AI-Assisted Option Generation Without Anchoring

A Non-Anchoring Brainstorm Protocol for ADR and Negotiation Practitioners

This article is the first instalment in a three-part practice series on controlling anchoring risk when using generative AI in negotiation: (1) a non-anchoring workflow protocol; (2) non-anchoring prompt engineering patterns and a template; and (3) provides facilitator-grade de-anchoring interventions for real-time, AI-assisted brainstorming in negotiation.

Generative AI can speed up option generation in negotiation and mediation, but the first plausible option an AI produces can function like a “first offer,” anchoring expectations and narrowing the solution space. A preregistered meta-analysis by Petrowsky et al. (2025)  Link shows that making a very high first offer can help the person who makes it get a better deal when an agreement is reached.At the same time, it also makes deals less likely and leaves the other side feeling less satisfied – partly because the first number shapes what seems “reasonable” and partly because it can trigger anger, especially in more complex negotiations. Put simply: anchoring can “work,” while simultaneously harming deal sustainability and procedural justice.

In human–AI decision-making, AI suggestions can “pull” people’s judgments in a particular direction, functioning like an anchor. Pausing, deliberately considering an alternative, or requiring a quick sanity-check step can help people think for themselves instead of blindly following the AI. This article translates academic evidence into a practical, negotiator-usable deliverable: a Non-Anchoring Brainstorm Protocol for AI-Assisted Negotiation (NABP). The core design is a “blinded brainstorming” workflow: parties generate options first, the AI generates options in parallel from sanitized (non-identifying) inputs. The AI output is disclosed only after the negotiator/mediator runs anti-anchoring steps at the negotiation table   and uses anti-anchoring prompt design to prevent the AI from presenting biased, anchoring solutions.

The problem: AI suggestions can become “procedural anchors”

Anchoring is the tendency for judgments to assimilate toward a previously presented value or framing – even if the anchor is arbitrary.  In negotiation, this is not just academic trivia: the first offer reliably shapes counteroffers, perceived fairness, and shaping where the final agreement lands within the ZOPA – and sometimes whether the parties reach agreement at all.

Now translate that into today’s practice: when a negotiator, mediator or lawyer asks an LLM (ChatGPT, Claude, Gemini, Mistral) to “suggest settlement options,” the AI’s first coherent proposal can function as a default reference point – especially when (a) parties are fatigued, (b) issues are complex, or (c) the AI’s output is presented with polished confidence. That is an anchoring hazard, not because professionals are naïve, but because anchoring is a fast cognitive shortcut that operates even under expertise.

Bracketing also generates a specific bias in this AI-assisted brainstorming workflow: if an AI output introduces ranges early, you may be injecting two or more anchors at once (“tandem anchoring”), shifting counterparts via both endpoints and via politeness norms that make assertive counteroffers feel less acceptable. (Ames & Mason, 2015). Link

Bias in Human – AI Collaboration: Known Effects, Proven Countermeasures

The studies listed below provide practical, evidence-based guidance. If you’d like to dig deeper, I’ve included a link after each reference so you can access the original source directly.

• In AI-assisted decision-making, Carter and Liu (2025) experimentally demonstrate anchoring effects from recommendations, but they also present a consider-the-opposite method as an effective debiasing approach we can use.  Link
• Rastogi et al. (2020) modeled cognitive biases in human – AI collaboration. They also tested a time-based de-anchoring strategy, showing that allocating time and effort to reflection can reduce anchoring and improve joint human – AI performance. Especially when the AI is wrong. Link
• Buçinca et al. (2021) found that cognitive forcing functions – adding built-in “speed bumps” that force people to pause and think – can reduce overreliance on AI. The downside is that people like it less and it takes more effort and time. Link
• A systematic review by Romeo and Conti (2026) synthesizes how automation bias – over-reliance on automated recommendations – manifests in modern human -AI work. They conclude that just showing AI outputs and explanations often isn’t enough; better workflow designs should prompt people to double-check and think independently. In negotiation, this bias can show up as: “If the tool suggested it, it must be reasonable.” Link
• A newer technical angle is that LLMs can show the same anchoring bias as humans in negotiation simulations. If the AI agent is anchored, your process can inherit that bias. Takenami et al. (2025) study anchoring effects in LLM-driven price negotiation simulations and report that models that do more explicit step-by-step reasoning are less affected by anchors. Link Related work (Bianchi et al., 2024; Shah et al., 2025) finds that LLM negotiators often open with unusually extreme numbers and stick to them across different scenarios, and this systematic “extreme anchoring” can quietly import the model’s anchoring bias into the negotiation. Link, Link

Bottom line: anchoring risk in AI-assisted option generation is real and measurable; however, well-designed workflows, non-anchoring prompt patterns, and facilitation guardrails can meaningfully mitigate it.

Non-Anchoring Brainstorm Protocol for AI-Assisted Negotiation

Here I offer a brief, “behind-the-scenes” look at my Non-Anchoring Brainstorm Protocol (NABP) – not as a one-size-fits-all recipe, but as a set of practical design principles you can adapt to your own negotiation and mediation context when using GenAI.  The protocol was developed for GenAI-assisted option generation in international high-asset divorce and cross-border financial settlement negotiations, where complexity and uncertainty make premature convergence particularly costly. Its underlying logic is transferable and can be adapted to commercial disputes, workplace conflict, and other multi-issue negotiations in which early AI outputs risk narrowing the option space.

1. Practically, NABP begins before anyone is in the room—with an explicit decision about the “AI lane.” In higher-risk matters, the cleanest lane is a no-upload workflow built on a sanitized, AI-ready brief (data minimization, no identifiers, no client documents). But if the parties insist on richer inputs, the alternative is a controlled private or enterprise environment with explicit retention, access, and audit controls. In other words, after consulting the parties, you must decide whether the session will use off-the-shelf public LLMs (with strict no-upload/sanitization rules) or a secured deployment (enterprise tenant or private instance) designed for sensitive professional use.
2. Session 1 is then engineered to prevent early fixation by mapping the landscape before generating solutions: you externalize jurisdictions, asset blocks, and constraints so the parties do not anchor on a single forum, a single jurisdiction, a single asset, or the first narrative that feels “complete.”
3. From there, you “freeze the unknowns” into an Open Questions Log before option generation begins. This step matters because parties often anchor to what is vivid rather than what is verified. You should pause any substantive brainstorming and run a short verification sprint instead, because anchoring to an incomplete factual picture is hard to unwind once the parties` story hardens (Glesner Fines, 2024). In other words, NABP treats uncertainty as a gating variable: when the unknowns are too large, the right move is not “more creativity,” but better inputs.
4. Only after that gate, generate the human brainstorming baseline through silent writing (NGT-lite) and round-robin “titles-only” harvesting, so the first detailed idea does not dominate simply because it arrives with more narrative force. If you need additional divergence without debate, add a short brainwriting round (“build on, don’t argue”) in which each participant silently improves two options and drafts one alternative for each.
5. A short, written consider-the-opposite checkpoint follows – asking each side to articulate why their current favorite could be wrong. At the same time, they should identify two options they do not want to lose from their list.
6. Between sessions, you can add a mini-Delphi loop to your protocol when cross-border expertise is needed, not to inflate complexity but to prevent “strongest voice wins” dynamics among professionals and to clarify feasibility constraints without turning one jurisdiction’s counsel into the de facto anchor.
7. Session 2 then runs the AI brainstorming privately “behind the screen”. You should repeat the bias gate, and reveal AI output only in layers – categories, then titles, then details – so the parties control the order of exposure and tag ideas before any polished explanation can take over. This layered reveal is a practical way to keep AI from becoming an authority cue, and it pairs naturally with time-based de-anchoring findings (Rastogi et al., 2020) showing that protecting reflection time before committing to an AI suggestion improves human -AI performance.  
8. A good NABP should also be intentionally conservative with numbers, because once a numeric anchor is “seen,” it is difficult to unsee, and ranges can be even more potent: range offers can produce tandem anchoring, where both endpoints influence counterparts via informational and politeness effects (Ames & Mason, 2015). In practice, that means delaying numeric talk until parties have articulated their own constraints and priorities and avoiding AI-generated ranges early because they can inject two or more anchors at once while still feeling “neutral.”
9. As an optional add-on after AI output is revealed, mediators can borrow from Edward de Bono’s CoRT attention-directing tools as “plug-in” micro-rounds—used only when helpful—to keep brainstorming divergent and to prevent a polished AI explanation from becoming the default. For example, PMI (Plus/Minus/Interesting) can quickly surface benefits, downsides, and curiosities without turning into advocacy; APC (Alternatives/Possibilities/Choices) can force additional routes to the same goal when the room is narrowing too fast; CAF (Consider All Factors) can flush out missing constraints (e.g., jurisdiction, enforcement, liquidity, timing, tax, valuation uncertainty) and convert them into contingencies; OPV (Other People’s Views) can generate options that would look reasonable from the other side or relevant third parties; and C&S (Consequences & Sequels) can add a light downstream check (immediate / short-term / long-term) before anyone “locks in”.
10. Finally, the protocol should slow down the “rush to decide” on purpose. I recommend to add small speed bumps—like a counter-anchor round (“what’s the opposite approach?”), two-package drafting (build two workable bundles instead of falling in love with one), and a premortem (“assume this deal failed—what went wrong?”). The point is not just to pick the first promising idea, but to keep exploring long enough to build a strong package that holds up when you test it against taxes, timing, enforcement, and human reactions. This matters even more because early research suggests that LLMs can show anchoring patterns in negotiation simulations, which means AI suggestions may come with built-in momentum unless you design the process to resist it.

The plain takeaway for practitioners is that your Non-Anchoring Brainstorm Protocol (NABP) should keep AI in its proper place: you are not choosing the AI’s ideas—you are using AI to expand the menu after humans establish an independent baseline, and then pressure-testing party-owned packages against verified constraints, with a clear Option Source Log (Party A / Party B / Joint / AI).

In Part 2, I translate this protocol into copy-paste prompt template and a “prompt hygiene” checklist designed for mediators, ADR and Negotiation Practitioners.

Dr. Blanka Illés

References:

Ames, D. R., & Mason, M. F. (2015). Tandem anchoring: Informational and politeness effects of range offers in social exchange. Journal of Personality and Social Psychology, 108(2), 254–274. https://doi.org/10.1037/pspi0000016

Bianchi, F., Chia, P. J., Yuksekgonul, M., Tagliabue, J., Jurafsky, D., & Zou, J. (2024). How well can LLMs negotiate? NegotiationArena platform and analysis [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2402.05863

Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 188. https://doi.org/10.1145/3449287

Carter, L., & Liu, D. (2025). How was my performance? Exploring the role of anchoring bias in AI-assisted decision making. International Journal of Information Management, 82, 102875. https://doi.org/10.1016/j.ijinfomgt.2025.102875

de Bono. (n.d.). de Bono thinking lessons 1. Retrieved February 15, 2026, from https://www.debono.com/de-bono-thinking-lessons-1

Glesner Fines, B. (2024). Biases & mediation practice. Journal of the American Academy of Matrimonial Lawyers, 37, 117–145. https://aaml.org/wp-content/uploads/biases-and-mediation-practice-glesner-fines-vol-37-2024.pdf

Petrowsky, H. M., Boecker, L., Escher, Y. A., Frech, M.-L., Friese, M., Galinsky, A. D., Gunia, B., Lee, A. J., Schaerer, M., Schweinsberg, M., Soliman, M., Swaab, R., Troll, E. S., Weber, M., & Loschelder, D. D. (2025). The power and peril of first offers in negotiations: A conceptual, meta-analytic, and experimental synthesis. Organizational Behavior and Human Decision Processes, 191, 104448. https://doi.org/10.1016/j.obhdp.2025.104448

Rastogi, C., Zhang, Y., Wei, D., Varshney, K. R., Dhurandhar, A., & Tomsett, R. (2020). Deciding fast and slow: The role of cognitive biases in AI-assisted decision-making [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2010.07938

Romeo, G., & Conti, D. (2026). Exploring automation bias in human–AI collaboration: A review and implications for explainable AI. AI & Society, 41, 259–278. https://doi.org/10.1007/s00146-025-02422-7

Shah, C., Agarwal, A., Garg, K., & Heddaya, M. (2025). LLM Rationalis? Measuring bargaining capabilities of AI negotiators [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2512.13063

Takenami, Y., Huang, Y. J., Murawaki, Y., & Chu, C. (2025). How does cognitive bias affect large language models? A case study on the anchoring effect in price negotiation simulations. In Findings of the Association for Computational Linguistics: EMNLP 2025 (pp. 4481–4498). Association for Computational Linguistics. https://doi.org/10.18653/v1/2025.findings-emnlp.240

Sample, J. A. (1984). Nominal group technique: An alternative to brainstorming. Journal of Extension, 22(2), Article 10. https://commons.joe.org/joe/vol22/iss2/10

Voß, M., Bozkurt, H., Sauer, T., & Nutzmann, M. (2022). Group ideation with brainwriting – A comparison of co-located and distance collaboration. International Conference on Engineering and Product Design Education (E&PDE 2022).

Blanka Illés

Dr. Blanka Illés is an international family lawyer and mediator with 26+ years of experience in complex, multi-jurisdictional family disputes, specialising in cross-border matrimonial property division and high-asset settlements. She graduated cum laude from Eötvös Loránd University (Budapest), founded her own law firm focused on international family and inheritance matters,… MORE