm_campaign=une-ia-qui-n-apprend-que-de-vous) ยท [Book a demo](https://www.optivalue.ai/en/demo/?utm_source=agentic-rfp-software.ai&utm_medium=satellite&utm_campaign=une-ia-qui-n-apprend-que-de-vous) Written by the compliance and presales team at Optivalue.ai. Last reviewed: 5 September 2026. [Markdown version](/en/une-ia-qui-n-apprend-que-de-vous.md) ### Sources cited - A. R. Doshi, O. P. Hauser, "Generative AI enhances individual creativity but reduces the collective diversity of novel content," *Science Advances*, vol. 10, no. 28, 2024. https://www.science.org/doi/10.1126/sciadv.adn5290 - V. Padmakumar, H. He, "Does Writing with Language Models Reduce Content Diversity?," ICLR 2024. https://arxiv.org/abs/2309.05196 - J. Kleinberg, M. Raghavan, "Algorithmic monoculture and social welfare," *PNAS*, vol. 118, no. 22, 2021. - R. Bommasani, K. A. Creel, A. Kumar, D. Jurafsky, P. Liang, "Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome Homogenization?," NeurIPS 2022. https://proceedings.neurips.cc/paper_files/paper/2022/hash/17a234c91f746d9625a75cf8a8731ee2-Abstract-Conference.html --- ## Further reading - [How artificial intelligence helps you respond to a tender or RFP](/en/) - [Are your tender responses protected when you use AI](/en/vos-reponses-sont-un-secret-des-affaires) - [How to choose tender-response software](/en/comment-evaluer-un-agent) - [How to respond to a tender for an e-sourcing platform](/en/comment-repondre-appel-offres-plateforme-sourcing)