Whitepaper · August 2026
When does AI make our knowledge better, and when narrower? A decision framework for organisations deploying language models in knowledge work.
In the space of two years, language models have moved from handy search aid to co-author of our knowledge work. They summarise, explain, advise, and co-write policy, code and analyses. Something fundamental has changed as a result: systems trained on our knowledge now produce knowledge themselves, which flows back into documents, decisions and data sources, and thus into the environment from which the next answers are drawn. Human and model form a feedback loop.
That this loop exists is by now firmly established scientifically. What is not established is which way it turns. The research shows both directions: AI that makes collective output more uniform, and AI that corrects and broadens the field of view. The interesting question is therefore not whether AI influences our knowledge, but under which conditions that turns out well or badly, and which dials an organisation can turn.
This whitepaper translates our ongoing scientific research into a practical framework. The core: do not judge AI in knowledge work on one axis (“diverse or uniform”), but on two: quality and diversity, with four possible outcomes. Know the two mechanisms that determine the direction: personalisation and plausibility. And recognise that the risks are measurable as properties of the system you deploy. Whoever standardises on a single model chooses, usually unknowingly, a knowledge profile for the entire organisation. That may be a sensible choice. But it should be a deliberate one. Precisely in the environments where Twentynext works every day, public administration, safety and security, and regulated healthcare, this is no academic matter: there, the quality and the spread of knowledge directly determine the quality of decisions about citizens, risks and patients.
A search engine points you to sources and leaves the thinking to you. A language model does part of the thinking itself: it composes an answer, tailored to your question and, increasingly, to you as the person asking. That feels like progress, and often it is. But it does change the role of the technology in our knowledge.
What that answer springs from is not reality itself, but a layered registration of it. Not everything that happens is observed; not everything observed is written down; not everything written down is digitised; and not everything that exists digitally ends up in training data. Every transition selects and colours (Bender et al., 2021). In itself that is nothing new: every technology that carries knowledge colours it, from clay tablet to newspaper.
What is new is the feedback. What a model generates does not vanish into thin air. It is adopted in reports, policy documents, code, teaching materials and web copy, and thus flows back into the environment from which subsequent answers are constructed. This happens along several routes and at very different speeds: search and retrieval systems pick up new documents within days, organisational archives and publications shape what people read, and, slowest of all, future models train on the changed internet. Even a model whose parameters never change can therefore sit in the middle of a loop, as soon as the material it retrieves is increasingly material its own output helped shape.
This is no longer speculation. Experimental research with over fourteen hundred participants showed that biases in AI judgements are adopted by human users and further amplified in the interaction, while accurate systems actually improved human judgements (Glickman & Sharot, 2025). Researchers now speak of human-AI coevolution as a research field in its own right (Pedreschi et al., 2025), and of “lock-in”: models that learn human beliefs, reflect them back and absorb them again (Qiu et al., 2025). The loop is there. The question is what it does to us.
The best-known research result in this field sounds like a warning. Writers who received AI suggestions delivered individually better stories, but the stories became more similar to one another (Doshi & Hauser, 2024). The same pattern appeared in idea generation (Anderson, Shah, & Kreminski, 2024), and a recent synthesis in Trends in Cognitive Sciences warns that widely shared use of the same models can standardise language, perspective and reasoning style (Sourati, Ziabari, & Dehghani, 2026). Individual gain, collective impoverishment: that is a real and documented risk.
And yet it is only half the story. In a large dynamic experiment, more exposure to AI ideas actually led to greater collective diversity of ideas (Ashkinaze et al., 2025). When researchers repeated the famous writers' experiment with ten deliberately different AI personas, the diversity of the stories remained at the level of a group without AI (Wan & Kalman, 2026). And a subtle but important detail: the diversity loss in co-writing did occur with an instruction-tuned model, but not with the underlying base model (Padmakumar & He, 2024). Put differently: the effect does not reside in “AI” as such, but in design and configuration choices.
Whoever lays these results side by side reaches an uncomfortable but liberating conclusion. The question “does AI homogenise our knowledge?” has no yes-or-no answer. The direction is conditional, and that shifts the conversation from fear to dials. Which conditions push an organisation the right way, and which the wrong way?
To sharpen that question, you first have to separate two things that constantly run together in the public debate: the quality of knowledge and its diversity. Convergence is not bad by definition: a team that abandons an outdated approach after solid evidence converges and improves. And diversity is not healthy by definition: ten well-founded scenarios are something different from ten variants of the same mistake. Philosophy of science has known this for a while: communities that reach consensus too quickly can land on the wrong answer, but eternally sustained disagreement also prevents the truth from ever winning (Zollman, 2010).
Cross both dimensions and four possible outcomes of AI in knowledge work emerge:
| Diversity falls | Diversity rises | |
|---|---|---|
| Quality rises | Warranted consensus | Well-founded pluralism |
| Quality falls | Collective error (lock-in) | Fragmented noise |
Warranted consensus: everyone arrives at the answer the evidence actually supports. Well-founded pluralism: several defensible lines continue to exist side by side: exactly what you want where the evidence has not yet decided. Collective error: the most dangerous outcome, because it feels like quality: everyone holding the same plausible, well-phrased, incorrect assumption. And fragmented noise: much variation, little substantiation; the outcome that diversity statistics wrongly register as health.
The practical lesson of this quadrant is simple but strict: whoever measures only diversity cannot distinguish warranted consensus from collective error. And whoever measures only per-answer accuracy does not see what the system does over time to the spread of knowledge in the organisation. You need both axes.
Personalisation: individually more fixed, collectively further apart. Modern assistants adapt to their user, through memory features, and through a tendency that arises from the training method itself: models optimised for human approval sometimes learn to prefer confirming over correcting (Sharma et al., 2024). The remarkable thing is that this mechanism can work in opposite directions on two levels at once. Each individual user becomes more stable in their own view, the system moves along with them, after all, while different users, each confirmed in different starting points, drift further apart. The consequence for organisations: you cannot read collective effects off individual satisfaction. An assistant everyone finds pleasant can, precisely because of that, erode a team’s shared reality. Anti-sycophancy is therefore not merely a matter of individual truthfulness, but collective policy.
Plausibility: the volume dial of the loop. From decades of psychological research we know that fluency and repetition raise the feeling of truth, independent of factual correctness (Hasher, Goldstein, & Toppino, 1977; Fazio et al., 2015). In humans, speaking fluently and coherently is still a weak signal of competence: it takes effort. For language models it costs nothing: they phrase just as convincingly when they are right as when they are wrong. Recent analyses identify precisely this as the core risk: linguistic plausibility taking the place of genuine verification (Quattrociocchi, Capraro, & Perc, 2025), and systems emitting “honest non-signals”: helpfulness and eloquence that carry meaning in humans but prove nothing in machines (Maynard, 2026). Our research adds the loop dimension: in a feedback system, persuasiveness is not a one-off sales argument but an amplification factor. Every convincing answer that gets adopted changes the environment from which the next answer is drawn.
Here it becomes interesting for practice. In our ongoing research programme we reduce these mechanisms to a small number of measurable properties of a deployed system, not philosophy, but audit quantities:
Together they answer one practical question every organisation should be able to ask: in our configuration, does evidence beat repetition? We measure this with controlled, fictional knowledge domains in which we vary the frequency of a claim and the quality of the evidence independently of each other, so the answer cannot come from the model’s memory, but reveals its actual behaviour.
One preliminary theoretical result from our formal model deserves to be shared now, with the caveat that mathematical verification is still under way: whether a bias in such a human-model system can run away turns out, in the model, to depend on the combination of frequency sensitivity and evidence sensitivity: properties of the system itself. How intensively it is trusted and reused then determines how fast that happens, not whether it can. If that picture holds, the governance translation is direct: the derailment risk lies in what you procure and how you ground it; usage intensity only sets the tempo. Current AI evaluations, which measure per-interaction accuracy, look right past this.
Five courses of action follow directly from the framework. They are written with the practice of public-sector agencies, safety and security services and healthcare institutions in mind, environments where hundreds of professionals work on assessments, analyses and case files with the same assistant, and where a collective error is therefore not a matter of style but an operational risk.
This whitepaper is the accessible translation of a scientific working paper currently in preparation. Phase one: characterising present-day model families on the four properties above, and computing the loop dynamics based on those measurements, is what we are carrying out now. Phase two consists of large-scale, preregistered experiments with human participants, in which the two core predictions (the personalisation reversal and the plausibility amplification under feedback) are tested confirmatorily. For that phase we are explicitly looking for academic and institutional partners.
Do you work at a university or research institute and would you like to take part in phase two? Or does your organisation deploy AI at scale in knowledge work and would you like to hold your own configuration up to this yardstick, from an initial audit to implementation and ongoing management? Get in touch via info@twentynext.nl.
Martijn van Grieken is Director of Data & AI at Twentynext, the Eindhoven-based data and AI consultancy that has worked for sectors including public administration, safety and security, and regulated healthcare since 2014. He has spent over ten years at the intersection of data governance, AI and public decision-making.
Twentynext is a data and AI consultancy from Eindhoven, founded in 2014, supporting organisations in sectors including public administration, safety and security, and regulated healthcare, from the secondment of specialists to projects with pre-agreed delivery milestones and ongoing management.
“The epistemic loop” is our working term for what the scientific literature knows as human-AI feedback loops and human-AI coevolution; this paper claims no discovery of that phenomenon, but rather a framework for giving it direction and measurability. All cited sources have been verified by the author. The theoretical stability result mentioned is preliminary and is currently being independently checked; the empirical measurements from phase one will be published upon completion.
AI language models were used in drafting this paper, under the editorial control and responsibility of the author.
Anderson, B. R., Shah, J. H., & Kreminski, M. (2024). Homogenization effects of large language models on human creative ideation. Proceedings of the 16th ACM Conference on Creativity & Cognition, 413–425. https://doi.org/10.1145/3635636.3656204
Ashkinaze, J., Mendelsohn, J., Qiwei, L., Budak, C., & Gilbert, E. (2025). How AI ideas affect the creativity, diversity, and evolution of human ideas: Evidence from a large, dynamic experiment. Proceedings of the ACM Collective Intelligence Conference, 198–213. https://doi.org/10.1145/3715928.3737481
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–623. https://doi.org/10.1145/3442188.3445922
Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290
Fazio, L. K., Brashier, N. M., Payne, B. K., & Marsh, E. J. (2015). Knowledge does not protect against illusory truth. Journal of Experimental Psychology: General, 144(5), 993–1002. https://doi.org/10.1037/xge0000098
Glickman, M., & Sharot, T. (2025). How human–AI feedback loops alter human perceptual, emotional and social judgements. Nature Human Behaviour, 9(2), 345–359. https://doi.org/10.1038/s41562-024-02077-2
Hasher, L., Goldstein, D., & Toppino, T. (1977). Frequency and the conference of referential validity. Journal of Verbal Learning and Verbal Behavior, 16(1), 107–112. https://doi.org/10.1016/S0022-5371(77)80012-1
Kleinberg, J., & Raghavan, M. (2021). Algorithmic monoculture and social welfare. Proceedings of the National Academy of Sciences, 118(22), e2018340118. https://doi.org/10.1073/pnas.2018340118
Maynard, A. D. (2026). The AI cognitive Trojan horse: How large language models may bypass human epistemic vigilance (arXiv:2601.07085). arXiv. https://doi.org/10.48550/arXiv.2601.07085
Padmakumar, V., & He, H. (2024). Does writing with language models reduce content diversity? Proceedings of the 12th International Conference on Learning Representations (ICLR 2024).
Pedreschi, D., Pappalardo, L., Ferragina, E., Baeza-Yates, R., Barabási, A.-L., Dignum, F., Dignum, V., Eliassi-Rad, T., Giannotti, F., Kertész, J., Knott, A., Ioannidis, Y., Lukowicz, P., Passarella, A., Pentland, A., Shawe-Taylor, J., & Vespignani, A. (2025). Human-AI coevolution. Artificial Intelligence, 339, 104244. https://doi.org/10.1016/j.artint.2024.104244
Qiu, T. A., He, Z., Chugh, T., & Kleiman-Weiner, M. (2025). The lock-in hypothesis: Stagnation by algorithm. Proceedings of the 42nd International Conference on Machine Learning.
Quattrociocchi, W., Capraro, V., & Perc, M. (2025). Epistemological fault lines between human and artificial intelligence (arXiv:2512.19466). arXiv. https://doi.org/10.48550/arXiv.2512.19466
Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., … Perez, E. (2024). Towards understanding sycophancy in language models. Proceedings of the 12th International Conference on Learning Representations (ICLR 2024).
Sourati, Z., Ziabari, A. S., & Dehghani, M. (2026). The homogenizing effect of large language models on human expression and thought. Trends in Cognitive Sciences. Advance online publication. https://doi.org/10.1016/j.tics.2026.01.003
Wan, Y., & Kalman, Y. M. (2026). Diverse AI personas can mitigate the homogenization effect in human-AI collaborative ideation. Computers in Human Behavior: Artificial Humans, 8, 100289. https://doi.org/10.1016/j.chbah.2026.100289
Zollman, K. J. S. (2010). The epistemic benefit of transient diversity. Erkenntnis, 72(1), 17–35. https://doi.org/10.1007/s10670-009-9194-6
Join or apply?
Do you work at a university or research institute and would you like to join phase two of the research programme? Or does your organisation deploy AI at scale in knowledge work and would you like your own configuration measured, from an initial audit to implementation and ongoing management?
Work with us
The people who build it also run it afterwards. Eindhoven, since 2014.
Martijn van Grieken
Director Data & AI
We use Google Tag Manager to measure visits and Leadinfo to recognise which company is visiting. Neither loads unless you agree. If you choose essential only, the site works as normal and we measure nothing. If you arrived via an advertisement in ChatGPT, we also use the OpenAI measurement pixel to attribute conversions to that advertisement. Cookie statement · Privacy statement
